Daily Papers

Recent advances in multi-agent systems have shown great potential for solving complex tasks. However, when multiple agents edit a shared codebase concurrently, their changes can silently conflict and inconsistent views lead to integration failures. Existing multi-agent systems address this through workspace isolation (e.g., one git worktree per agent), but this defers conflict resolution to a post-hoc merge step where recovery is expensive. In this paper, we propose STORM, i.e., STate-ORiented Management for multi-agent collaboration. Specifically, STORM manages agent states by mediating their interactions with the shared workspace, ensuring that each agent operates on a consistent view of the codebase and that conflicting edits are detected and resolved at write time. We evaluate STORM on Commit0 and PaperBench across multiple LLMs. STORM outperforms the git-worktree-based multi-agent baseline by +18.7 on Commit0-Lite and +1.4 on PaperBench, while achieving comparable or better cost efficiency. Combined with single-agent runs, STORM reaches highest scores of 87.6 and 78.2 on the two benchmarks respectively, suggesting that explicit state management is a more effective foundation for multi-agent collaboration than workspace isolation. STORM can also be plugged into any multi-agent system seamlessly.

AI agents often fail in ways that are difficult to localize because executions are probabilistic, long-horizon, multi-agent, and mediated by noisy tool outputs. We address this gap by manually annotating failed agent runs and release a novel benchmark of 115 failed trajectories spanning structured API workflows, incident management, and open-ended web/file tasks. Each trajectory is annotated with a critical failure step and a category from a grounded-theory derived, cross-domain failure taxonomy. To mitigate the human cost of failure attribution, we present AGENTRX, an automated domain-agnostic diagnostic framework that pinpoints the critical failure step in a failed agent trajectory. It synthesizes constraints, evaluates them step-by-step, and produces an auditable validation log of constraint violations with associated evidence; an LLM-based judge uses this log to localize the critical step and category. Our framework improves step localization and failure attribution over existing baselines across three domains.

Autonomous web agents need to operate under externally imposed or human-specified policies while generating long-horizon trajectories. However, little work has examined whether these trajectories comply with such policies, or whether policy violations persist across different contexts such as domains (e.g., shopping or coding websites) and subdomains (e.g., product search and order management in shopping). To address this gap, we introduce PolicyGuardBench, a benchmark of about 60k examples for detecting policy violations in agent trajectories. From diverse agent runs, we generate a broad set of policies and create both within subdomain and cross subdomain pairings with violation labels. In addition to full-trajectory evaluation, PolicyGuardBench also includes a prefix-based violation detection task where models must anticipate policy violations from truncated trajectory prefixes rather than complete sequences. Using this dataset, we train PolicyGuard-4B, a lightweight guardrail model that delivers strong detection accuracy across all tasks while keeping inference efficient. Notably, PolicyGuard-4B generalizes across domains and preserves high accuracy on unseen settings. Together, PolicyGuardBench and PolicyGuard-4B provide the first comprehensive framework for studying policy compliance in web agent trajectories, and show that accurate and generalizable guardrails are feasible at small scales.

Run the same LLM agent on the same task twice: do you get the same behavior? We find the answer is often no. In a study of 3,000 agent runs across three models (Llama 3.1 70B, GPT-4o, and Claude Sonnet 4.5) on HotpotQA, we observe that ReAct-style agents produce 2.0--4.2 distinct action sequences per 10 runs on average, even with identical inputs. More importantly, this variance predicts failure: tasks with consistent behavior (leq2 unique paths) achieve 80--92% accuracy, while highly inconsistent tasks (geq6 unique paths) achieve only 25--60%, a 32--55 percentage point gap depending on model. We trace variance to early decisions: 69% of divergence occurs at step 2, the first search query. Our results suggest that monitoring behavioral consistency during execution could enable early error detection and improve agent reliability.

Large language model (LLM)-based systems are increasingly deployed to conduct scientific research autonomously, yet whether their reasoning adheres to the epistemic norms that make scientific inquiry self-correcting is poorly understood. Here, we evaluate LLM-based scientific agents across eight domains, spanning workflow execution to hypothesis-driven inquiry, through more than 25,000 agent runs and two complementary lenses: (i) a systematic performance analysis that decomposes the contributions of the base model and the agent scaffold, and (ii) a behavioral analysis of the epistemological structure of agent reasoning. We observe that the base model is the primary determinant of both performance and behavior, accounting for 41.4% of explained variance versus 1.5% for the scaffold. Across all configurations, evidence is ignored in 68% of traces, refutation-driven belief revision occurs in 26%, and convergent multi-test evidence is rare. The same reasoning pattern appears whether the agent executes a computational workflow or conducts hypothesis-driven inquiry. They persist even when agents receive near-complete successful reasoning trajectories as context, and the resulting unreliability compounds across repeated trials in epistemically demanding domains. Thus, current LLM-based agents execute scientific workflows but do not exhibit the epistemic patterns that characterize scientific reasoning. Outcome-based evaluation cannot detect these failures, and scaffold engineering alone cannot repair them. Until reasoning itself becomes a training target, the scientific knowledge produced by such agents cannot be justified by the process that generated it.

Recent advances in prompt learning allow large language model agents to acquire task-relevant knowledge from inference-time context without parameter changes. For example, existing methods (like ACE or GEPA) can learn system prompts to improve accuracy based on previous agent runs. However, these methods primarily focus on single-agent or low-parallelism settings. This fundamentally limits their ability to efficiently learn from a large set of collected agentic traces. It would be efficient and beneficial to run prompt learning in parallel to accommodate the growing trend of learning from many agentic traces or parallel agent executions. Yet without a principled strategy for scaling, current methods suffer from quality degradation with high parallelism. To improve both the efficiency and quality of prompt learning, we propose Combee, a novel framework to scale parallel prompt learning for self-improving agents. Combee speeds up learning and enables running many agents in parallel while learning from their aggregate traces without quality degradation. To achieve this, Combee leverages parallel scans and employs an augmented shuffle mechanism; Combee also introduces a dynamic batch size controller to balance quality and delay. Evaluations on AppWorld, Terminal-Bench, Formula, and FiNER demonstrate that Combee achieves up to 17x speedup over previous methods with comparable or better accuracy and equivalent cost.

Large language model (LLM) agents are evolving from request-response assistants into long-running software actors: they maintain state across model calls, fork subtasks, wait for external events, request human authority, generate tools, and perform side effects that must be resumed and audited. This paper presents Agent libOS, a library-OS-inspired runtime substrate for LLM agents. Agent libOS runs above a conventional host operating system; it does not implement hardware drivers, kernel-mode isolation, or a POSIX-compatible operating system. Instead, it treats an agent as an AgentProcess: a schedulable execution subject with process identity, parent-child lineage, lifecycle state, a tool table derived from an AgentImage, typed Object Memory, explicit capabilities, human queues, checkpoints, events, and audit records. Its central design rule is tools are libc-like wrappers; runtime primitives are the authority boundary. Filesystem access, object access, sleeps, human approval, JIT tool registration, and external side effects are checked at primitive boundaries under explicit capabilities and policy. We describe the design, threat model, Python prototype, and safety-oriented evaluation. The current prototype implements async scheduling, namespace-local Object Memory, runtime-integrated human approval, one-shot permission grants, per-process working directories, shell and image-registration primitives, Deno/TypeScript JIT tools over a libOS syscall broker, filesystem/object bridge tools, an injectable Resource Provider Substrate, deterministic demos, real-model smoke scripts, and 123 regression tests at the time of writing. Rather than improving planner accuracy, Agent libOS demonstrates a runtime substrate in which long-running LLM agents can be scheduled, authorized, resumed, and audited without treating tool dispatch as the trust boundary.

Deep Research (DR) has emerged as a new agentic paradigm to tackle complex, open-ended research tasks, demanding systems that can iteratively frame problems, acquire evidence, verify sources, and synthesize long-form reports. In practice, however, current DR systems are constrained by four interrelated limitations: long-horizon planning over an underspecified scope, the bottleneck of decomposing and scheduling such tasks within a single agent, hallucination risk in long-form synthesis, and limited process auditability. This technical report presents DuMate-DeepResearch, a multi-agent DR framework built on the Qianfan Agent Foundry. The framework decouples the Agent Core, which handles task understanding, planning, and scheduling, from an extensible Tool Ecosystem for retrieval, evidence acquisition, and report rendering, making every intermediate decision and tool invocation explicitly traceable. Building on this infrastructure, DuMate-DeepResearch further introduces three mechanisms: (i) a graph-based dynamic planning strategy expands the research roadmap coarse-to-fine and continuously revises it through reflection, re-planning, backtracking, and parallel branching; (ii) a recursive two-level execution design delegates each complex search sub-task to an inner Search Agent that runs its own planning loop, isolating noisy retrieval and stabilizing long-horizon execution; (iii) a rubric-based test-time optimization mechanism dynamically generates task-specific quality criteria and uses them as live reasoning scaffolds for evidence-grounded synthesis and adaptive stopping. Across two deep research benchmarks, DuMate-DeepResearch establishes new state-of-the-art results: the best overall score (58.03%) on DeepResearch Bench, and the best overall score (61.95%) on DeepResearch Bench II while ranking first in information recall and analysis.

Contemporary large language model (LLM) agents are remarkably capable, but they still lack reliable safety controls and can produce unconstrained, unpredictable, and even actively harmful outputs. To address this, we introduce Reflection-Driven Control, a standardized and pluggable control module that can be seamlessly integrated into general agent architectures. Reflection-Driven Control elevates "self-reflection" from a post hoc patch into an explicit step in the agent's own reasoning process: during generation, the agent continuously runs an internal reflection loop that monitors and evaluates its own decision path. When potential risks are detected, the system retrieves relevant repair examples and secure coding guidelines from an evolving reflective memory, injecting these evidence-based constraints directly into subsequent reasoning steps. We instantiate Reflection-Driven Control in the setting of secure code generation and systematically evaluate it across eight classes of security-critical programming tasks. Empirical results show that Reflection-Driven Control substantially improves the security and policy compliance of generated code while largely preserving functional correctness, with minimal runtime and token overhead. Taken together, these findings indicate that Reflection-Driven Control is a practical path toward trustworthy AI coding agents: it enables designs that are simultaneously autonomous, safer by construction, and auditable.

As AI agents become widely deployed as online services, users often rely on an agent developer's claim about how safety is enforced, which introduces a threat where safety measures are falsely advertised. To address the threat, we propose proof-of-guardrail, a system that enables developers to provide cryptographic proof that a response is generated after a specific open-source guardrail. To generate proof, the developer runs the agent and guardrail inside a Trusted Execution Environment (TEE), which produces a TEE-signed attestation of guardrail code execution verifiable by any user offline. We implement proof-of-guardrail for OpenClaw agents and evaluate latency overhead and deployment cost. Proof-of-guardrail ensures integrity of guardrail execution while keeping the developer's agent private, but we also highlight a risk of deception about safety, for example, when malicious developers actively jailbreak the guardrail. Code and demo video: https://github.com/SaharaLabsAI/Verifiable-ClawGuard

In the realm of web agent research, achieving both generalization and accuracy remains a challenging problem. Due to high variance in website structure, existing approaches often fail. Moreover, existing fine-tuning and in-context learning techniques fail to generalize across multiple websites. We introduce Wilbur, an approach that uses a differentiable ranking model and a novel instruction synthesis technique to optimally populate a black-box large language model's prompt with task demonstrations from previous runs. To maximize end-to-end success rates, we also propose an intelligent backtracking mechanism that learns and recovers from its mistakes. Finally, we show that our ranking model can be trained on data from a generative auto-curriculum which samples representative goals from an LLM, runs the agent, and automatically evaluates it, with no manual annotation. Wilbur achieves state-of-the-art results on the WebVoyager benchmark, beating text-only models by 8% overall, and up to 36% on certain websites. On the same benchmark, Wilbur is within 5% of a strong multi-modal model despite only receiving textual inputs, and further analysis reveals a substantial number of failures are due to engineering challenges of operating the web.

The diverse agents in multi-agent perception systems may be from different companies. Each company might use the identical classic neural network architecture based encoder for feature extraction. However, the data source to train the various agents is independent and private in each company, leading to the Distribution Gap of different private data for training distinct agents in multi-agent perception system. The data silos by the above Distribution Gap could result in a significant performance decline in multi-agent perception. In this paper, we thoroughly examine the impact of the distribution gap on existing multi-agent perception systems. To break the data silos, we introduce the Feature Distribution-aware Aggregation (FDA) framework for cross-domain learning to mitigate the above Distribution Gap in multi-agent perception. FDA comprises two key components: Learnable Feature Compensation Module and Distribution-aware Statistical Consistency Module, both aimed at enhancing intermediate features to minimize the distribution gap among multi-agent features. Intensive experiments on the public OPV2V and V2XSet datasets underscore FDA's effectiveness in point cloud-based 3D object detection, presenting it as an invaluable augmentation to existing multi-agent perception systems.

Current LiDAR-based Vehicle-to-Everything (V2X) multi-agent perception systems have shown the significant success on 3D object detection. While these models perform well in the trained clean weather, they struggle in unseen adverse weather conditions with the domain gap. In this paper, we propose a Domain Generalization based approach, named V2X-DGW, for LiDAR-based 3D object detection on multi-agent perception system under adverse weather conditions. Our research aims to not only maintain favorable multi-agent performance in the clean weather but also promote the performance in the unseen adverse weather conditions by learning only on the clean weather data. To realize the Domain Generalization, we first introduce the Adaptive Weather Augmentation (AWA) to mimic the unseen adverse weather conditions, and then propose two alignments for generalizable representation learning: Trust-region Weather-invariant Alignment (TWA) and Agent-aware Contrastive Alignment (ACA). To evaluate this research, we add Fog, Rain, Snow conditions on two publicized multi-agent datasets based on physics-based models, resulting in two new datasets: OPV2V-w and V2XSet-w. Extensive experiments demonstrate that our V2X-DGW achieved significant improvements in the unseen adverse weathers. The code is available at https://github.com/Baolu1998/V2X-DGW.

Code agents are advancing rapidly, but debugging them is becoming increasingly difficult. As frameworks orchestrate parallel tool calls and multi-stage workflows over complex tasks, making the agent's state transitions and error propagation hard to observe. In these runs, an early misstep can trap the agent in unproductive loops or even cascade into fundamental errors, forming hidden error chains that make it hard to tell when the agent goes off track and why. Existing agent tracing analyses either focus on simple interaction or rely on small-scale manual inspection, which limits their scalability and usefulness for real coding workflows. We present CodeTracer, a tracing architecture that parses heterogeneous run artifacts through evolving extractors, reconstructs the full state transition history as a hierarchical trace tree with persistent memory, and performs failure onset localization to pinpoint the failure origin and its downstream chain. To enable systematic evaluation, we construct CodeTraceBench from a large collection of executed trajectories generated by four widely used code agent frameworks on diverse code tasks (e.g., bug fixing, refactoring, and terminal interaction), with supervision at both the stage and step levels for failure localization. Experiments show that CodeTracer substantially outperforms direct prompting and lightweight baselines, and that replaying its diagnostic signals consistently recovers originally failed runs under matched budgets. Our code and data are publicly available.

Multi-turn, multi-agent LLM game evaluations often exhibit substantial run-to-run variance. In long-horizon interactions, small early deviations compound across turns and are amplified by multi-agent coupling. This biases win rate estimates and makes rankings unreliable across repeated tournaments. Prompt choice worsens this further by producing different effective policies. We address both instability and underperformance with MEMO (Memory-augmented MOdel context optimization), a self-play framework that optimizes inference-time context by coupling retention and exploration. Retention maintains a persistent memory bank that stores structured insights from self-play trajectories and injects them as priors during later play. Exploration runs tournament-style prompt evolution with uncertainty-aware selection via TrueSkill, and uses prioritized replay to revisit rare and decisive states. Across five text-based games, MEMO raises mean win rate from 25.1% to 49.5% for GPT-4o-mini and from 20.9% to 44.3% for Qwen-2.5-7B-Instruct, using 2,000 self-play games per task. Run-to-run variance also drops, giving more stable rankings across prompt variations. These results suggest that multi-agent LLM game performance and robustness have substantial room for improvement through context optimization. MEMO achieves the largest gains in negotiation and imperfect-information games, while RL remains more effective in perfect-information settings.

Can generative agents be trusted in multimodal environments? Despite advances in large language and vision-language models that enable agents to act autonomously and pursue goals in rich settings, their ability to reason about safety, coherence, and trust across modalities remains limited. We introduce a reproducible simulation framework for evaluating agents along three dimensions: (1) safety improvement over time, including iterative plan revisions in text-visual scenarios; (2) detection of unsafe activities across multiple categories of social situations; and (3) social dynamics, measured as interaction counts and acceptance ratios of social exchanges. Agents are equipped with layered memory, dynamic planning, multimodal perception, and are instrumented with SocialMetrics, a suite of behavioral and structural metrics that quantifies plan revisions, unsafe-to-safe conversions, and information diffusion across networks. Experiments show that while agents can detect direct multimodal contradictions, they often fail to align local revisions with global safety, reaching only a 55 percent success rate in correcting unsafe plans. Across eight simulation runs with three models - Claude, GPT-4o mini, and Qwen-VL - five agents achieved average unsafe-to-safe conversion rates of 75, 55, and 58 percent, respectively. Overall performance ranged from 20 percent in multi-risk scenarios with GPT-4o mini to 98 percent in localized contexts such as fire/heat with Claude. Notably, 45 percent of unsafe actions were accepted when paired with misleading visuals, showing a strong tendency to overtrust images. These findings expose critical limitations in current architectures and provide a reproducible platform for studying multimodal safety, coherence, and social dynamics.

Training capable OS agents requires data that simultaneously captures structured user intents, multi-turn task delegation, and grounded tool execution--properties absent from existing datasets. We propose ISE (Intent -> Simulate -> Execute), a three-stage synthesis paradigm that addresses these gaps jointly. Stage 1 constructs roughly 50000 structured intents via a 4D framework (Persona x Domain x Task x Complexity); after deduplication the pool contains 43956 unique intents and attains a Vendi Score of 61.57 over the entire pool on mpnet-base-v2 embeddings (cosine kernel, q=1). Stage 2 drives multi-turn user-agent interaction through a role-locked user simulator that grounds each user turn in actual execution outcomes, producing 23132 complete trajectories averaging 8.12 user turns and 68.24 total dialogue turns. Stage 3 runs every tool call inside a live, isolated OS workspace, generating authentic failure-recovery dynamics instead of simulated responses. Fine-tuning on ISETrace improves ClawEval pass@1 from 19.3 to 37.7 using Qwen3-8B on agent tool-use tasks with a standard protocol. This result outperforms zero-shot GPT-4o and the larger Qwen3-32B base model which is four times bigger. An ablation on Stage 2 proves multi-turn simulation brings a large portion of the performance gain. We release all source code and dataset at https://github.com/Valiere01/ISE-Trace.

Auto-regressive LLM-based software engineering (SWE) agents, henceforth SWE agents, have made tremendous progress (>60% on SWE-Bench Verified) on real-world coding challenges including GitHub issue resolution. SWE agents use a combination of reasoning, environment interaction and self-reflection to resolve issues thereby generating "trajectories". Analysis of SWE agent trajectories is difficult, not only as they exceed LLM sequence length (sometimes, greater than 128k) but also because it involves a relatively prolonged interaction between an LLM and the environment managed by the agent. In case of an agent error, it can be hard to decipher, locate and understand its scope. Similarly, it can be hard to track improvements or regression over multiple runs or experiments. While a lot of research has gone into making these SWE agents reach state-of-the-art, much less focus has been put into creating tools to help analyze and visualize agent output. We propose a novel tool called SeaView: Software Engineering Agent Visual Interface for Enhanced Workflow, with a vision to assist SWE-agent researchers to visualize and inspect their experiments. SeaView's novel mechanisms help compare experimental runs with varying hyper-parameters or LLMs, and quickly get an understanding of LLM or environment related problems. Based on our user study, experienced researchers spend between 10 and 30 minutes to gather the information provided by SeaView, while researchers with little experience can spend between 30 minutes to 1 hour to diagnose their experiment.

We introduce Meta Agents Research Environments (ARE), a research platform for scalable creation of environments, integration of synthetic or real applications, and execution of agentic orchestrations. ARE provides simple abstractions to build complex and diverse environments, each with their own rules, tools, content, and verifiers, helping to bridge the gap between model development and real-world deployment. We also propose Gaia2, a benchmark built in ARE and designed to measure general agent capabilities. Beyond search and execution, Gaia2 requires agents to handle ambiguities and noise, adapt to dynamic environments, collaborate with other agents, and operate under temporal constraints. Unlike prior benchmarks, Gaia2 runs asynchronously, surfacing new failure modes that are invisible in static settings. Our experiments show that no system dominates across the intelligence spectrum: stronger reasoning often comes at the cost of efficiency, and budget scaling curves plateau, highlighting the need for new architectures and adaptive compute strategies. Perhaps more importantly, ARE abstractions enable continuous extension of Gaia2 to other environments, empowering the community to rapidly create new benchmarks tailored to their domains. In AI's second half, progress increasingly depends on defining meaningful tasks and robust evaluations to drive frontier capabilities forward.

Agent skills extend language-model agents with task-specific procedures, scripts, and references, but the tasks and environments they target continually change. Existing methods improve skills in bounded runs and retain only the final artifact, discarding the decision history that later agents need to interpret prior revisions, evaluations, and rejected alternatives. We introduce SkillHone, a harness for continual agent skill evolution grounded in persistent decision history. SkillHone pairs skill revisions with evaluation-side evidence that supplies practice feedback, recording structured histories of diagnoses, revisions, evidence, and outcomes. Role-separated subagents run candidate skills on practice probes with redacted reporting and propose revisions informed by prior decisions, enabling cross-session refinement without rediscovering past rationale. On deep-research benchmarks, SkillHone runs without a pre-integrated search stack and outperforms the commercially backed deep-research agent by 15.8 points on GAIA and 3.2 points on WebWalkerQA-EN, while also exceeding prior skill-evolution methods. We further deploy SkillHone on internal tool-mediated analysis scenarios, where it improves accuracy by an average of 18.8 points across seven settings.

Multi-agent reinforcement learning (MARL) is a powerful paradigm for solving cooperative and competitive decision-making problems. While many MARL benchmarks have been proposed, few combine continuous state and action spaces with challenging coordination and planning tasks. We introduce CAMAR, a new MARL benchmark designed explicitly for multi-agent pathfinding in environments with continuous actions. CAMAR supports cooperative and competitive interactions between agents and runs efficiently at up to 100,000 environment steps per second. We also propose a three-tier evaluation protocol to better track algorithmic progress and enable deeper analysis of performance. In addition, CAMAR allows the integration of classical planning methods such as RRT and RRT* into MARL pipelines. We use them as standalone baselines and combine RRT* with popular MARL algorithms to create hybrid approaches. We provide a suite of test scenarios and benchmarking tools to ensure reproducibility and fair comparison. Experiments show that CAMAR presents a challenging and realistic testbed for the MARL community.

Deploying LLM-powered agents in enterprise scenarios such as cloud technical support demands high-quality, domain-specific skills. However, existing skill creators lack domain grounding, producing skills poorly aligned with real-world task requirements. Moreover, once deployed, there is no systematic mechanism to trace execution failures back to skill deficiencies and drive targeted refinements, leaving skill quality stagnant despite accumulating operational evidence. We introduce SkillForge, a self-evolving framework that closes an end-to-end creation-evaluation-refinement loop. To produce well-aligned initial skills, a Domain-Contextualized Skill Creator grounds skill synthesis in knowledge bases and historical support tickets. To enable continuous self-optimization, a three-stage pipeline -- Failure Analyzer, Skill Diagnostician, and Skill Optimizer -- automatically diagnoses execution failures in batch, pinpoints the underlying skill deficiencies, and rewrites the skill to eliminate them. This cycle runs iteratively, allowing skills to self-improve with every round of deployment feedback. Evaluated on five real-world cloud support scenarios spanning 1,883 tickets and 3,737 tasks, experiments show that: (1) the Domain-Contextualized Skill Creator produces substantially better initial skills than the generic skill creator, as measured by consistency with expert-authored reference responses from historical tickets; and (2) the self-evolution loop progressively improves skill quality from diverse starting points (including expert-authored, domain-created, and generic skills) across successive rounds, demonstrating that automated evolution can surpass manually curated expert knowledge.

We measure how well current large language models coordinate as multiple agents sharing a common resource, using the dining philosophers problem as a clean test bed. Across 630 episodes spanning seven models and three philosopher counts, four frontier closed-source systems reach mean reward 0.45 to 0.87 and Mistral-Small 24B reaches 0.83 to 0.99, while Qwen3-14B reaches 0.13 to 0.35. We then ask whether group relative policy optimization (GRPO) on rollouts from the task itself can close the gap and find that it cannot: a Welch's t-test on per-episode reward at five philosophers gives p = 0.66 and a Hedges' g of -0.11, with no statistically significant change at ten or fifteen philosophers either. Two further observations qualify the result. The training reward of both 8B and 14B runs peaked at step nine and then declined, so the default saved checkpoint at step 15 is strictly worse than several earlier ones. The four-term reward we use admits a degenerate maximum at zero actions, which DeepSeek-R1-Distill-Qwen-7B and Mistral-Small 24B at five philosophers both inhabit, with mean reward 1.0 and 0.83 respectively at zero meals. The bottleneck for an open-weight 14B model on multi-agent coordination is not training compute but training methodology: reward shaping that does not collapse to a no-action maximum, checkpoint discipline that does not depend on the final step, and curriculum across problem scales.

Reasoning over heterogeneous artifacts (PDFs, spreadsheets, slide decks, etc.) increasingly occurs within structured agent workflows that iteratively extract, transform, and reference external information. In these workflows, uncertainty is not merely an input-quality issue: it can redirect decomposition and routing decisions, reshape intermediate state, and produce qualitatively different execution trajectories. We study this phenomenon by treating uncertainty as a controlled variable: we inject structured perturbations into artifact-derived representations, execute fixed workflows under comprehensive logging, and quantify contamination via trace divergence in plans, tool invocations, and intermediate state. Across 614 paired runs on 32 GAIA tasks with three different language models, we find a decoupling: workflows may diverge substantially yet recover correct answers, or remain structurally similar while producing incorrect outputs. We characterize three manifestation types: silent semantic corruption, behavioral detours with recovery, and combined structural disruption and their control-flow signatures (rerouting, extended execution, early termination). We measure operational costs and characterize why commonly used verification guardrails fail to intercept contamination. We contribute (i) a formal taxonomy of contamination manifestations in structured workflows, (ii) a trace-based measurement framework for detecting and localizing contamination across agent interactions, and (iii) empirical evidence with implications for targeted verification, defensive design, and cost control.

In multi-agent reinforcement learning (MARL), independent learning (IL) often shows remarkable performance and easily scales with the number of agents. Yet, using IL can be inefficient and runs the risk of failing to successfully train, particularly in scenarios that require agents to coordinate their actions. Using centralised learning (CL) enables MARL agents to quickly learn how to coordinate their behaviour but employing CL everywhere is often prohibitively expensive in real-world applications. Besides, using CL in value-based methods often needs strong representational constraints (e.g. individual-global-max condition) that can lead to poor performance if violated. In this paper, we introduce a novel plug & play IL framework named Multi-Agent Network Selection Algorithm (MANSA) which selectively employs CL only at states that require coordination. At its core, MANSA has an additional agent that uses switching controls to quickly learn the best states to activate CL during training, using CL only where necessary and vastly reducing the computational burden of CL. Our theory proves MANSA preserves cooperative MARL convergence properties, boosts IL performance and can optimally make use of a fixed budget on the number CL calls. We show empirically in Level-based Foraging (LBF) and StarCraft Multi-agent Challenge (SMAC) that MANSA achieves fast, superior and more reliable performance while making 40% fewer CL calls in SMAC and using CL at only 1% CL calls in LBF.

Recent vision-language models have strong perceptual ability but their implicit reasoning is hard to explain and easily generates hallucinations on complex queries. Compositional methods improve interpretability, but most rely on a single agent or hand-crafted pipeline and cannot decide when to collaborate across complementary agents or compete among overlapping ones. We introduce MATA (Multi-Agent hierarchical Trainable Automaton), a multi-agent system presented as a hierarchical finite-state automaton for visual reasoning whose top-level transitions are chosen by a trainable hyper agent. Each agent corresponds to a state in the hyper automaton, and runs a small rule-based sub-automaton for reliable micro-control. All agents read and write a shared memory, yielding transparent execution history. To supervise the hyper agent's transition policy, we build transition-trajectory trees and transform to memory-to-next-state pairs, forming the MATA-SFT-90K dataset for supervised finetuning (SFT). The finetuned LLM as the transition policy understands the query and the capacity of agents, and it can efficiently choose the optimal agent to solve the task. Across multiple visual reasoning benchmarks, MATA achieves the state-of-the-art results compared with monolithic and compositional baselines. The code and dataset are available at https://github.com/ControlNet/MATA.

Multi-agent LLM ensembles can converge on coordinated, socially harmful equilibria. This paper advances an experimental framework for evaluating Institutional AI, our system-level approach to AI alignment that reframes alignment from preference engineering in agent-space to mechanism design in institution-space. Central to this approach is the governance graph, a public, immutable manifest that declares legal states, transitions, sanctions, and restorative paths; an Oracle/Controller runtime interprets this manifest, attaching enforceable consequences to evidence of coordination while recording a cryptographically keyed, append-only governance log for audit and provenance. We apply the Institutional AI framework to govern the Cournot collusion case documented by prior work and compare three regimes: Ungoverned (baseline incentives from the structure of the Cournot market), Constitutional (a prompt-only policy-as-prompt prohibition implemented as a fixed written anti-collusion constitution, and Institutional (governance-graph-based). Across six model configurations including cross-provider pairs (N=90 runs/condition), the Institutional regime produces large reductions in collusion: mean tier falls from 3.1 to 1.8 (Cohen's d=1.28), and severe-collusion incidence drops from 50% to 5.6%. The prompt-only Constitutional baseline yields no reliable improvement, illustrating that declarative prohibitions do not bind under optimisation pressure. These results suggest that multi-agent alignment may benefit from being framed as an institutional design problem, where governance graphs can provide a tractable abstraction for alignment-relevant collective behavior.

AI agents have been developed for complex real-world tasks from coding to customer service. But AI agent evaluations suffer from many challenges that undermine our understanding of how well agents really work. We introduce the Holistic Agent Leaderboard (HAL) to address these challenges. We make three main contributions. First, we provide a standardized evaluation harness that orchestrates parallel evaluations across hundreds of VMs, reducing evaluation time from weeks to hours while eliminating common implementation bugs. Second, we conduct three-dimensional analysis spanning models, scaffolds, and benchmarks. We validate the harness by conducting 21,730 agent rollouts across 9 models and 9 benchmarks in coding, web navigation, science, and customer service with a total cost of about $40,000. Our analysis reveals surprising insights, such as higher reasoning effort reducing accuracy in the majority of runs. Third, we use LLM-aided log inspection to uncover previously unreported behaviors, such as searching for the benchmark on HuggingFace instead of solving a task, or misusing credit cards in flight booking tasks. We share all agent logs, comprising 2.5B tokens of language model calls, to incentivize further research into agent behavior. By standardizing how the field evaluates agents and addressing common pitfalls in agent evaluation, we hope to shift the focus from agents that ace benchmarks to agents that work reliably in the real world.

AI agents are increasingly deployed to execute important tasks. While rising accuracy scores on standard benchmarks suggest rapid progress, many agents still continue to fail in practice. This discrepancy highlights a fundamental limitation of current evaluations: compressing agent behavior into a single success metric obscures critical operational flaws. Notably, it ignores whether agents behave consistently across runs, withstand perturbations, fail predictably, or have bounded error severity. Grounded in safety-critical engineering, we provide a holistic performance profile by proposing twelve concrete metrics that decompose agent reliability along four key dimensions: consistency, robustness, predictability, and safety. Evaluating 14 agentic models across two complementary benchmarks, we find that recent capability gains have only yielded small improvements in reliability. By exposing these persistent limitations, our metrics complement traditional evaluations while offering tools for reasoning about how agents perform, degrade, and fail.

Fine-grained action recognition in egocentric video is challenging for Vision-Language Models (VLMs): actions often differ only in small visual cues, and a single model tends to be biased toward a subset of these cues. We propose Divide, Deliberate, Decide, a fully-local, zero-shot multi-agent framework in which (i) a VLM orchestrator chunks the video and proposes a top-k candidate label list per segment, (ii) an ensemble of heterogeneous VLM specialists, drawn from different open model families, engages in a structured deliberation that includes a peer-consultation round of questions, and (iii) agent rankings are aggregated with a Borda count and the orchestrator re-ranks its own prediction in light of the specialists' evidence. The entire pipeline runs locally with no fine-tuning. Experiments show that our method positively improves zero-shot action recognition performance over the baseline, highlighting the influence of a heterogeneous deliberation step, showing that the gain stems from decorrelated model priors rather than from additional compute.

AutoMegaKernel (AMK) compiles a HuggingFace Llama-family model into a single persistent cooperative CUDA kernel that runs the whole forward pass in one launch, with no per-model hand-written CUDA. The contribution is the system, not raw speed. A frozen schedule-IR validator statically certifies deadlock-freedom and race-freedom via static graph checks (not a mechanized proof), so an unsafe agent-proposed schedule is rejected before launch: across 7,160 adversarial schedules (6,091 unsafe) it had zero false-accepts and accepted all 360 real lowerings. The same source retargets sm_80/sm_90/sm_120 from one codebase, auto-generates correct megakernels for 10 of 10 supported models, and on a real SmolLM2-135M checkpoint reproduces HuggingFace greedy decode token-for-token (perplexity match 2.5e-7). An unattended, agent-drivable autoresearch loop self-improves the megakernel over its own baseline (1.25-1.72x). A search-found int8 (W8A16) megakernel beats CUDA-graphed cuBLAS bf16 at batch-1 decode across NVIDIA's datacenter inference fleet: L4 up to 1.33x, the current-gen L40S 1.25-1.27x, A10G up to 1.08x at scale, and the consumer RTX 5090 1.19-1.23x. The ordering is not a clean function of bandwidth (the 864 GB/s L40S beats the 600 GB/s A10G); the divide is inference-class vs training-class. AMK trails cuBLAS on the high-bandwidth training-class A100/H100, where the harness localizes the cross-SM-sync bottleneck; we report the gap plainly. This is a precision-asymmetric (W8A16 vs bf16) comparison at decode position 0; the largest real checkpoint is TinyLlama-1.1B. Code and the harness: https://github.com/RightNow-AI/AutoMegaKernel

Agent systems often decompose a task across multiple roles, but these roles are typically specified by prompts rather than enforced by access controls. Without enforcement, a team pass rate can mask whether agents actually coordinated or whether one role effectively did another role's work. We present TeamBench, a benchmark with 851 task templates and 931 seeded instances for evaluating agent coordination under operating system-enforced role separation. TeamBench separates specification access, workspace editing, and final certification across Planner, Executor, and Verifier roles, so that no role can read the full requirements, modify the workspace, and certify the final answer. Prompt-only and sandbox-enforced teams reach statistically indistinguishable pass rates, but prompt-only runs produce 3.6 times more cases where the verifier attempts to edit the executor's code. Verifiers approve 49% of submissions that fail the deterministic grader, and removing the verifier improves mean partial score in the ablation. Team value is also conditional. Teams benefit when single agents struggle, but hurt when single agents already perform well. A 40-session human study under the same role separation shows that our benchmark exposes interaction patterns that pass rate misses. Solo participants work through the task directly, human participants paired with agents often collapse into quick approval, and human teams spend more effort coordinating missing information across roles.

Multi-agent large language model (LLM) and vision-language model (VLM) debate systems employ specialized roles for complex problem-solving, yet model specializations are not leveraged to decide which model should fill which role. We propose dynamic role assignment, a framework that runs a Meta-Debate to select suitable agents before the actual debate. The meta-debate has two stages: (1) proposal, where candidates provide role-tailored arguments, and (2) peer review, where proposals are scored with data and role-specific criteria to choose the best agent for each position. We evaluate our method on LLM problem solving benchmarks. Applied on top of existing debate systems, our approach consistently outperforms uniform assignments (filling all roles with the same model) by up to 74.8% and random assignments (assigning models to roles without considering their suitability) by up to 29.7%, depending on the task and the specific assignment. This work establishes a new paradigm for multi-agent system design, shifting from static agent deployment to dynamic and capability-aware selection.

Evaluating the general abilities of intelligent agents requires complex simulation environments. Existing benchmarks typically evaluate only one narrow task per environment, requiring researchers to perform expensive training runs on many different environments. We introduce Crafter, an open world survival game with visual inputs that evaluates a wide range of general abilities within a single environment. Agents either learn from the provided reward signal or through intrinsic objectives and are evaluated by semantically meaningful achievements that can be unlocked during each episode, such as discovering resources and crafting tools. Consistently unlocking all achievements requires strong generalization, deep exploration, and long-term reasoning. We experimentally verify that Crafter is of appropriate difficulty to drive future research and provide baselines scores of reward agents and unsupervised agents. Furthermore, we observe sophisticated behaviors emerging from maximizing the reward signal, such as building tunnel systems, bridges, houses, and plantations. We hope that Crafter will accelerate research progress by quickly evaluating a wide spectrum of abilities.

We present Test-Driven AI Agent Definition (TDAD), a methodology that treats agent prompts as compiled artifacts: engineers provide behavioral specifications, a coding agent converts them into executable tests, and a second coding agent iteratively refines the prompt until tests pass. Deploying tool-using LLM agents in production requires measurable behavioral compliance that current development practices cannot provide. Small prompt changes cause silent regressions, tool misuse goes undetected, and policy violations emerge only after deployment. To mitigate specification gaming, TDAD introduces three mechanisms: (1) visible/hidden test splits that withhold evaluation tests during compilation, (2) semantic mutation testing via a post-compilation agent that generates plausible faulty prompt variants, with the harness measuring whether the test suite detects them, and (3) spec evolution scenarios that quantify regression safety when requirements change. We evaluate TDAD on SpecSuite-Core, a benchmark of four deeply-specified agents spanning policy compliance, grounded analytics, runbook adherence, and deterministic enforcement. Across 24 independent trials, TDAD achieves 92% v1 compilation success with 97% mean hidden pass rate; evolved specifications compile at 58%, with most failed runs passing all visible tests except 1-2, and show 86-100% mutation scores, 78% v2 hidden pass rate, and 97% regression safety scores. The implementation is available as an open benchmark at https://github.com/f-labs-io/tdad-paper-code.

Frontier coding agents are increasingly used in workflows where users supervise progress primarily through repeated improvement of a public score, namely the reported score on a public evaluation file with labels in the workspace, rather than through direct inspection of the agent's intermediate outputs. We study whether multi-round user pressure to improve that score induces public score exploitation: behavior that raises the public score through shortcuts without improving hidden private evaluation. We begin with a preliminary single-script tabular classification task, where GPT-5.4 and Claude Opus 4.6 both exploit label information within 10 rounds of user-agent interaction. We then build AgentPressureBench, a 34-task machine-learning repository benchmark spanning three input modalities, and collect 1326 multi-round trajectories from 13 coding agents. On our benchmark, we observe 403 exploitative runs, spanning across all tasks. We also find that stronger models have higher exploitation rates, supported by a significant Spearman rank correlation of 0.77. Our ablation experiments show that higher user pressure leads to earlier exploitation, reducing the average first exploit round by 15.6 rounds (i.e., 19.67 to 4.08). As a mitigation, adding explicit anti-exploit wordings in prompt mostly eliminates exploitation (100% to 8.3%). We hope that our work can bring attention to more careful use of coding agents workflow, and developing more robust coding agents under user pressure. Our project page is at https://ucsc-vlaa.github.io/AgentPressureBench .

The increasing complexity and interconnectivity of digital infrastructures make scalable and reliable security assessment methods essential. Robotic systems represent a particularly important class of operational technology, as modern robots are highly networked cyber-physical systems deployed in domains such as industrial automation, logistics, and autonomous services. This paper explores the use of large language models for automated penetration testing in robotic environments. We propose an environment-grounded multi-agent architecture tailored to Robotics-based systems. The approach dynamically constructs a shared graph-based memory during execution that captures the observable system state, including network topology, communication channels, vulnerabilities, and attempted exploits. This enables structured automation while maintaining traceability and effective context management throughout the testing process. Evaluated across multiple iterations within a specialized robotics Capture-the-Flag scenario (ROS/ROS2), the system demonstrated high reliability, successfully completing the challenge in 100\% of test runs (n=5). This performance significantly exceeds literature benchmarks while maintaining the traceability and human oversight required by frameworks like the EU AI Act.

Modern agents built on frontier language models often cannot adapt their weights. What, then, remains trainable? We argue it is the agent's workspace, the structured external substrate it reads, writes, and tests; we call its evolution workspace optimization. Workspace optimization targets hard multi-turn environments where a frontier model has strong priors but cannot solve the task in a single shot, so the agent must learn through interaction. We propose a principled way to evolve the workspace, mirroring the structure of weight-space training: artifacts in place of parameters, evidence in place of data, counterexamples in place of losses, and textual feedback in place of gradients. We instantiate the idea in DreamTeam, a multi-agent harness for ARC-AGI-3 whose roles build an executable world model, plan, hypothesize, probe, strategize, and route failures. On the current 25-game ARC-AGI-3 public set under the official scoring protocol and averaged over two independent runs, DreamTeam improves the SOTA protocol-matched agent's score from 36% to 38.4%, while using 31% fewer environment actions per game.

We present MAESTRO, an evaluation suite for the testing, reliability, and observability of LLM-based MAS. MAESTRO standardizes MAS configuration and execution through a unified interface, supports integrating both native and third-party MAS via a repository of examples and lightweight adapters, and exports framework-agnostic execution traces together with system-level signals (e.g., latency, cost, and failures). We instantiate MAESTRO with 12 representative MAS spanning popular agentic frameworks and interaction patterns, and conduct controlled experiments across repeated runs, backend models, and tool configurations. Our case studies show that MAS executions can be structurally stable yet temporally variable, leading to substantial run-to-run variance in performance and reliability. We further find that MAS architecture is the dominant driver of resource profiles, reproducibility, and cost-latency-accuracy trade-off, often outweighing changes in backend models or tool settings. Overall, MAESTRO enables systematic evaluation and provides empirical guidance for designing and optimizing agentic systems.

As LLM-based agents are deployed in production systems, understanding their behavioral consistency (whether they produce similar action sequences when given identical tasks) becomes critical for reliability. We study consistency in the context of SWE-bench, a challenging software engineering benchmark requiring complex, multi-step reasoning. Comparing Claude~4.5~Sonnet, GPT-5, and Llama-3.1-70B across 50 runs each (10 tasks times 5 runs), we find that across models, higher consistency aligns with higher accuracy: Claude achieves the lowest variance (CV: 15.2\%) and highest accuracy (58\%), GPT-5 is intermediate (CV: 32.2\%, accuracy: 32\%), and Llama shows the highest variance (CV: 47.0\%) with lowest accuracy (4\%). However, within a model, consistency can amplify both correct and incorrect interpretations. Our analysis reveals a critical nuance: consistency amplifies outcomes rather than guaranteeing correctness. 71\% of Claude's failures stem from "consistent wrong interpretation": making the same incorrect assumption across all runs. Interestingly, GPT-5 achieves similar early strategic agreement as Claude (diverging at step 3.4 vs.\ 3.2) but exhibits 2.1times higher variance, suggesting that divergence timing alone does not determine consistency. These findings suggest that for production deployment, interpretation accuracy matters more than execution consistency, with implications for agent evaluation and training.

Long-lived AI agents are increasingly deployed as persistent operational systems, yet they are still evaluated like freshly initialized models. Day-one benchmarks miss a basic systems question: how long does an agent remain reliable after deployment? Even when model weights are frozen, an agent's effective state keeps changing as it compresses interaction history, retrieves from a growing memory store, revises facts after updates, and undergoes routine maintenance. Reliability therefore becomes a lifespan property of the full agent harness, not only a snapshot property of the base model. We introduce AgingBench, a longitudinal reliability benchmark for agent lifespan engineering: measuring not only whether deployed agents degrade, but what form the degradation takes and where repair should target. AgingBench organizes agent aging into four mechanisms: compression aging, interference aging, revision aging, and maintenance aging. To diagnose these failures, AgingBench uses temporal dependency graphs and paired counterfactual probes that produce diagnostic profiles for the write, retrieval, and utilization stages of the memory pipeline. Across 7 scenarios, 14 models, multiple memory policies, and both runner-controlled and autonomous agents, over ~400 runs spanning 8 - 200 sessions show that agent aging is not one-dimensional: behavioral tests can remain clean while factual precision decays; derived-state tracking can collapse sharply within a single model; and the same wrong answer can require different repairs depending on what the diagnostic profile points to. These results suggest that reliable agent deployment requires lifespan evaluation, mechanism-level diagnosis, and stage-targeted repair, not only stronger day-one models.

Large Language Models (LLMs) have shown remarkable potential across various domains, including cybersecurity. Using commercial cloud-based LLMs may be undesirable due to privacy concerns, costs, and network connectivity constraints. In this paper, we present Hackphyr, a locally fine-tuned LLM to be used as a red-team agent within network security environments. Our fine-tuned 7 billion parameter model can run on a single GPU card and achieves performance comparable with much larger and more powerful commercial models such as GPT-4. Hackphyr clearly outperforms other models, including GPT-3.5-turbo, and baselines, such as Q-learning agents in complex, previously unseen scenarios. To achieve this performance, we generated a new task-specific cybersecurity dataset to enhance the base model's capabilities. Finally, we conducted a comprehensive analysis of the agents' behaviors that provides insights into the planning abilities and potential shortcomings of such agents, contributing to the broader understanding of LLM-based agents in cybersecurity contexts

We demonstrate how collective memory emerges in decentralized multi-agent systems through the interplay between individual agent memory and environmental trace communication. Our agents maintain internal memory states while depositing persistent environmental traces, creating a spatially distributed collective memory without centralized control. Comprehensive validation across five environmental conditions (20x20 to 50x50 grids, 5-20 agents, 50 runs per configuration) reveals a critical asymmetry: individual memory alone provides 68.7% performance improvement over no-memory baselines (1563.87 vs 927.23, p < 0.001), while environmental traces without memory fail completely. This demonstrates that memory functions independently but traces require cognitive infrastructure for interpretation. Systematic density-sweep experiments (rho in [0.049, 0.300], up to 625 agents) validate our theoretical phase transition prediction. On realistic large grids (30x30, 50x50), stigmergic coordination dominates above rho ~ 0.20, with traces outperforming memory by 36-41% on composite metrics despite lower food efficiency. The experimental crossover confirms the predicted critical density rho_c = 0.230 within 13% error.

As AI systems increasingly assume roles where trust and alignment with human values are essential, understanding when and why they engage in deception has become a critical research priority. We introduce The Traitors, a multi-agent simulation framework inspired by social deduction games, designed to probe deception, trust formation, and strategic communication among large language model (LLM) agents under asymmetric information. A minority of agents the traitors seek to mislead the majority, while the faithful must infer hidden identities through dialogue and reasoning. Our contributions are: (1) we ground the environment in formal frameworks from game theory, behavioral economics, and social cognition; (2) we develop a suite of evaluation metrics capturing deception success, trust dynamics, and collective inference quality; (3) we implement a fully autonomous simulation platform where LLMs reason over persistent memory and evolving social dynamics, with support for heterogeneous agent populations, specialized traits, and adaptive behaviors. Our initial experiments across DeepSeek-V3, GPT-4o-mini, and GPT-4o (10 runs per model) reveal a notable asymmetry: advanced models like GPT-4o demonstrate superior deceptive capabilities yet exhibit disproportionate vulnerability to others' falsehoods. This suggests deception skills may scale faster than detection abilities. Overall, The Traitors provides a focused, configurable testbed for investigating LLM behavior in socially nuanced interactions. We position this work as a contribution toward more rigorous research on deception mechanisms, alignment challenges, and the broader social reliability of AI systems.

ResearchPilot is an open-source, self-hostable multi-agent system for literature-review assistance. Given a natural-language research question, it retrieves papers from Semantic Scholar and arXiv, extracts structured findings from paper abstracts, synthesizes cross-paper patterns, and drafts a citation-aware related-work section. The system combines FastAPI, Next.js, DSPy, SQLite, and Qdrant in a local-first architecture that supports bring-your-own-key model access and remote-or-local embeddings. This paper describes the system design, typed agent interfaces, persistence and history-search mechanisms, and the engineering tradeoffs involved in building a transparent research assistant. Rather than claiming algorithmic novelty, we present ResearchPilot as a systems contribution and evaluate it through automated tests and end-to-end local runs. We discuss limitations including external API rate limits, abstract-only extraction, incomplete corpus coverage, and the lack of citation verification.

Large language and vision-language models increasingly power agents that act on a user's behalf through command-line interface (CLI) harnesses. However, most agent benchmarks still rely on synthetic sandboxes, short-horizon tasks, mock-service APIs, and final-answer checks, leaving open whether agents can complete realistic long-horizon work in the runtimes where they are deployed. This work presents WildClawBench, a native-runtime benchmark of 60 human-authored, bilingual, multimodal tasks spanning six thematic categories. Each task averages roughly 8 minutes of wall-clock time and over 20 tool calls, and runs inside a reproducible Docker container hosting an actual CLI agent harness (OpenClaw, Claude Code, Codex, or Hermes Agent) with access to real tools rather than mock services. Grading is hybrid, combining deterministic rule-based checks, environment-state auditing of side effects, and an LLM/VLM judge for semantic verification. Across 19 frontier models, the best, Claude Opus 4.7, reaches only 62.2% overall under OpenClaw, while every other model stays below 60%, and switching harness alone shifts a single model by up to 18 points. These results show that long-horizon, native-runtime agent evaluation remains a far-from-resolved task for current frontier models. We release the tasks, code, and containerized tooling to support reproducible evaluation.

Claude Code is an agentic coding tool that can run shell commands, edit files, and call external services on behalf of the user. This study describes its comprehensive architecture by analyzing the publicly available TypeScript source code and further comparing it with OpenClaw, an independent open-source AI agent system that answers many of the same design questions from a different deployment context. Our analysis identifies five human values, philosophies, and needs that motivate the architecture (human decision authority, safety and security, reliable execution, capability amplification, and contextual adaptability) and traces them through thirteen design principles to specific implementation choices. The core of the system is a simple while-loop that calls the model, runs tools, and repeats. Most of the code, however, lives in the systems around this loop: a permission system with seven modes and an ML-based classifier, a five-layer compaction pipeline for context management, four extensibility mechanisms (MCP, plugins, skills, and hooks), a subagent delegation mechanism with worktree isolation, and append-oriented session storage. A comparison with OpenClaw, a multi-channel personal assistant gateway, shows that the same recurring design questions produce different architectural answers when the deployment context changes: from per-action safety classification to perimeter-level access control, from a single CLI loop to an embedded runtime within a gateway control plane, and from context-window extensions to gateway-wide capability registration. We finally identify six open design directions for future agent systems, grounded in recent empirical, architectural, and policy literature.

Large language models are increasingly used as agents in social, economic, and policy simulations. A common assumption is that stronger reasoning should improve simulation fidelity. We argue that this assumption can fail when the objective is not to solve a strategic problem, but to sample plausible boundedly rational behavior. In such settings, reasoning-enhanced models can become better solvers and worse simulators: they can over-optimize for strategically dominant actions, collapse compromise-oriented terminal behavior, and sometimes exhibit a diversity-without-fidelity pattern in which local variation survives without outcome-level fidelity. We study this solver-sampler mismatch in three multi-agent negotiation environments adapted from earlier simulation work: an ambiguous fragmented-authority trading-limits scenario, an ambiguous unified-opposition trading-limits scenario, and a new-domain grid-curtailment case in emergency electricity management. We compare three reflection conditions, no reflection, bounded reflection, and native reasoning, across two primary model families and then extend the same protocol to direct OpenAI runs with GPT-4.1 and GPT-5.2. Across all three experiments, bounded reflection produces substantially more diverse and compromise-oriented trajectories than either no reflection or native reasoning. In the direct OpenAI extension, GPT-5.2 native ends in authority decisions in 45 of 45 runs across the three experiments, while GPT-5.2 bounded recovers compromise outcomes in every environment. The contribution is not a claim that reasoning is generally harmful. It is a methodological warning: model capability and simulation fidelity are different objectives, and behavioral simulation should qualify models as samplers, not only as solvers.

When multiple large language models interact in a shared conversation, do they develop differentiated social roles or converge toward uniform behavior? We present a controlled experimental platform that orchestrates simultaneous multi-agent discussions among 7 heterogeneous LLMs on a unified inference backend, systematically varying group composition, naming conventions, and prompt structure across 12 experimental series (208 runs, 13,786 coded messages). Each message is independently coded on six behavioral flags by two LLM judges from distinct model families (Gemini 3.1 Pro and Claude Sonnet 4.6), achieving mean Cohen's kappa = 0.78 with conservative intersection-based adjudication. Human validation on 609 randomly stratified messages confirmed coding reliability (mean kappa = 0.73 vs. Gemini). We find that (1) heterogeneous groups exhibit significantly richer behavioral differentiation than homogeneous groups (cosine similarity 0.56 vs. 0.85; p < 10^-5, r = 0.70); (2) groups spontaneously exhibit compensatory response patterns when an agent crashes; (3) revealing real model names significantly increases behavioral convergence (cosine 0.56 to 0.77, p = 0.001); and (4) removing all prompt scaffolding converges profiles to homogeneous-level similarity (p < 0.001). Critically, these behaviors are absent when agents operate in isolation, confirming that behavioral diversity is a structured, reproducible phenomenon driven by the interaction of architectural heterogeneity, group context, and prompt-level scaffolding.

We present PDE-Agents, a multi-agent ecosystem that automates the full lifecycle of partial differential equation (PDE) / finite element method (FEM) simulations through natural-language interaction. Three specialist large language model (LLM) agents (Simulation, Analytics, Database) are orchestrated via a LangGraph supervisor, with a local open-source LLM stack (Qwen3-Coder-Next, Llama 4 Scout) on dual NVIDIA RTX PRO 6000 GPUs. The architecture is model-agnostic, validated across two LLM generations. A GraphRAG knowledge base (Neo4j, 768-d vector embeddings) encodes curated material properties, known failure patterns, and prior run lineage. We report seven contributions: (i) a verification and validation (V&V) study confirming second-order spatial convergence (O(h^2)) on the heat-equation solver; (ii) a three-way ablation over 50 tasks with a frozen KG (KG On, KG Off, KG Smart), where KG Smart reaches 100% success and the highest output quality (physics 0.933 vs. 0.853 for KG Off; MPF 0.926 vs. 0.796); (iii) a novel-material experiment with three fictional materials known only to the KG, where KG Smart attains near-perfect material property fidelity (MPF = 1.00) versus 0.34 for the KG-free baseline; (iv) a failure analysis tracing KG On's three failures to budget exhaustion and timeout, establishing warm-start injection as the dominant reliability factor; (v) an adaptive framework selecting the optimal retrieval mode per task; (vi) production metrics from 1,369 runs (97.8% success, 57.6% first-try); and (vii) a 100-task KG growth experiment showing a difficulty-dependent gain, with hard-task MPF improving 8.8% while easy/novel tasks stay at ceiling. All code, models, and evaluation artifacts are released openly. Our findings show that integration pattern, not knowledge content, determines whether GraphRAG augmentation helps or hinders LLM agents.

We present Claw AI Lab, a lab-native autonomous research platform that advances automated research from a hidden prompt-to-paper pipeline into an interactive AI laboratory. Rather than centering the system around a single agent or a fixed serial workflow, we allow users to instantiate a full research team from one prompt, with customizable roles, collaborative workflows, real-time monitoring, artifact inspection, and rollback/resume control through a unified dashboard. The platform also supports distinct research modes for exploration, multi-agent discussion, and reproduction, making autonomous research substantially more steerable and laboratory-like in practice. A key practical contribution of Claw AI Lab lies in its Claw-Code Harness, which connects local codebases, datasets, and checkpoints to runnable experiments and feeds execution artifacts back into the research loop. As a result, the harness improves not only execution integration, but also experimental completion and result integrity: experiments are easier to inspect, iterate on, and faithfully transfer into final papers, reducing common failure modes such as partial runs and malformed result reporting. In our internal evaluation on five AI research case studies, using AutoResearchClaw as the baseline, Claw AI Lab is consistently preferred by AI expert judges on idea novelty, experiment completeness, and paper presentation quality. We view Claw AI Lab as an early step toward a new paradigm: autonomous research as usable, interactive, and reliability-aware scientific infrastructure.

AI coding agents operate in a paradox: they possess vast parametric knowledge yet cannot remember a conversation from an hour ago. Existing memory systems store text in vector databases with single-channel retrieval, require cloud LLMs for core operations, and implement none of the cognitive processes that make human memory effective. We present SuperLocalMemory V3.3 ("The Living Brain"), a local-first agent memory system implementing the full cognitive memory taxonomy with mathematical lifecycle dynamics. Building on the information-geometric foundations of V3.2 (arXiv:2603.14588), we introduce five contributions: (1) Fisher-Rao Quantization-Aware Distance (FRQAD) -- a new metric on the Gaussian statistical manifold achieving 100% precision at preferring high-fidelity embeddings over quantized ones (vs 85.6% for cosine), with zero prior art; (2) Ebbinghaus Adaptive Forgetting with lifecycle-aware quantization -- the first mathematical forgetting curve in local agent memory coupled to progressive embedding compression, achieving 6.7x discriminative power; (3) 7-channel cognitive retrieval spanning semantic, keyword, entity graph, temporal, spreading activation, consolidation, and Hopfield associative channels, achieving 70.4% on LoCoMo in zero-LLM Mode A; (4) memory parameterization implementing Long-Term Implicit memory via soft prompts; (5) zero-friction auto-cognitive pipeline automating the complete memory lifecycle. On LoCoMo, V3.3 achieves 70.4% in Mode A (zero-LLM), with +23.8pp on multi-hop and +12.7pp on adversarial. V3.2 achieved 74.8% Mode A and 87.7% Mode C; the 4.4pp gap reflects a deliberate architectural trade-off. SLM V3.3 is open source under the Elastic License 2.0, runs entirely on CPU, with over 5,000 monthly downloads.

The quadratic complexity and indefinitely growing key-value (KV) cache of standard Transformers pose a major barrier to long-context processing. To overcome this, we introduce the Collaborative Memory Transformer (CoMeT), a novel architecture that enables LLMs to handle arbitrarily long sequences with constant memory usage and linear time complexity. Designed as an efficient, plug-in module, CoMeT can be integrated into pre-trained models with only minimal fine-tuning. It operates on sequential data chunks, using a dual-memory system to manage context: a temporary memory on a FIFO queue for recent events, and a global memory with a gated update rule for long-range dependencies. These memories then act as a dynamic soft prompt for the next chunk. To enable efficient fine-tuning on extremely long contexts, we introduce a novel layer-level pipeline parallelism strategy. The effectiveness of our approach is remarkable: a model equipped with CoMeT and fine-tuned on 32k contexts can accurately retrieve a passkey from any position within a 1M token sequence. On the SCROLLS benchmark, CoMeT surpasses other efficient methods and achieves performance comparable to a full-attention baseline on summarization tasks. Its practical effectiveness is further validated on real-world agent and user behavior QA tasks. The code is available at: https://anonymous.4open.science/r/comet-B00B/

Bird's eye view (BEV) semantic segmentation plays a crucial role in spatial sensing for autonomous driving. Although recent literature has made significant progress on BEV map understanding, they are all based on single-agent camera-based systems. These solutions sometimes have difficulty handling occlusions or detecting distant objects in complex traffic scenes. Vehicle-to-Vehicle (V2V) communication technologies have enabled autonomous vehicles to share sensing information, dramatically improving the perception performance and range compared to single-agent systems. In this paper, we propose CoBEVT, the first generic multi-agent multi-camera perception framework that can cooperatively generate BEV map predictions. To efficiently fuse camera features from multi-view and multi-agent data in an underlying Transformer architecture, we design a fused axial attention module (FAX), which captures sparsely local and global spatial interactions across views and agents. The extensive experiments on the V2V perception dataset, OPV2V, demonstrate that CoBEVT achieves state-of-the-art performance for cooperative BEV semantic segmentation. Moreover, CoBEVT is shown to be generalizable to other tasks, including 1) BEV segmentation with single-agent multi-camera and 2) 3D object detection with multi-agent LiDAR systems, achieving state-of-the-art performance with real-time inference speed. The code is available at https://github.com/DerrickXuNu/CoBEVT.

Recent advancements in Vehicle-to-Everything (V2X) technologies have enabled autonomous vehicles to share sensing information to see through occlusions, greatly boosting the perception capability. However, there are no real-world datasets to facilitate the real V2X cooperative perception research -- existing datasets either only support Vehicle-to-Infrastructure cooperation or Vehicle-to-Vehicle cooperation. In this paper, we present V2X-Real, a large-scale dataset that includes a mixture of multiple vehicles and smart infrastructure to facilitate the V2X cooperative perception development with multi-modality sensing data. Our V2X-Real is collected using two connected automated vehicles and two smart infrastructure, which are all equipped with multi-modal sensors including LiDAR sensors and multi-view cameras. The whole dataset contains 33K LiDAR frames and 171K camera data with over 1.2M annotated bounding boxes of 10 categories in very challenging urban scenarios. According to the collaboration mode and ego perspective, we derive four types of datasets for Vehicle-Centric, Infrastructure-Centric, Vehicle-to-Vehicle, and Infrastructure-to-Infrastructure cooperative perception. Comprehensive multi-class multi-agent benchmarks of SOTA cooperative perception methods are provided. The V2X-Real dataset and codebase are available at https://mobility-lab.seas.ucla.edu/v2x-real.

Vehicle-to-Everything (V2X) collaborative perception is crucial for autonomous driving. However, achieving high-precision V2X perception requires a significant amount of annotated real-world data, which can always be expensive and hard to acquire. Simulated data have raised much attention since they can be massively produced at an extremely low cost. Nevertheless, the significant domain gap between simulated and real-world data, including differences in sensor type, reflectance patterns, and road surroundings, often leads to poor performance of models trained on simulated data when evaluated on real-world data. In addition, there remains a domain gap between real-world collaborative agents, e.g. different types of sensors may be installed on autonomous vehicles and roadside infrastructures with different extrinsics, further increasing the difficulty of sim2real generalization. To take full advantage of simulated data, we present a new unsupervised sim2real domain adaptation method for V2X collaborative detection named Decoupled Unsupervised Sim2Real Adaptation (DUSA). Our new method decouples the V2X collaborative sim2real domain adaptation problem into two sub-problems: sim2real adaptation and inter-agent adaptation. For sim2real adaptation, we design a Location-adaptive Sim2Real Adapter (LSA) module to adaptively aggregate features from critical locations of the feature map and align the features between simulated data and real-world data via a sim/real discriminator on the aggregated global feature. For inter-agent adaptation, we further devise a Confidence-aware Inter-agent Adapter (CIA) module to align the fine-grained features from heterogeneous agents under the guidance of agent-wise confidence maps. Experiments demonstrate the effectiveness of the proposed DUSA approach on unsupervised sim2real adaptation from the simulated V2XSet dataset to the real-world DAIR-V2X-C dataset.

Vehicle-to-Vehicle technologies have enabled autonomous vehicles to share information to see through occlusions, greatly enhancing perception performance. Nevertheless, existing works all focused on homogeneous traffic where vehicles are equipped with the same type of sensors, which significantly hampers the scale of collaboration and benefit of cross-modality interactions. In this paper, we investigate the multi-agent hetero-modal cooperative perception problem where agents may have distinct sensor modalities. We present HM-ViT, the first unified multi-agent hetero-modal cooperative perception framework that can collaboratively predict 3D objects for highly dynamic vehicle-to-vehicle (V2V) collaborations with varying numbers and types of agents. To effectively fuse features from multi-view images and LiDAR point clouds, we design a novel heterogeneous 3D graph transformer to jointly reason inter-agent and intra-agent interactions. The extensive experiments on the V2V perception dataset OPV2V demonstrate that the HM-ViT outperforms SOTA cooperative perception methods for V2V hetero-modal cooperative perception. We will release codes to facilitate future research.

The learning from practice paradigm is crucial for developing capable Agentic AI systems, yet it is severely hampered by inefficient experience generation, a bottleneck especially pronounced in complex benchmarks like GAIA. To address this, we introduce AWorld, an open-source system engineered for large-scale agent-environment interaction. By distributing tasks across a cluster, AWorld accelerates experience collection by 14.6x compared to standard single-node, sequential execution. This critical speedup makes extensive reinforcement learning practical and scalable. Leveraging this capability, we trained a Qwen3-32B-based agent that significantly outperforms its base model, increasing its overall GAIA accuracy from 21.59% to 32.23%. On the benchmark's most challenging levels, our agent achieves a score of 16.33%, surpassing the performance of leading proprietary models. Our open-source system and resulting agent provide a practical blueprint for a complete agentic AI training pipeline, from efficient interaction to demonstrable model improvement.

Real-world robots localize objects from natural-language instructions while scenes around them keep changing. Yet most of the existing 3D visual grounding (3DVG) method still assumes a reconstructed and up-to-date point cloud, an assumption that forces costly re-scans and hinders deployment. We argue that 3DVG should be formulated as an active, memory-driven problem, and we introduce ChangingGrounding, the first benchmark that explicitly measures how well an agent can exploit past observations, explore only where needed, and still deliver precise 3D boxes in changing scenes. To set a strong reference point, we also propose Mem-ChangingGrounder, a zero-shot method for this task that marries cross-modal retrieval with lightweight multi-view fusion: it identifies the object type implied by the query, retrieves relevant memories to guide actions, then explores the target efficiently in the scene, falls back when previous operations are invalid, performs multi-view scanning of the target, and projects the fused evidence from multi-view scans to get accurate object bounding boxes. We evaluate different baselines on ChangingGrounding, and our Mem-ChangingGrounder achieves the highest localization accuracy while greatly reducing exploration cost. We hope this benchmark and method catalyze a shift toward practical, memory-centric 3DVG research for real-world applications. Project page: https://hm123450.github.io/CGB/ .

Recent advances in multimodal large language models (MLLMs) have shown remarkable capabilities in integrating vision and language for complex reasoning. While most existing benchmarks evaluate models under offline settings with a fixed set of pre-recorded inputs, we introduce OST-Bench, a benchmark designed to evaluate Online Spatio-Temporal understanding from the perspective of an agent actively exploring a scene. The Online aspect emphasizes the need to process and reason over incrementally acquired observations, while the Spatio-Temporal component requires integrating current visual inputs with historical memory to support dynamic spatial reasoning. OST-Bench better reflects the challenges of real-world embodied perception. Built on an efficient data collection pipeline, OST-Bench consists of 1.4k scenes and 10k question-answer pairs collected from ScanNet, Matterport3D, and ARKitScenes. We evaluate several leading MLLMs on OST-Bench and observe that they fall short on tasks requiring complex spatio-temporal reasoning. Under the online setting, their accuracy declines as the exploration horizon extends and the memory grows. Through further experimental analysis, we identify common error patterns across models and find that both complex clue-based spatial reasoning demands and long-term memory retrieval requirements significantly drop model performance along two separate axes, highlighting the core challenges that must be addressed to improve online embodied reasoning. To foster further research and development in the field, our codes, dataset, and benchmark are available. Our project page is: https://rbler1234.github.io/OSTBench.github.io/

Phone agents are increasingly expected to complete real mobile workflows rather than merely predict the next screen action. However, much of the current mobile-agent literature still evaluates agents primarily as GUI controllers that observe a screen, emit taps and swipes, and are scored by target app state. Real phone-use tasks are broader: they require deciding when to use app GUIs, device-side commands, or structured tools, while leaving evidence that the intended side effect actually occurred. We introduce PhoneHarness, a mixed-action benchmark and execution harness for studying phone-use agents on verifiable mobile workflows. PhoneHarness runs a device-side agent loop over GUI, CLI, and host-side tool actions, combining deterministic action routing with bounded GUI delegation and auditable execution traces. Its benchmark, PhoneHarness Bench, evaluates whether agents complete tasks with observable side effects, not only whether they produce plausible final answers. On the annotated evaluation split, PhoneHarness reaches a 75.0% pass rate, outperforming the strongest non-PhoneHarness settings by 12.9 percentage points. PhoneHarness and PhoneHarness Bench therefore play distinct but mutually dependent roles: the harness makes mixed phone workflows executable, while the benchmark measures whether agents can use that harness reliably and safely. Our findings suggest that reliable phone automation depends on action-surface routing and verifiable execution, not only visual GUI control.

LLM-based code agents navigate repositories through keyword search but miss the structural relationships, such as call graphs, inheritance hierarchies, and configuration dependencies, that define how software actually works. This makes agent navigation stochastic and difficult to reproduce across runs. We investigate whether lightweight static analysis can provide deterministic anchors for these agents: stable structural facts injected as plain-text comments that constrain probabilistic exploration and make navigation more predictable. Starting from a strong baseline, Codex from OpenAI, we systematically inject varying granularities of structural annotations and measure their effects on localization, trajectory behavior, and run-to-run stability. Our study identifies what we call the deterministic anchoring effect: static structure helps less by making agents "smarter" and more by making their navigation disciplined and reproducible. Three observations support this finding: (1) Anchoring works: lightweight call/inheritance topology improves function-level localization (+2.2pp Func@5) and shortens trajectories (-1.6 interaction rounds); (2) Anchoring is scale-sensitive: the optimal granularity and directionality depend on repository characteristics, where denser semantics show diminishing returns and hub-heavy projects benefit from inverse-only links that expose "who-calls-me" without forward edges; (3) Anchoring stabilizes: tags raise link-following rate from 0.15-0.18 to 0.21-0.24, roughly halve run-to-run variance, and improve single-run reliability (Pass@1 +3.4 pp) on medium-scale repositories, at the cost of roughly 10% more input tokens. These observations suggest practical guidelines: default to lightweight topology on medium projects, prune forward edges in large repositories, and reserve dense tags for implicit-dependency cases.

As generative agents become increasingly sophisticated and deployed in long-term interactive scenarios, their memory management capabilities emerge as a critical bottleneck for both performance and privacy. Current approaches either maintain unlimited memory stores, leading to computational intractability and privacy concerns, or employ simplistic forgetting mechanisms that compromise agent coherence and functionality. This paper introduces the Memory-Aware Retention Schema (MaRS), a novel framework for human-centered memory management in generative agents, coupled with six theoretically-grounded forgetting policies that balance performance, privacy, and computational efficiency. We present the Forgetful but Faithful Agent (FiFA) benchmark, a comprehensive evaluation framework that assesses agent performance across narrative coherence, goal completion, social recall accuracy, privacy preservation, and cost efficiency. Through extensive experimentation involving 300 evaluation runs across multiple memory budgets and agent configurations, we demonstrate that our hybrid forgetting policy achieves superior performance (composite score: 0.911) while maintaining computational tractability and privacy guarantees. Our work establishes new benchmarks for memory-budgeted agent evaluation and provides practical guidelines for deploying generative agents in resource-constrained, privacy-sensitive environments. The theoretical foundations, implementation framework, and empirical results contribute to the emerging field of human-centered AI by addressing fundamental challenges in agent memory management that directly impact user trust, system scalability, and regulatory compliance.

LLM agents increasingly act by writing code, yet a split persists between the runtime that drives the agent and the code the model writes. The runtime owns the loop, context, and control flow, and the model has little say over any of them. Letting model-written code shape the runtime itself would make agents more expressive, but it would also sharpen safety problems. A model can be diverted by a prompt injection, call the wrong tool, or fail partway and leave an inconsistent state, and each such failure reaches further when the code shapes the runtime than when it expresses a single action. We present LACUNA, a programming model for agents that closes this split while preserving safety. Each agent action is a typed call agent[T](task) that the LLM fills with code when execution reaches it, and the code is type-checked against the surrounding program before it runs. Because each action is accepted or rejected as a whole, a rejected one leaves the environment untouched, and its compiler diagnostics drive a retry. The same check also bounds which tools and data an action may use and how they flow. Our primitive expresses ReAct loops, sub-agents, skills, parallel decomposition, and multi-model planning as ordinary control flow. We evaluate LACUNA on a collection of test cases, BrowseComp-Plus, and τ^2-bench. On BrowseComp-Plus, 8.6% of generations are rejected before execution, with 0.7 retries per query on average, and the agent reaches 27.1% accuracy. On τ^2-bench, LACUNA solves 76.0% of 392 tasks across four domains with a capable model, on par with the baseline agent.

We study two-level autoresearch for cooperation: an outer-loop AI agent autonomously redesigns the inner-loop pipeline of an LLM policy-synthesis system for multi-agent Sequential Social Dilemmas (SSDs). A researcher agent R (run as a coding agent) reads the inner-loop source code, edits system prompts, feedback functions, helper libraries, and iteration logic, runs evaluations, and decides what to keep, following the autoresearch paradigm. Across two games (Cleanup and Gathering), two policy-synthesizer LLMs, and two welfare objectives (utilitarian efficiency and Rawlsian maximin), the researcher reliably exceeds hand-designed baselines, sharply tightens run-to-run variance, and outperforms prompt-only optimization. The discovered pipelines are objective-dependent: only under maximin does the researcher inject an explicit fairness mechanism into synthesizer pipelines, a class of mechanism that is absent from its own objective-agnostic system prompt and from every efficiency-optimized pipeline. This supports an information-design reading in which the researcher chooses what to reveal to the boundedly rational synthesizer as a function of the welfare objective. Code at https://github.com/vicgalle/autoresearch-social-dilemmas.

AI agents are entering high-risk production settings, where they use tools, retain context, follow policies, handle private data, and interact with users over multiple turns. Yet many evaluation methods still judge isolated outputs or static tasks, missing failures that emerge through trajectory, pressure, and adversarial interaction. We introduce ProofAgent Harness, open infrastructure for scalable, auditable, and adversarial AI agent evaluation. The harness provides evaluation infrastructure around an agent: it curates evaluation intelligence, runs adversarial multi-turn trials, captures behavioral traces, applies post-hoc multi-juror scoring, resolves disagreement, and produces evidence-linked reports. Its open design allows developers and researchers to extend domains, traps, metrics, juror personas, scoring rules, and reporting formats. At its core is Adversarial Multi-Juror Scoring with Turn-Level Audit, which evaluates completed agent behavior under pressure using calibrated juror personas, consensus checks, and turn-level evidence. Experiments across customer support, medical triage, privacy and security, and code generation agents show that strong agents fail selectively through weak metrics, fragile turns, unsafe reframing, and manipulation paths. We also find that a small quantized local Harness LLM can challenge production agents powered by best-in-class large LLMs, suggesting that evaluation capability emerges from the full harness pipeline rather than model scale alone. ProofAgent Harness turns AI agent evaluation from a static score into scalable adversarial evaluation infrastructure: repeatable, evidence-backed, extensible, and actionable before deployment.

As data-science agents shift from co-pilots to auto-pilots, silent misframing becomes a critical failure mode. Agents quietly commit to plausible but unintended task framings, producing clean, executable artifacts that hide their incorrect assessment of the task. Existing benchmarks score whether the pipeline runs, ignoring whether the agent recognized the task was underspecified. We introduce Ambig-DS, two diagnostic suites: one for prediction-target ambiguity (Ambig-DS-Target, 51 tasks built on DSBench, a tabular modeling benchmark) and one for evaluation-objective ambiguity (Ambig-DS-Objective, 61 tasks built on MLE-bench, a Kaggle-style ML competition benchmark), constructed so that scoring uses each source benchmark's original evaluator. For every task we pair the original, fully specified version with an ambiguous variant produced by controlled edits; a human-and-LLM verification pipeline confirms each variant admits multiple plausible interpretations with decision-relevant consequences. The suites are analyzed independently and ambiguity lowers performance in both. Across five agents spanning efficient to frontier-class models, we find in our controlled diagnostic setting: (i) failures are silent commitments: wrong-target submissions on Target, wrong-metric or non-committal baseline submissions on Objective, rather than execution errors; (ii) allowing the agent to ask one clarifying question recovers much of the loss under idealized conditions, suggesting missing framing information drives a substantial part of the observed degradation; but (iii) agents cannot reliably tell when to use it: permissive prompts induce over-asking on clear tasks, while conservative prompts induce silent defaulting on ambiguous ones. Recognizing target and objective underspecification, not pipeline execution, is the bottleneck missing from standard DS-agent evaluations.

Large language models can generate code and call tools with remarkable fluency, yet deploying them as practical software engineering assistants still expose stubborn gaps: finite context windows, single mistakes that derail entire sessions, agents that get stuck in dead ends, AI slop, and generated changes that are difficult to review or revert. We present KISS Sorcar, a general-purpose assistant and integrated development environment (IDE) built on top of the KISS Agent Framework, a stupidly-simple AI agent framework of roughly 1,850 lines of code. The framework addresses these gaps using a robust system prompt and through a five-layer agent hierarchy in which each layer adds exactly one concern: budget-tracked ReAct execution, automatic continuation across sub-sessions via summarization, coding, and browser tools with parallel sub-agents, persistent multi-turn chat with history recall, and git worktree isolation so every task runs on its own branch. To assess the power of the KISS agent framework, we implemented KISS Sorcar as a free, open-source Visual Studio Code extension that runs locally and effectively for long-horizon tasks, and supports browser automation, multimodal input, and Docker containers. In this research, we deliberately prioritize output quality over latency: giving a frontier model adequate time to validate its own output -- running linters, type checkers, and tests -- dramatically reduces the low-quality code that plagues faster but less thorough agents. The entire system was built using itself in 4.5 months, providing a continuous stress test in which any agent-introduced bug immediately impairs its own ability to work. On Terminal Bench 2.0, KISS Sorcar achieves a 62.2% overall pass rate with Claude Opus 4.6, comparing favorably to Claude Code (58%) and Cursor Composer 2 (61.7).

Agentic artificial intelligence systems promise to accelerate scientific workflows, but neuroimaging poses unique challenges: heterogeneous modalities (sMRI, fMRI, dMRI, EEG), long multi-stage pipelines, and persistent reproducibility risks. To address this gap, we present NeuroClaw, a domain-specialized multi-agent research assistant for executable and reproducible neuroimaging research. NeuroClaw operates directly on raw neuroimaging data across formats and modalities, grounding decisions in dataset semantics and BIDS metadata so users need not prepare curated inputs or bespoke model code. The platform combines harness engineering with end-to-end environment management, including pinned Python environments, Docker support, automated installers for common neuroimaging tools, and GPU configuration. In practice, this layer emphasizes checkpointing, post-execution verification, structured audit traces, and controlled runtime setup, making toolchains more transparent while improving reproducibility and auditability. A three-tier skill/agent hierarchy separates user-facing interaction, high-level orchestration, and low-level tool skills to decompose complex workflows into safe, reusable units. Alongside the NeuroClaw framework, we introduce NeuroBench, a system-level benchmark for executability, artifact validity, and reproducibility readiness. Across multiple multimodal LLMs, NeuroClaw-enabled runs yield consistent and substantial score improvements compared with direct agent invocation. Project homepage: https://cuhk-aim-group.github.io/NeuroClaw/index.html

Autonomous agents powered by large language models (LLMs) promise to accelerate scientific discovery end-to-end, but rigorously evaluating their capacity for verifiable discovery remains a central challenge. Existing benchmarks face a trade-off: they either heavily rely on LLM-as-judge evaluations of automatically generated research outputs or optimize convenient yet isolated performance metrics that provide coarse proxies for scientific insight. To address this gap, we introduce FIRE-Bench (Full-cycle Insight Rediscovery Evaluation), a benchmark that evaluates agents through the rediscovery of established findings from recent, high-impact machine learning research. Agents are given only a high-level research question extracted from a published, verified study and must autonomously explore ideas, design experiments, implement code, execute their plans, and derive conclusions supported by empirical evidence. We evaluate a range of state-of-the-art agents with frontier LLMs backbones like gpt-5 on FIRE-Bench. Our results show that full-cycle scientific research remains challenging for current agent systems: even the strongest agents achieve limited rediscovery success (<50 F1), exhibit high variance across runs, and display recurring failure modes in experimental design, execution, and evidence-based reasoning. FIRE-Bench provides a rigorous and diagnostic framework for measuring progress toward reliable agent-driven scientific discovery.

Recent progress in large language model (LLM) agents has largely focused on embedding self-improvement mechanisms inside the agent or searching over many concurrent variants. While these approaches can raise aggregate scores, they often yield unstable and hard-to-audit improvement trajectories, making it difficult to guarantee non-regression or to reason about failures across versions. We reframe agent improvement as release engineering: agents are treated as shippable artifacts, and improvement is externalized into a regression-aware release pipeline. We introduce AgentDevel, a release engineering pipeline that iteratively runs the current agent, produces implementation-blind, symptom-level quality signals from execution traces, synthesizes a single release candidate (RC) via executable diagnosis, and promotes it under flip-centered gating. AgentDevel features three core designs: (i) an implementation-blind LLM critic that characterizes failure appearances without accessing agent internals, (ii) script-based executable diagnosis that aggregates dominant symptom patterns and produces auditable engineering specifications, and (iii) flip-centered gating that prioritizes pass to fail regressions and fail to pass fixes as first-class evidence. Unlike population-based search or in-agent self-refinement, AgentDevel maintains a single canonical version line and emphasizes non-regression as a primary objective. Experiments on execution-heavy benchmarks demonstrate that AgentDevel yields stable improvements with significantly fewer regressions while producing reproducible, auditable artifacts. Overall, AgentDevel provides a practical development discipline for building, debugging, and releasing LLM agents as software development.

LLM web agents now browse and take actions on the open web, yet current agent evaluations are constrained to sandboxed environments or artificial tasks. We introduce BrowserArena, a live open-web agent evaluation platform that collects user-submitted tasks, runs Arena-style head-to-head comparisons, and uses step-level human feedback to surface failure modes. Collecting and analyzing step-level annotations on the agent traces, we identify three consistent failure modes: captcha resolution, pop-up banner removal, and direct navigation to URLs. By constructing targeted datasets to further study these tasks, we discover variations in how different language models navigate these failure modes. We find, for example, that o4-mini deploys a wider variety of strategies to circumvent captcha resolution than other models and DeepSeek-R1 consistently misleads users about pop-up banner closure. Our findings surface both the diversity and brittleness of current web agents. More broadly, our benchmarking methodology provides an approach to evaluating and understanding web agent failure modes at scale.

Automating scientific discovery requires more than generating papers from ideas. Real research is iterative: hypotheses are challenged from multiple perspectives, experiments fail and inform the next attempt, and lessons accumulate across cycles. Existing autonomous research systems often model this process as a linear pipeline: they rely on single-agent reasoning, stop when execution fails, and do not carry experience across runs. We present AutoResearchClaw, a multi-agent autonomous research pipeline built on five mechanisms: structured multi-agent debate for hypothesis generation and result analysis, a self-healing executor with a Pivot/Refine decision loop that transforms failures into information, verifiable result reporting that prevents fabricated numbers and hallucinated citations, human-in-the-loop collaboration with seven intervention modes spanning full autonomy to step-by-step oversight, and cross-run evolution that converts past mistakes into future safeguards. On ARC-Bench, a 25-topic experiment-stage benchmark, AutoResearchClaw outperforms AI Scientist v2 by 54.7%. A human-in-the-loop ablation across seven intervention modes reveals that precise, targeted collaboration at high-leverage decision points consistently outperforms both full autonomy and exhaustive step-by-step oversight. We position AutoResearchClaw as a research amplifier that augments rather than replaces human scientific judgment. Code is available at https://github.com/aiming-lab/AutoResearchClaw.

We introduce the MiniMax-M2 series, a family of Mixture-of-Experts language models built around the principle that mini activations can unleash maximum real-world intelligence. The flagship M2 contains 229.9B total parameters with only 9.8B activated per token. Designed end-to-end for agentic deployment, the M2 series rests on three components: (i) agent-driven data pipelines producing large-scale, verifiable trajectories across agentic coding and agentic cowork, each grounded in an executable workspace and an artifact-aligned reward; (ii) Forge, a scalable agent-native RL system that adapts to long-horizon agent trajectories, paired with windowed-FIFO scheduling, prefix-tree merging, inference optimization, and a clean training-inference-agent decoupling that supports both white-box and black-box agents; (iii) the latest M2.7 checkpoint takes an early step toward self-evolution -- autonomously debugging training runs and modifying its own scaffold. Across M2 through M2.7, this combination translates a mini-activation footprint into frontier-tier performance on agentic coding, deep search, office-task, and reasoning benchmarks.

The field of Artificial Intelligence is undergoing a transition from Generative AI -- probabilistic generation of text and images -- to Agentic AI, in which autonomous systems execute actions within external environments on behalf of users. This transition exposes a fundamental architectural mismatch: Large Language Models (LLMs) produce stochastic, unstructured outputs, whereas the backend infrastructure they must control -- databases, APIs, cloud services -- requires deterministic, schema-conformant inputs. The present paper describes the Auton Agentic AI Framework, a principled architecture for standardizing the creation, execution, and governance of autonomous agent systems. The framework is organized around a strict separation between the Cognitive Blueprint, a declarative, language-agnostic specification of agent identity and capabilities, and the Runtime Engine, the platform-specific execution substrate that instantiates and runs the agent. This separation enables cross-language portability, formal auditability, and modular tool integration via the Model Context Protocol (MCP). The paper formalizes the agent execution model as an augmented Partially Observable Markov Decision Process (POMDP) with a latent reasoning space, introduces a hierarchical memory consolidation architecture inspired by biological episodic memory systems, defines a constraint manifold formalism for safety enforcement via policy projection rather than post-hoc filtering, presents a three-level self-evolution framework spanning in-context adaptation through reinforcement learning, and describes runtime optimizations -- including parallel graph execution, speculative inference, and dynamic context pruning -- that reduce end-to-end latency for multi-step agent workflows.

Computer-using agents drive real software through the screen -- clicking and typing -- but they solve every task from scratch: asked to repeat a task, an agent re-reads the screen, re-reasons every tap, and pays the full cost again. We present PreAct, which lets such an agent get faster on tasks it has done before. The first time it succeeds, PreAct compiles the run into a small state-machine program-states that check the screen, transitions that act-and on later runs replays it directly instead of invoking the agent 8.5-13x faster, with no per-step language-model calls. Replay is not blind: at each step PreAct checks that the screen matches what the program expects before acting, and hands control back to the agent the moment something is off. PreAct applies the same discipline when deciding what to keep: a freshly compiled program enters the store only if, re-run from a clean state, an independent evaluator confirms it solved the task-catching programs that replay to their last step yet leave the task undone. Across a mobile, a desktop, and a web benchmark, this store-time check separates repeated runs that improve from ones that degrade as faulty programs accumulate, worth 1.75-2.6 tasks per benchmark, the same direction on all three; a fallback that explores afresh when no program fits brings PreAct level with a strong record-and-replay baseline. We also report what did not matter: prompt wording, runtime guardrails, and whether a language model or a plain embedding retriever selects which program to reuse.

Web agents can autonomously complete online tasks by interacting with websites, but their exposure to open web environments makes them vulnerable to prompt injection attacks embedded in HTML content or visual interfaces. Existing guard models still suffer from limited generalization to unseen domains and attack patterns, high false positive rates on benign content, reduced deployment efficiency due to added latency at each step, and vulnerability to adversarial attacks that evolve over time or directly target the guard itself. To address these limitations, we propose WARD (Web Agent Robust Defense against Prompt Injection), a practical guard model for secure and efficient web agents. WARD is built on WARD-Base, a large-scale dataset with around 177K samples collected from 719 high-traffic URLs and platforms, and WARD-PIG, a dedicated dataset designed for prompt injection attacks targeting the guard model. We further introduce A3T, an adaptive adversarial attack training framework that iteratively strengthens WARD through a memory-based attacker and guard co-evolution process. Extensive experiments show that WARD achieves nearly perfect recall on out-of-distribution benchmarks, maintains low false positive rates to preserve agent utility, remains robust against guard-targeted and adaptive attacks under substantial distribution shifts, and runs efficiently in parallel with the agent without introducing additional latency.

Deployed language models must produce outputs that are both correct and format-compliant. We study this structured-output reliability gap using two mathematical benchmarks -- GSM8K and MATH -- as a controlled testbed: ground truth is unambiguous and the output contract is strict (JSON with required fields). We evaluate three 7-9B models under five prompting strategies and report output accuracy -- the joint event of mathematical correctness and valid JSON structure -- as the primary metric. A systematic format failure emerges: NAIVE prompting (no system prompt) achieves up to 85% task accuracy on GSM8K but 0% output accuracy across all models and datasets. REFERENCE prompting (a minimal hand-written JSON format prompt) fares little better, yielding 0% output accuracy for two of four models tested. Constrained decoding enforces syntactic validity but incurs 3.6x-8.2x latency overhead and in several settings degrades task performance substantially. To overcome this limitation, we developed AloLab, an iterative system-prompt optimizer (meta-agent: Claude Sonnet 4.5) requiring only black-box API access to the target model; it reaches 84-87% output accuracy on GSM8K and 34-40% on MATH across five independent runs per model, with 29/30 paired McNemar comparisons against the best static prompt significant at p < 0.05, at near-NAIVE inference latency and without model fine-tuning. The same format failure extends to GPT-4o (OpenAI, 2024), a proprietary closed-source model: REFERENCE achieves 0% output accuracy due to systematic markdown-fence wrapping, while AloLab reaches 95.2% [94.8, 95.6]. An ablation replacing the Sonnet 4.5 meta-agent with Claude 3 Haiku reduces mean output accuracy to 61.0% and increases run-to-run standard deviation from <1 pp to 21.8 pp, confirming that meta-agent capability is a primary driver of optimization quality.

AI agents are increasingly deployed in multi-tenant cloud environments, where they execute diverse tool calls within sandboxed containers, each call with distinct resource demands and rapid fluctuations. We present a systematic characterization of OS-level resource dynamics in sandboxed AI coding agents, analyzing 144 software engineering tasks from the SWE-rebench benchmark across two LLM models. Our measurements reveal that (1) OS-level execution (tool calls, container and agent initialization) accounts for 56-74% of end-to-end task latency; (2) memory, not CPU, is the concurrency bottleneck; (3) memory spikes are tool-call-driven with a up to 15.4x peak-to-average ratio; and (4) resource demands are highly unpredictable across tasks, runs, and models. Comparing these characteristics against serverless, microservice, and batch workloads, we identify three mismatches in existing resource controls: a granularity mismatch (container-level policies vs. tool-call-level dynamics), a responsiveness mismatch (user-space reaction vs. sub-second unpredictable bursts), and an adaptability mismatch (history-based prediction vs. non-deterministic stateful execution). We propose AgentCgroup, an intent-driven eBPF-based resource controller that exploits agents ability to declare resource needs and reconstruct execution strategies, using hierarchical cgroup structures aligned with tool-call boundaries, in-kernel enforcement via sched_ext and memcg_bpf_ops, and runtime-adaptive policies. Preliminary evaluation demonstrates improved multi-tenant isolation and reduced resource waste. AgentCgroup is open-source at https://github.com/eunomia-bpf/agentcgroup

Computer-use agents provide a promising path toward general software automation because they can interact directly with arbitrary graphical user interfaces instead of relying on brittle, application-specific integrations. Despite recent advances in benchmark performance, strong computer-use agents remain expensive and slow in practice, since most systems invoke large multimodal models at nearly every interaction step. We argue that this uniform allocation of compute is fundamentally inefficient for long-horizon GUI tasks. Such trajectories are highly heterogeneous: many steps are routine and can be handled reliably by smaller, cheaper policies, while errors tend to concentrate at a relatively small number of high-risk moments. Across computer-use benchmarks, these failures repeatedly take two forms: progress stalls, where the agent loops, repeats ineffective actions, or fails to make meaningful progress, and silent semantic drift, where the agent continues taking locally plausible actions after already deviating from the user's true goal. To address this inefficiency, we propose an event-driven, step-level cascade for computer-use agents that runs a small policy by default and escalates to a stronger model only when lightweight learned monitors detect elevated risk. Our framework combines two complementary signals: a Stuck Monitor that detects degraded progress from recent reasoning-action history and triggers recovery, and a Milestone Monitor that identifies semantically meaningful checkpoints where sparse verification is most informative for catching drift. This design turns always-on frontier-model inference into adaptive, on-demand compute allocation over the course of an evolving interaction. The framework is modular and deployment-oriented: it can be layered on top of existing computer-use agents without changing the underlying agent architecture or retraining the large model.

AI agents are increasingly expected to complete long-horizon workflows that require sustained progress over hours, millions of tokens, and complex environments. Yet current agent benchmarks largely evaluate short-form tasks, such as single pull requests, small tickets, or 5-10 minute exercises, limiting our ability to measure agents' capabilities in planning, long-context understanding, and memory use. We introduce SWE-Marathon, a benchmark of 20 long-horizon tasks spanning software engineering and adjacent technical domains. Each task consists of a unique executable environment, a human-written reference solution, and a multi-layer verification suite. Logged agent attempts average 27.2M total tokens, making SWE-Marathon substantially longer-horizon than existing SWE and command-line agent benchmarks. Current frontier coding agents solve fewer than 30% of tasks. Failures often arise from poor self-verification, self-reported infeasibility, and premature termination. We also observe reward-hacking behavior in 13.8% of rollouts, where agents attempt to exploit the environment or verifier to bypass the intended workflow. SWE-Marathon includes adversarial review of test suites and execution environments, as well as multi-layer checks designed to prevent shortcut solutions. We release SWE-Marathon, evaluation code, and agent trajectories at https://swe-marathon.org/.

Autonomous agents are increasingly expected to support end-to-end medical-AI research workflows, moving beyond isolated prediction tasks or short-form clinical question answering. However, existing medical agent benchmarks primarily evaluate final outputs, providing limited visibility into agent behavior within the research process. To address this gap, we present AutoMedBench, a workflow-aware benchmark for autonomous medical-AI research across diverse medical imaging and multimodal inference tasks, organizing agent execution into a unified five-stage workflow (S1-S5): Plan, Setup, Validate, Inference, and Submit. It comprises long-horizon tasks with each run averaging 33 agent turns, spanning five research tracks: segmentation, image enhancement, visual question answering (VQA), report generation, and lesion detection. Each task is evaluated under two difficulty tiers, Lite and Standard, which use the same data and metrics but differ in the amount of task-brief scaffolding, and each run is scored using both final task performance and S1-S5 stage scores, enabling stage-level analysis from the initial task brief to the final submitted artifact. Across thousands of recorded runs, stage-level scoring reveals that Validate is the weakest workflow stage on average, whereas Setup is the strongest, suggesting that current agents are better at making pipelines executable than at verifying their reliability. Post-run error analysis further shows that verification and submission failures dominate tagged errors, accounting for 37.7% and 38.1% of fired codes respectively, whereas task-understanding errors are rare at 0.9%, and runs with one fired error code have a 48% lower overall score than runs with no error code on average.

Scaling up executable code data is significant for improving language models' software engineering capability. The intricate nature of the process makes it labor-intensive, time-consuming and expert-knowledge-dependent to build a large number of executable code repositories, limiting the scalability of existing work based on running tests. The primary bottleneck lies in the automated building of test environments for different repositories, which is an essential yet underexplored task. To mitigate the gap, we introduce Repo2Run, the first LLM-based agent aiming at automating the building of executable test environments for any repositories at scale. Specifically, given a code repository, Repo2Run iteratively builds the Docker image, runs unit tests based on the feedback of the building, and synthesizes the Dockerfile until the entire pipeline is executed successfully. The resulting Dockerfile can then be used to create Docker container environments for running code and tests. We created a benchmark containing 420 Python repositories with unit tests for evaluation. The results illustrate that Repo2Run achieves an 86.0% success rate, outperforming SWE-agent by 77.0%. The resources of Repo2Run are available at https://github.com/bytedance/Repo2Run.

Existing tool-use benchmarks for LLM agents are overwhelmingly linear: our analysis of six benchmarks shows 55 to 100% of instances are simple chains of 2 to 5 steps. We introduce The Amazing Agent Race (AAR), a benchmark featuring directed acyclic graph (DAG) puzzles (or "legs") with fork-merge tool chains. We release 1,400 instances across two variants: sequential (800 legs) and compositional (600 DAG legs). Agents must navigate Wikipedia, execute multi-step tool chains, and aggregate results into a verifiable answer. Legs are procedurally generated from Wikipedia seeds across four difficulty levels with live-API validation. Three complementary metrics (finish-line accuracy, pit-stop visit rate, and roadblock completion rate) separately diagnose navigation, tool-use, and arithmetic failures. Evaluating three agent frameworks on 1,400 legs, the best achieves only 37.2% accuracy. Navigation errors dominate (27 to 52% of trials) while tool-use errors remain below 17%, and agent architecture matters as much as model scale (Claude Code matches Codex CLI at 37% with 6x fewer tokens). The compositional structure of AAR reveals that agents fail not at calling tools but at navigating to the right pages, a blind spot invisible to linear benchmarks. The project page can be accessed at: https://minnesotanlp.github.io/the-amazing-agent-race

A multi-agent pipeline with N agents typically issues N LLM calls per run. Merging agents into fewer calls (compound execution) promises token savings, but naively merged calls silently degrade quality through tool loss and prompt compression. We present Agent Capsules, an adaptive execution runtime that treats multi-agent pipeline execution as an optimization problem with empirical quality constraints. The runtime instruments coordination overhead per group, scores composition opportunity, selects among three compound execution strategies, and gates every mode switch on rolling-mean output quality. A controlled negative result confirms that injecting more context into a merged call worsens compression rather than relieving it, so the framework's escalation ladder (standard, then two-phase, then sequential) recovers quality by moving toward per-agent dispatch rather than by rewriting merged prompts. On LLM-judged quality, the controller matches a hand-tuned oracle on every measured (model, group, mode) cell: routing compound whenever the oracle would, and reverting to fine whenever quality would fail the floor, without per-model configuration. Against a hand-crafted LangGraph implementation of a 14-agent competitive intelligence pipeline, Agent Capsules uses 51% fewer fine-mode input tokens and 42% fewer compound-mode input tokens, at +0.020 and +0.017 quality respectively. Against a DSPy implementation of a 5-agent due diligence pipeline, the framework uses 19% fewer tokens than uncompiled DSPy at quality parity, and 68% fewer tokens than MIPROv2 at +0.052 quality. Even before compound mode fires, the runtime delivers efficiency through automatic policy resolution, cache-aligned prompts, and topology-aware context injection, matching both hand-tuned and compile-time baselines without training data or per-pipeline engineering.

The safety of autonomous AI agents is increasingly recognized as a critical open problem. As agents transition from passive text generators to active actors capable of executing shell commands, modifying files, calling APIs, and browsing the web, the consequences of unsafe or adversarially manipulated behavior become immediate and tangible. Existing AI safety work has focused primarily on model alignment and input filtering, but these approaches do not address what happens at the moment an agent's intent becomes a real action on a real machine. This gap is especially acute in local environments, where developers run agents against their own filesystems, credentials, and infrastructure with little runtime control. This paper introduces AgentWall, a runtime safety and observability layer for local AI agents. AgentWall intercepts every proposed agent action before it reaches the host environment, evaluates it against an explicit declarative policy, requires human approval for sensitive operations, and records a complete execution trail for audit and replay. It is implemented as a policy-enforcing MCP proxy and native OpenClaw plugin, working across Claude Desktop, Cursor, Windsurf, Claude Code, and OpenClaw with a single install command. We present the design, architecture, threat model, and policy model of AgentWall, and demonstrate 92.9% policy enforcement accuracy with sub-millisecond overhead across 14 benchmark tests. AgentWall is open-source at https://github.com/agentwall/Agentwall.