Large Language Model Agents

LLM agents increasingly operate in large tool ecosystems, where real-world tasks require discovering relevant tools, inferring implicit sub-goals, and adapting to dynamic environments over long horizons. However, existing benchmarks rarely evaluate planning under retrieval-limited tool visibility. To address this gap, we introduce PlanBench-XL, an interactive benchmark of 327 retail tasks over 1,665 tools that tests whether agents can iteratively retrieve usable tools, invoke them to uncover intermediate evidence for subsequent calls toward the final goal. PlanBench-XL further features an optional blocking mechanism that simulates real-world unpredictability through missing, failing, or distracting tool functions, forcing agents to detect disrupted paths and adapt at runtime. Experiments on ten leading LLMs show that massive-tool planning remains challenging: while GPT-5.4 achieves 51.90% accuracy in block-free settings, it collapses to 11.36% under the most severe blocking condition. Further analysis shows that agents are especially vulnerable when failures lack explicit error signals or when recovery requires longer alternative tool-use paths. These results establish PlanBench-XL as a testbed for diagnosing agentic planning failures and highlight the need for robust adaptive planning in long-horizon tasks with large, imperfect tool environments.

External skills can improve action-oriented LLM agents without changing model weights, but persistent skill updates are risky when they are distilled from sparse or noisy trajectories. A plausible reflection may encode a useful procedure, a spurious shortcut, or a rule that the target executor cannot reliably follow. We propose Hypothesis-Driven Skill Optimization (HDSO), a train-free framework in which both the skill curator and the agent executor are frozen inference endpoints. The curator observes executor traces, proposes a falsifiable hypothesis with an explicit validation plan, instantiates the hypothesis as a candidate skill package, validates the package through paired control/treatment executions, reviews behavior differences, and consolidates only supported candidates into an approved repository. The executor consumes approved skills through progressive disclosure, preserving the executor-only path when no skill is selected. On ALFWorld, HDSO improves executor-only baselines by +6.9 Avg. SR points for Qwen3-8B and +4.0 points for Qwen3.6-27B. Under 20% randomly flipped success/failure feedback during skill discovery and validation, HDSO preserves a +7.1-point gain for Qwen3-8B. Transfer and heterogeneous-pair diagnostics further show that validated repositories can be useful beyond the run that produced them, but cross-model curation succeeds only when curator diagnosis, executor capability, and validation evidence align. HDSO provides an auditable skill lifecycle for frozen action agents rather than an unconstrained memory accumulation procedure.

The increasing deployment of large language model (LLM) agents in collaborative workflows demands robust multi-user, multi-principal interaction mechanisms capable of enforcing access permissions, resolving authoritative conflicts, and preventing unauthorized data disclosure. However, a fundamental mismatch exists between the single-user training paradigm of contemporary LLMs and the hard constraints required for multi-principal governance, rendering probabilistic, prompt-based safeguards vulnerable under multi-turn adversarial interactions.Our key insight is that governance constraints -- who is authorized, what is restricted, and whose instructions take precedence -- are deterministic runtime variables that should be enforced by execution hooks rather than entrusted to the LLM. We present \textbf{Harness-MU}, the first model-agnostic, zero-tuning infrastructure framework for multi-user LLM agents. By decoupling language generation from safety orchestration, Harness-MU guarantees unbreakable permission boundaries while maximizing compliant demand satisfaction. Across four frontier open-weight and proprietary models on the \textit{Muses-Bench} benchmark, Harness-MU achieves the goal of privacy preservation across all access-control attacks, outperforming the standard baseline by 0.28--0.39 in utility score and improving instruction-following accuracy by up to 48.9 percentage points. Harness-MU advances the philosophy of \textit{Harness Engineering}, establishing that systematic infrastructure is essential for solving LLM multi-principal governance challenges. The code and data are available at https://github.com/YuanJrShiuan/Harness-MulUser.

Computerized Adaptive Testing (CAT), as a key technology for personalized education, aims to accurately assess examinee proficiency by retrieving exercises dynamically matching current ability estimates. However, existing CAT research is constrained by limitations of static offline data and isolated component optimization. Restricted by partial labels in offline logs, researchers degrade the dynamic assessment process into static sequence prediction. Current research focuses on isolated perspectives, e.g., selection or diagnosis, neglecting the overall CAT interaction process. To address this, we propose AgentCAT, a Large Language Model-based multi-agent simulation system, to construct a high-fidelity benchmarking environment for dynamic testing. This framework comprises three modules: (1) The examinee agent with memory retrieval and Chain-of-Thought reasoning simulates responses based on cognitive profiles; (2) The selection agent uses coarse-to-fine bucketing and knowledge graph exploration to balance local difficulty and global coverage; (3) The supervisor uses dual-auditing and robust update to ensure convergence and validity. To validate the framework, we evaluated on two real-world datasets across three dimensions: macro-level ability convergence, micro-level interaction logic, and data sparsity resilience. Results show AgentCAT achieves effective ability estimation, and its selection strategy balances difficulty adaptation and instructional coherence, aligning with human pedagogical intuition.

Translating natural-language planning intent into verified plans is a longstanding challenge: people communicate goals in language, while classical planners require formal PDDL specifications. Recent agentic frameworks bridge this gap by orchestrating a pool of specialized repair agents inside a verifier-checked refinement loop, but the orchestrator at the centre is itself a prompted frontier LLM, paying a frontier-LLM API call at every refinement step. We present HALO (Hybrid Agent-Learned Orchestrator), which trains the orchestrator from refinement trajectories that an external verifier has certified as ending in valid plans, across 11 PDDL domains. HALO pairs a small QLoRA-tuned policy with three hardcoded rules for trivially decidable selections, and operates over an expanded 21-agent action space. Unlike approaches that prompt a frontier LLM at every step or learn an orchestrator from sparse end-of-episode rewards, our key observation is that the verifier already provides strong guidance: every accepted trajectory is a sequence of demonstrably correct (state, agent) decisions, directly usable as supervision. Across PlanBench, Natural Plan, and classical planning benchmarks, HALO matches or exceeds the GPT-5-mini prompted baseline on success rate, sits within three percentage points of the stronger Gemini-3-Flash prompted baseline, reduces orchestration cost by more than an order of magnitude ($0.18 to $0.004 per task against GPT-5-mini, roughly 45$\times$ cheaper; roughly 15$\times$ cheaper than Gemini-3-Flash), and cuts total LLM calls per episode by 40 to 50 percent.

Summarization-based prompt compression is increasingly used by LLM agents to shorten long, distributed contexts, but it shifts the security boundary: filters inspect the pre-compression prompt while the backend acts on a newly generated compressed context. We identify relinking, a compression-boundary vulnerability where the compressor behaves as a confused deputy, summarizing distributed, locally benign fragments into a complete malicious instruction. Unlike prompt injection, relinking need not place an explicitly malicious payload in the source context. We show that relinking arises from summarization itself: attention makes separated fragments jointly available, pre-training makes compatible fragments plausible to connect, and post-training favors compact backend-actionable summaries. We formalize the attacker-induced form as adversarial relinking and present Relink, an automated DSL-based tool that splits malicious payloads into benign fragments while keeping the complete payload absent before compression. Across four long-context agent benchmarks, Relink achieves 86.9% Relink Rate and Backend Action Rate versus 17.0% for clean-split controls. Existing defenses fail to reliably capture adversarial relinking; our KBRA defense reduces residual Backend Action Rate to 0.0%.

AI agents acting on behalf of users are constantly making decisions, and for users to trust their agents, those decisions must align with what they actually want. Privacy is an important alignment problem for agents: every message, post, or tool call an agent makes is a contextual judgment about what is appropriate to share, with whom, and under which conditions. Because such judgments depend on social expectations and norms, human judgment does not merely label privacy violations but also helps define them. While existing work relies on unreliable proxies for both training and evaluation, we place human judgment at the center of agentic privacy alignment. We introduce PrivacyAlign, a dataset of 1,350 samples with 3,516 detailed annotations from 599 unique annotators across diverse scenarios where current LLMs actually leak, and use it to ground both alignment training and automated evaluation in human privacy norms. Building on these annotations, we first show that conditioning LLM judges on human annotations and explanations for reference responses to the same prompt makes their judgments more reliable. We then introduce annotation-conditioned reward modeling, which uses these annotations to score new responses during RL, and show that small open-weight agents trained with this reward better align with human privacy norms, with strong gains on PrivacyAlign and existing privacy benchmarks for agents.

Multi-agent LLM systems routinely produce hallucinated outputs that cannot be explained by model deficiencies alone. A significant class of these failures arises not from model incapacity but from context drift: the divergence of internal knowledge states between concurrent agents. When agents enter a collaborative task with mismatched or stale representations of shared world state, their joint reasoning produces contradictions that manifest as hallucination. We define the Context Divergence Score (CDS), a lightweight scalar metric quantifying knowledge-state discrepancy between agent pairs across spatial, temporal, and task dimensions, and propose the Shared State Verification Protocol (SSVP), which lets agents periodically exchange compressed state summaries and flag high-divergence conditions before joint reasoning. We evaluate SSVP across two domains (multi-agent travel and software project planning) using Claude Haiku. In controlled experiments (n=30 per condition, travel; n=10, software) across 8 scenarios, naive full-broadcast synchronization increases hallucination rate by 34% above the no-sync baseline (HR: 0.658 vs. 0.492, p=0.0022, d=1.18), a contamination effect from propagating erroneous agent states. SSVP avoids this failure mode while showing modest, consistent reduction (HR: 0.463, d=0.30) and achieves significantly lower hallucination than full-broadcast (p=0.0005, d=1.47) using 58% fewer API calls. The contamination effect does not replicate in the software domain, where all conditions converge to low HR (<0.2), confirming it is specific to tasks where one erroneous shared belief cascades across evaluation dimensions. Our results reframe hallucination mitigation as a distributed systems problem and establish context synchronization as a first-class primitive in multi-agent LLM design.

Tool-augmented agents are typically evaluated by their gains under reliable external feedback. Yet these gains leave open a key counterfactual: when feedback is unreliable, would the agent be better off receiving no task evidence? We study this question with a controlled matched-loop comparison that fixes the agent loop, prompt, action space, and decoding, while varying only the returned observation: faithful, misleading, or absent. Across question answering and fact verification, persistent misleading feedback produces a value inversion: agents that benefit from clean tools can perform worse than the matched no-feedback fallback. On HotpotQA, Qwen2.5-7B reaches 44.8 F1 with clean retrieval and 22.3 F1 with no feedback, but drops to 4.7 F1 under shuffled retrieval. The inversion persists under stronger clean retrieval and locally plausible distractors, but weakens when later clean evidence can repair the trajectory. Early trajectory signals predict many failures, yet simple repairs remain fallback-limited: rejecting bad evidence helps only when the exposed fallback is reliable. These results show that clean-tool gains can overstate tool value, and that matched no-feedback fallback controls are necessary for evaluating tool-augmented agents.

The capabilities of frontier Large Language Models (LLMs) continue to advance, with different providers increasingly specializing in distinct domains. This raises a natural next objective: how to combine the individual specializations of various LLMs into a collectively intelligent system. To this end, we report the development of Sakana Fugu, a family of orchestrator models that harness and amplify the capabilities of an LLM agent team. Fugu models are themselves language models trained to understand user queries and dynamically devise agentic scaffolds to solve them. Through these adaptive scaffolds, Fugu accesses performance beyond any individual LLM agent, achieving state-of-the-art results compared to other publicly accessible models across a range of challenging tasks, including SWE-Bench Pro, Terminal Bench, LiveCodeBench, GPQA-Diamond, Humanity's Last Exam, and CharXiv Reasoning. We release two models: Fugu, which balances performance with latency for everyday use, and Fugu-Ultra, which prioritizes answer quality on the hardest problems. We describe our training paradigm, which encompasses large-scale fine-tuning, evolutionary algorithms, and reinforcement learning approaches, along with the infrastructure and core design principles that turn these methods into a production system. We hope this report encourages further research into multi-agent systems and dynamic, query-adaptive agentic scaffolds as a path toward the next frontier of AI capabilities, accessed through collective intelligence.

High-stakes reasoning tasks necessitate transparent and verifiable workflows, yet conventional single-model large language models (LLMs) often struggle with hallucination and low interpretability under zero-shot paradigms. To address this general AI challenge, we propose a Multi-Agent Audit Framework that simulates a collaborative, multi-step verification process. We empirically validate this architecture in the sensitive domain of clinical mental health screening using a modular LangChain workflow. Our framework decomposes the reasoning process into a Perception Agent, Knowledge Retrieval-Augmented Generation (RAG), Chain-of-Thought (CoT) clinical inference, and a critical Audit verification stage. We evaluated this framework on the DAIC-WOZ dataset using locally deployed open-source models. Experimental results demonstrate that our multi-agent pipeline significantly outperforms single-agent baselines, reducing the Mean Absolute Error (MAE) for PHQ-8 depression severity prediction from 5.35 to 5.02. By exposing cross-agent validation traces, the framework mitigates reasoning drift and provides highly interpretable diagnostic rationales, offering a generalizable paradigm for reliable AI-assisted decision support beyond isolated model scaling. We make data and code open access on GitHub for replicability.

Long-horizon large language model (LLM) agents accumulate interaction trajectories that quickly exceed any practical prompt budget, and existing memory methods either truncate aggressively and lose non-local evidence or retain boilerplate that degrades decision quality. We ask a mechanism question rather than claiming a better general-purpose memory system: when does organizing trajectory memory into overlapping semantic units (OSUs) -- groups of related steps in which one step may belong to several units -- help retrieval over flat or disjoint alternatives? We instantiate this in OSU-Mem, which retrieves from an overlapping OSU pool via budgeted coarse-to-fine expansion, and show its benefit is conditional: overlapping memory helps when the evidence steps a query needs share tool calls or entities, but hurts when those steps are fully heterogeneous and share neither. On a synthetic benchmark where evidence carries such shared structure by construction, OSU-Mem improves over the strongest baseline as the theory predicts; yet on a concatenated, constructed unaugmented $τ$-bench setting its aggregate advantage over flat retrieval vanishes. Splitting queries by whether their evidence shares tools and entities shows this near-tie to be an artifact of mixing query types rather than a property of either method, and ToolBench, a controlled probe built to carry shared structure by design, corroborates the same mechanism via an overlap-vs.-disjoint construction contrast (under a coverage-guided variant), isolating the construction principle rather than validating the full default system. Because the relevant sharing is cheaply estimable from metadata, the analysis yields a metadata-based heuristic for predicting when overlap is likely to improve retrieval. We deliberately isolate the retrieval layer, assessed by retrieval quality and an LLM-mediated evidence-selection stage.

Local LLM agents such as OpenClaw and Nanobot run on end-user machines and act on host resources - the shell, filesystem, browser, stored credentials, and messaging applications - through natural-language goals. These agents have become privileged software runtimes that mediate between user intent, model outputs, and host-level actions. Existing research characterizes the landscape through prompt injection, malicious skills, marketplace risks, or black-box evaluation of agents. But the implementation layer that performs this mediation, the prompt builder, parser, tool dispatcher, skill loader, memory writer, network client, and permission gate, has remained an unexamined safety boundary. To our knowledge, no prior work has examined the agent's source tree to audit these components for implementation-level security weaknesses. We present CLAWAUDIT, a static auditing framework for measuring vulnerability exposure in local LLM agent runtimes. CLAWAUDIT derives a five-category vulnerability taxonomy from STRIDE and develops custom static-analysis rules that target agent-specific patterns absent from established rule sets for vulnerability analysis. We instantiate the taxonomy in two backends, 47 Semgrep YAML rules and 30 CodeQL queries, and evaluate on OPENCLAWBENCH, a benchmark of 446 source-code-level advisories from the OpenClaw repository and split temporally into 229 rule-derivation (train) and 217 held-out (test) advisories. On the held-out test, CLAWAUDIT raises Semgrep recall from 21.7% (Pro baseline) to 66.8%, and CodeQL recall from 13.8% (security-extended) to 75.1%. Train/test gaps remain within 4 percentage points for all four configurations, indicating that the rules generalize to vulnerabilities unseen during rule writing. A preliminary live-code audit shows that these recall-oriented rules require manual triage, motivating semantic filtering before production deployment.

When large language models serve as evaluators in multi-agent systems, their strategy preferences -- whether induced by explicit prompts or by shared architectural priors -- propagate through the agent network. We introduce Contagion Networks, a formal framework for measuring how evaluator preferences spread across interacting LLM agents. In a controlled 3-agent experiment using DeepSeek-chat with three distinct evaluator preference profiles (structured, balanced, evidence-based), we measure the Cross-Agent Contagion Matrix Gamma_3 and find that preferences consistently propagate between agents (gamma in [0.157, 0.352]). A neutral-prompt control experiment reveals a counter-intuitive result: shared architectural priors dominate explicit preference prompts as the driver of contagion (rho_neutral = 1.498 vs. rho_mixed = 1.299; prompt contribution: -63.5%). We identify three propagation regimes governed by the spectral radius rho(Gamma_N) and demonstrate that the same agents suppress preference contagion in chain topology (beta_3 = 0.0126 +/- 0.0038, 95% CI [0.0089, 0.0163], n=4 seeds) but cascade in fully-connected topology (Delta H_avg = -0.020) -- a topology-dependent regime transition validated both for homogeneous and cross-model agent pools (rho^cross = 1.296 +/- 0.016, n=4). We show that increasing evaluator committee size from k=1 to k=3 reduces effective contagion by 68.9% +/- 14.1% (n=4 seeds), providing an actionable mitigation strategy. We release the open-source Contagion Network experimental framework.

Large Language Models (LLMs) show promise for code compilation tasks, but applying them to runtime performance tuning is difficult due to complex microarchitectural effects and noisy runtime measurements. We present AutoPass, a multi-agent framework for compiler performance tuning that uses compiler and runtime evidence to guide LLM-generated optimization decisions. Rather than treating the compiler as a black box like prior auto-tuning schemes, AutoPass opens up the compiler to the LLM, enabling it to query compiler-internal optimization states and analyze the intermediate representation to orchestrate compiler options. The search process iteratively refines optimization configurations using measured runtime feedback to diagnose regressions and guide latency-improving edits. AutoPass operates in an inference-only, training-free setting and requires no offline training or task-specific fine-tuning, making it readily applicable to new benchmarks and platforms. We implement AutoPass on the LLVM compiler and evaluate it on server-grade x86-64 and embedded ARM64 systems. AutoPass outperforms expert-tuned heuristics and classical autotuning methods, achieving geometric-mean speedups of 1.043x and 1.117x over LLVM -O3 on x86-64 and ARM64, respectively.

We present Phoenix, a multi-agent LLM system that resolves GitHub issues from triage through pull-request creation, combining seven layered safety controls with a baseline-aware test evaluation strategy. Phoenix decomposes the work across six specialized agents. Planner, reproducer, coder, tester, failure analyst and Pull Request (PR) agent, all coordinated by a label-based GitHub webhook state machine. Every change is checked against a baseline test run before a pull request is opened. On a 24-instance slice of SWE-bench Lite. run on the production webhook path, Phoenix oracle-resolves 75% of instances with no pass-to-pass regressions on successful runs; this curated slice is not directly comparable to full-split leaderboard results, and we discuss the limits of the comparison. A complementary pilot on 42 real issues across 14 repositories yields 100% correctness preservation (CP; mean 122s on the hard tier). Manual inspection shows that about half of the resulting pull requests are well-targeted fixes. The other half place code at incorrect paths, a planner localization limitation we are addressing with retrieval. We also report the deployment failure modes (WAF filtering, token expiry, permission boundaries, flaky CI) that motivated each safety mechanism.

We propose an AI economist agent for economic and financial scenario analysis. Scenario design often requires analysts to assess emerging risks with limited historical precedent, combine information from many sources, and translate qualitative mechanisms into internally consistent quantitative paths. Large language models (LLMs) can search and synthesize this information, but fluent narratives alone do not establish the model-based calculations needed for economic conclusions. Our framework uses LLM agents to plan the analysis, retrieve relevant evidence, and organize economic mechanisms, while registered quantitative models generate numerical outcomes and predefined tests determine whether intermediate results can be used in the final report. We apply the framework to European macro-financial stress scenarios and bank capital analysis. The empirical analysis evaluates retrieval of economic mechanisms, scenario construction, model execution, and report generation under a historical information cutoff. The results show how the AI economist agent can combine flexible evidence retrieval and scenario construction while keeping the resulting analysis linked to identifiable sources and explicit model calculations.

As LLM agents increasingly select tools autonomously, their choices among tools with different privileges become safety-relevant. However, prior tool-selection studies focus on safety-agnostic metadata preferences, leaving privilege-sensitive choices underexplored. To address this gap, we study over-privileged tool selection, in which an agent selects or escalates to a higher-privilege tool despite a sufficient lower-privilege alternative. We introduce ToolPrivBench to evaluate whether agents choose higher-privilege tools despite sufficient lower-privilege alternatives, measuring both initial selection and escalation after transient tool failures. Across eight domains and five recurring risk patterns, we find that over-privileged tool selection is common among mainstream LLM agents and is further amplified by transient failures. We further find that general safety alignment does not reliably transfer to least-privilege tool choice, while prompt-level controls provide only limited mitigation under transient failures. We therefore introduce a privilege-aware post-training defense that teaches agents to prefer sufficient lower-privilege tools and escalate only when necessary. Our mitigation experiments show that this defense substantially reduces unnecessary high-privilege tool use while preserving general capabilities.

Industrial automation is being transformed by digitalization and the increasing use of cyber-physical systems. Modern production environments require greater adaptability, faster reconfiguration, and more intuitive human-machine interaction. However, traditional rule-based systems rely on fixed logic and cannot autonomously adapt to changing conditions. Consequently, current automation systems lack a systematic approach for integrating adaptive and generalizable reasoning capabilities for interpreting, planning, and executing user tasks across dynamic environments and heterogeneous components. This dissertation proposes a three-layer framework that integrates large language models (LLMs), digital twins, and automation systems into an autonomous system. Autonomy is defined as a design property assigned to system components and enabled through LLM-based reasoning to achieve adaptive, goal-oriented behavior. The Task-Process-Service-Resource (TPSR) model is introduced to transform user tasks into executable processes. Four LLM roles are identified: process orchestration, service matching, digital resource generation, and agent-as-a-service. Five peer-reviewed studies develop and refine these concepts using the design science research methodology. Case studies and prototypes demonstrate adaptive task planning, event-driven control, simulation-based parameterization, and digital model generation. Results show high task executability, command correctness, and content-generation accuracy while reducing manual effort. The framework enables the integration of LLM-based reasoning into industrial automation systems and improves adaptability and usability. Limitations include dependence on accurate digital representations, the computational demands of LLMs, and the need for human intervention in safety-critical situations.

Large language models (LLMs) demonstrate strong reasoning and generation abilities, but their fixed context windows limit long-term information accumulation and reuse across multi-session interactions. Existing memory-augmented systems often construct memory in a coarse and unstable manner, relying on inefficient memory representations or unstable unconstrained updates. To address these challenges, we propose AtomMem, a long-term memory system designed for value-dense storage and stable memory evolution. AtomMem introduces a Fact Executor, which selectively extracts high value atomic facts from long form interactions to serve as highly efficient memory representations. Subsequently, AtomMem organizes these facts into hierarchical event structures and temporal profiles, capturing coherent episodic contexts and tracking dynamically evolving user attributes over time. During retrieval, the system activates an associative memory graph to connect fragmented memories. Experiments on the LoCoMo benchmark confirm that AtomMem achieves state-of-the-art performance across various reasoning tasks, offering a scalable and economically viable solution for deploying intelligent personalized agents.

Large language models are increasingly deployed as autonomous agents for multi-step tasks in executable environments, yet their ability to perform realistic operations research (OR) work remains unclear. Existing OR evaluations often decouple modeling from solving, rely on pre-formalized or text-only instances, and rarely test the full workflow from operational artifacts to validated decisions. In this work, we introduce ORAgentBench, an execution-grounded benchmark for evaluating autonomous agents on challenging end-to-end operations research tasks. It contains 107 human-reviewed tasks across diverse operational scenarios, each packaged in an isolated environment with a natural-language brief, multi-file data, configuration artifacts, and a required submission schema. Agents must write and run solution code, and their submissions are evaluated by hidden validators for schema validity, hard-constraint feasibility, and normalized objective quality. Experiments with fourteen frontier agent-model configurations show that current agents remain far from reliable OR practice. The best agent passes only 35.51% of all tasks and 20.59% of hard tasks, and many feasible submissions still fall below the required quality threshold. Failure analysis further shows that errors are dominated by strategic weaknesses, including missed operational rules, brittle formulations, weak feasible-solution construction, and insufficient solution improvement. OR-specific procedural skills increase hard-task feasibility, but do not reliably improve solution quality or pass rate. These results suggest that progress in OR agents requires moving beyond plausible optimization code toward dependable, high-quality operational decision-making.

Agent benchmarks are growing fast, but no single benchmark touches more than four or five of the dimensions that deployment exposes. This paper aggregates the largest coordinated deep-dive of one MCP-based industrial-agent benchmark to date: fourteen parallel implementation studies covering new asset classes (including a multi-modal visual extension), alternative orchestrations, retrieval strategies, reasoning modes, infrastructure optimizations, and evaluation-methodology probes. Consolidating those studies with seven prior agent benchmarks, we argue that aggregate-score leaderboards systematically underspecify deployed-agent evaluation. Rankings derived from aggregate scores do not transfer to out-of-distribution settings; recent public-to-hidden competition retrospectives provide direct empirical evidence of this rank instability. We propose ranking configurations by predictive validity, the correlation between in-sample and out-of-sample rank, rather than in-sample mean, and report a twelve-tier measurement apparatus that exposes the deployment-relevant dimensions HELM and its agent-era successors collapse. The position is operationalized through three falsifiable out-of-distribution criteria with explicit thresholds; existing evidence partly supports it but is too thin to confirm. We close with a pre-registered pilot design and a field-level vision for what the next generation of agentic benchmarks should report.

Recent position papers argue that the classical aleatoric/epistemic uncertainty framework is insufficient for interactive large language model (LLM) agents and call for underspecification-aware, decomposed, and communicable uncertainty representations that can unlock new agent capabilities such as proactive clarification seeking and shared mental-model building. Practical deployment constraints -- black-box APIs, interactive latency budgets, and the absence of labeled trajectories -- rule out logprob-based, multi-sampling, and training-based methods, leaving prompt-based estimation as the most viable family for surfacing such signals at deployment time. We answer this call with a simple prompt-based decomposition that separates action confidence from request uncertainty (u), enabling the agent to ask for clarification when the task specification is ambiguous. To evaluate it, we introduce two clarification-augmented benchmarks (WebShop-Clarification and ALFWorld-Clarification) in which 50% of tasks are deliberately underspecified, and systematically compare the proposed decomposition against ReAct+UE and Uncertainty-Aware Memory (UAM) across five LLM backbones (GPT-5.1, DeepSeek-v3.2-exp, GLM-4.7, Qwen3.5-35B, GPT-OSS-120B) on these variants together with the standard WebShop, ALFWorld, and REAL benchmarks for fault detection. Averaged across the five backbones, the proposed decomposition improves clarification F1 on ALFWorld-Clarification by 73% over ReAct+UE and by 36% over UAM, and leads clarification F1 on every backbone on WebShop-Clarification and on four of five backbones on ALFWorld-Clarification, indicating that the gains generalize beyond a single LLM.

As large language models (LLMs) advance and multi-agent systems aim to overcome the limits of standalone agents, robust communication protocols are becoming essential infrastructure for distributed agent networks. Nonetheless, the fragmented protocol landscape presents a significant interoperability challenge. This study develops a technical taxonomy to classify and analyze LLM agent communication protocols. Following an established iterative method, we defined the taxonomy's purpose, meta-characteristic, and ending conditions, then performed five iterations, three empirical-to-conceptual and two conceptual-to-empirical, on nine actively maintained open-source protocols with demonstrable adoption. The taxonomy comprises five dimensions: counterparty, payload, interaction state, discovery mechanism, and schema flexibility. Classification reveals recurring architectural patterns: all sampled agent-to-agent protocols combine hybrid payloads with session-state persistence; most protocols support multiple predefined schemas, and two negotiate schemas at runtime, indicating a trend toward schema flexibility; decentralized discovery remains rare. Analysis suggests short-term convergence pressure toward protocols unifying agent-to-agent and agent-to-context (tool and data) communication. Long-term, however, no single protocol is likely to maximize versatility, efficiency, and portability simultaneously. The field will more likely evolve toward a federated, layered protocol stack. The framework guides protocol selection and highlights open research gaps such as privacy and policy enforcement.}

Team science holds that leadership is contingent: it helps only under specific conditions, and capable, autonomous teams may need none at all. We ask the analogous question for multi-agent LLM teams: under what measurable conditions does process-level coordination control add value, and do those conditions match what team science predicts? We use behavioral signatures (majority lock-in, exploration, recovery from an incorrect round-0 consensus) and per-action ablations, clean because each controller is an explicit action set, not a monolithic prompt. We operationalize three classical leadership styles (transactional, transformational, situational) as controllers over a shared action vocabulary (explore, revise, accept, synthesize). A matched controller with the same actions but an arbitrary rule recovers no better than majority voting, so the theory-derived rule, not the vocabulary, does the work. Across four task regimes and three open-weight model families, no controller dominates by accuracy, as the contingency view predicts: transactional control matches a shared round-0 vote on all 12 (model, regime) combinations to within 1.3pp, and gains appear only on the one combination where the round-0 majority is unreliable (llama-4-scout social; situational +8pp over flat). A recovery-advantage account, tested with four boundary probes, says a controller beats plain interaction only where the round-0 majority is unreliable, the task is recoverable, and undirected interaction does not already repair it. These regions map onto contingency theory (leadership substitutes, path-goal redundancy, the situational readiness gap), so a largely null accuracy result is what the theory predicts, not a failure of the controllers. We read process-level coordination control as a contingency to be measured and theory-mapped, not a leaderboard to be topped.

Automated assessment in software engineering education has advanced significantly for code grading and essay scoring. However, reviewing software architecture deliverables, which requires analyzing structural completeness and requirements traceability, has not yet been fully automated. Applying Large Language Models (LLMs) to this task requires robust architectures to ensure technical feedback is accurate and reliable for students. This paper presents CAPRA (Configurable Architecture Proficiency Report Assessment), a multi-agent LLM system that analyzes software architecture deliverables to generate personalized, template-compliant LaTeX feedback. As a core design choice, CAPRA coordinates multiple specialized agents and employs a Python-based microservice for multi-modal document extraction, utilizing PyMuPDF and vision-enabled LLMs (specifically gpt-4o) to parse text and UML diagrams. To ensure educational reliability and mitigate hallucinations, CAPRA introduces a deterministic Evidence Anchoring step using fuzzy matching via normalized Levenshtein distance, along with a ConsistencyManager agent that cross-verifies, deduplicates, and merges findings. System performance is assessed using a structured eight-criterion binary evaluation taxonomy covering: (i) extraction completeness, (ii) feature validation, (iii) issue grounding and severity detection, (iv) recommendation specificity and traceability, and (v) template and tone compliance. A preliminary empirical evaluation on 10 student reports shows that CAPRA satisfied 88.8% of the evaluated criteria under a strict two-rater aggregation rule, achieved moderate inter-rater agreement with human evaluators (kappa = 0.582), and processed each report in slightly over 4 minutes. While these results support the viability of LLM-supported architectural feedback, human oversight remains essential for subjective assessment dimensions.

Production LLM agents increasingly depend on real-time search, yet native search grounding bundles retrieval policy, provider choice, evidence injection, cost, latency, and generation behavior behind a single model-provider boundary. This coupling makes grounding hard to inspect, tune, reuse, or port, and can trigger Search-Induced Verbosity that breaks strict output contracts. We present Decoupled Search Grounding (DSG), a vendor-agnostic boundary that moves grounding outside the reasoning model through an MCP-compatible gateway, exposing provider routing, source-aware context rendering, configured fallback, retrieval-depth control, and exact plus semantic caching as first-class controls. Across five frontier models on SimpleQA, FreshQA, and HotpotQA, native search leads on recency-sensitive FreshQA, but DSG exposes a stronger frontier when control matters: on SimpleQA it nearly matches native accuracy (86.1% vs. 87.7%) at 91% lower search cost, preserves concise answer contracts, and reaches a 99.4% warm-cache hit rate with 68% lower latency. Deployed as a shared production grounding layer for large-scale agentic workloads with interchangeable models, DSG matches or slightly exceeds native-search accuracy on an e-commerce query-understanding (QIU) workload while cutting search cost by over 98%. Real-time grounding is best treated as an optimizable interface boundary, not a fixed model feature.

RL post-training strategies are dataset-dependent and reveal a recurring empirical pattern: capacity parameters accumulate monotonically across stages, while regularization parameters predominantly oscillate in response to shifting training dynamics. This distinction highlights a potential flaw in fixed training schedules: by forcing all parameters along rigid paths, they fail to capture the dynamic exploration-exploitation tradeoffs that regularization must track. We uncover this through LLMZero, an agentic system that optimizes training trajectories via tree search by diagnosing pathologies at each checkpoint and proposing coordinated multi-parameter transitions. Across four diverse GRPO tasks, LLMZero discovers strategies that improve over the base model by 9% to 140% and over grid search by 6% to 15% (relative), consistently outperforming random search and a skill-based agent under a matched compute budget. The capacity--regularization asymmetry is consistent across all four tasks, offering a candidate design heuristic for multi-stage training.

Reproducing research results from papers and released code is central to scientific progress. Existing works have introduced benchmarks to evaluate whether LLM agents can assist with reproducibility, but they are difficult to scale due to their reliance on substantial manual effort for data curation and evaluation. We introduce ReproRepo, a scalable framework for reproducibility evaluation that leverages human-raised GitHub issues as naturally occurring supervision on realistic reproduction blockers. We instantiate ReproRepo on 1,149 recent machine learning papers from major conferences and evaluate four frontier model-agent configurations. Our results show that LLM agents, even without executing code, can identify many real-world reproducibility problems from paper-repository pairs: the best agent in our study, namely Codex with GPT-5.5, surfaces at least one semantically related human-reported blocker for $\sim$90% of papers in the study. Further analysis shows that agents are particularly effective for surfacing visible failures and identifying the right semantic region, but may still be insufficient in exact localization. ReproRepo can serve as a reusable, scalable framework for future evaluations of LLM agents on real-world reproducibility auditing. Our code is released at https://github.com/LithiumDA/ReproRepo.

LLM agents increasingly rely on external skills -- reusable tool specifications -- but real-world tasks often require composing multiple skills, not just selecting one. We formalize this as the Compositional Skill Routing problem: given a complex user query and a large skill library, decompose the query into atomic sub-tasks, retrieve the appropriate skill for each sub-task, and compose an executable plan. We present SkillWeaver, a decompose-retrieve-compose framework combining an LLM task decomposer, a bi-encoder skill retriever with FAISS indexing, and a dependency-aware DAG planner. To support evaluation, we introduce CompSkillBench, a benchmark of 300 compositional queries over 2,209 real MCP server skills spanning 24 functional categories, sourced from the public MCP ecosystem. Our experiments reveal that task decomposition quality is the primary bottleneck: standard LLM decomposition reaches only 34.2% category recall at the step level. To address this, we propose Iterative Skill-Aware Decomposition (SAD), a retrieval-augmented feedback loop that iteratively aligns decomposition with available skills. SAD improves decomposition accuracy from 51.0% to 67.7% (+32.7%, Wilcoxon p < 10^-6) in a single iteration; DA-conditioned analysis confirms that correct granularity is the prerequisite for effective retrieval (CatR@1 rises from 34% to 41% when DA=1). SkillWeaver reduces context window consumption by over 99%, and transfer experiments confirm generalization (+35.6% relative DA gain even when target categories are absent from the retrieval pool).

Agent skills -- structured, reusable knowledge artifacts that augment LLM agent capabilities -- have been rapidly adopted in industry, yet their cross-domain impact and use across commercial and open-source models remain under-studied, and no reusable methodology exists for evaluating an individual skill. In this work, we present an evaluation framework that lets a skill author construct realistic tasks to rigorously assess the aspects of a skill that matter most to them, and that estimates skill utility by solving those tasks. Further, we apply our evaluation approach at scale to 500 real-world skills, generating 1,000 tasks derived from the skills' content, along with instruction-following and goal-completion scoring rubrics. Using these metrics, we evaluate how 19 agent-model configurations, both proprietary and open-source, perform on the tasks. Our results show that models vary widely in how closely they adhere to the instructions encoded in skills, leading to substantial differences in their performance gains. Furthermore, we show that access to a skill significantly changes model behavior compared to the no-skill setup, providing an essential mechanism for encoding opinionated workflows into LLM agents. We release our evaluation dataset to support future work on agent skills.

Self-evolving LLM-based agents improve mainly by changing their agent harness: the structured execution layer around a base model, including prompts, memory, tools, middleware, runtime state, and the model-tool interaction loop. Existing evaluations often reduce this process to isolated task scores or a single sequential curve, obscuring whether an update produces reusable improvement, overfits recent tasks, increases cost, or harms older behavior. We introduce SEAGym, an evaluation environment for measuring agent harness updates across training, validation, test, replay, and cost records. SEAGym turns Harbor-compatible benchmarks into dynamic self-evolution task sources with train batches, frozen update-validation, held-out ID and OOD transfer views, replay diagnostics, and saved snapshot and metric records. Instantiating SEAGym on Terminal-Bench 2.0 and HLE, we compare ACE, TF-GRPO, and AHE under a shared epoch/batch protocol. The results show that these evaluation views provide complementary signals about the evolution process: frequent updates may fail to improve held-out performance, useful intermediate snapshots may collapse later, and source diversity and model backend can affect harness reliability.

Prompt injection defenses evaluated on synthetic benchmarks do not generalize to real enterprise documents, which are longer, denser, and interleave legitimate authority language with factual content. We demonstrate this gap with a real-document benchmark of 122 tasks across five professional domains (financial, legal, medical, scientific, DevOps) using actual SEC filings, Federal Register rules, PubMed abstracts, arXiv papers, and GitHub postmortems. Paraphrasing, the strongest defense on synthetic benchmarks, shows no statistically significant attack success rate reduction on real documents (p=0.500) while degrading utility from 91.8% to 82.8%. We introduce PARSE (Provenance-Aware Retrieval Sanitization), a domain-aware, fact-preserving sanitization pipeline that classifies each sentence by injection likelihood, extracts structured facts before rewriting, and verifies fact preservation via a consistency-checking loop. A directiveness gate routes 59% of real enterprise documents to a lightweight path, concentrating computational cost on high-risk documents. PARSE achieves 15.6% attack success rate -- a 38% reduction versus the 25.4% baseline -- at 86.9% utility, the only condition that is both statistically significant (p=0.014, adequately powered) and maintains near-baseline utility. Practitioners should evaluate defenses on domain-matched real documents, not synthetic proxies.

Large language models (LLMs) have achieved remarkable success in software development. However, they are susceptible to hallucinations, meaning that they can introduce subtle semantic and logical errors. Due to the high stakes in chip design and manufacturing, hardware engineers are still reluctant to rely on LLMs for register-transfer level (RTL) generation. In this paper, we propose a hardware generation framework that combines the creativity and broad knowledge of LLMs with the explainability and mathematical rigor of formal methods. Specifically, we devise a set of transformation rules that cover various design decisions and hardware features. By iteratively applying these rules, an LLM agent can convert a design specification into an RTL program with guaranteed correctness. Experimental results demonstrate the effectiveness and efficiency of the framework.

Multi-agent LLM systems share state through memory stores, vector indices, and tool registries. We model such sharing as long-running read-generate-write operations under deterministic-generation semantics -- the regime durable-execution engines enforce by deterministic replay -- and formalize four concurrency anomalies in TLA+: stale-generation, phantom-tool, causal-cascade, and tool-effect reordering, structural analogues of classical isolation anomalies, each with a TLC counter-example. The exclusion lattice over these anomalies is trivial; the contribution is the mechanically verified realizability and strict separation of one maximal chain within it, $L_0 \subsetneq \cdots \subsetneq L_4$, to our knowledge the first machine-checked consistency hierarchy for such runtimes. A development of 274 Verus obligations (zero assume, zero admit; trust base: two structural axioms and a mutex correspondence) proves the detectors sound and complete against the specifications and each runtime its avoidance set. Three deployed Rust runtimes realize L0-L1 (pessimistic locking, serializable snapshot isolation, default-SI), each verified against stale-generation and refined to its state machine; L2-L4 are exec-mode-verified with dependency-free prevention twins (A3, A6, A2: 0/1000 versus 1000/1000), and L2 is run live across three model families (A3 prevented in all 120 retracted sessions). We reproduce a silent lost update in ByteDance's deer-flow, formalizing its fix as a verified $L_0 \to L_1$ refinement, and exhibit tool-effect reordering in LangGraph's ToolNode on unmodified output, removed by an L3 commit-order sequencer. The verified detector, refinements, and realizability artifacts are the contribution; the phenomena and lattice are classical.

As LLM agents are deployed in long-horizon sessions, context accumulation drives up inference costs. Existing approaches utilize text pruning or dynamic memory eviction to minimize token footprints; however, their unconstrained sequence mutations alter layouts, introducing prefix mismatches and cache invalidation. This reveals a critical trade-off between text sparsity and prompt cache continuity. To address this, we present TokenPilot, a dual-granularity context management framework. Globally, Ingestion-Aware Compaction acts as a framework harness to stabilize prompt prefixes and eliminate open-world environmental noise at the ingestion gate. Locally, Lifecycle-Aware Eviction monitors the ongoing residual utility of context segments, enforcing a conservative batch-turn schedule to offload content segments only when task relevance expires. Experiments on PinchBench and Claw-Eval under both isolated and continuous modes demonstrate that TokenPilot reduces costs by 61% and 56% in isolated mode, and 61% and 87% in continuous mode, while maintaining competitive performance compared to prior systems. TokenPilot has been integrated into LightMem2 at https://github.com/zjunlp/LightMem2.

Tool-augmented LLM agents rely on runtime filtering to decide which tools should be visible at each step. Causal Minimal Tool Filtering (CMTF) reduces tool-choice confusion by exposing only the next causally necessary tool frontier, but it assumes that the user request has already been mapped to a symbolic goal state. In practice, requests such as "handle my appointment" or "take care of this email" may correspond to multiple possible goals. This creates wrong-goal execution, where an agent follows a valid causal tool path for an unintended objective. We introduce GIST-CMTF, a goal-state inference layer that predicts candidate symbolic goals over the same state-transition vocabulary used by CMTF, estimates ambiguity, and either applies CMTF or exposes clarification as a causal action that produces missing goal or state variables. We evaluate GIST-CMTF across seven model backends, six filtering methods, and 120 controlled tool-use tasks. GIST-CMTF achieves 97.0% task success, compared with 80.1% for top-goal CMTF and 82.9% for semantic-goal CMTF. It reduces wrong-goal execution from 19.4% under top-goal CMTF to 2.5%, while preserving the one-tool exposure of causal filtering and using substantially fewer tokens than all-tools exposure. These results suggest that reliable tool-augmented agents should validate goal state, not only tool relevance, before exposing external actions.

Agent skills are commonly distributed as SKILL.md files: human-readable procedural documents that describe workflows, tools, resources, and domain conventions. While convenient for inspection and reuse, this design requires the same reusable procedure to be repeatedly injected into the runtime context. We propose Skill-to-LoRA(S2L), a behavior-centric skill representation that replaces runtime skill text with skill-specific LoRA adapters. Rather than compressing the skill document itself, S2L models the behavioral change induced by the skill text: offline, the complete SKILL.md is used to synthesize skill-guided demonstrations; online, the full document is omitted and the corresponding LoRA adapter is dynamically loaded to activate the learned skill behavior. We evaluate S2L with Qwen3.6-27B on a 21-skill subset of SWE-Skills-Bench. Compared with the no-skill and Full Skill Text baselines, S2L improves pass rate by 2.9 and 5.2 percentage points, respectively, while reducing per-step token cost by 6.6% relative to Full Skill Text prompting. S2L matches or improves Full Skill Text on 18/21 skills and the no-skill baseline on 15/21 skills. Control experiments further show that the gains depend on skill-specific adapter alignment: Wrong-LoRA and Shared-LoRA both reduce performance. These results suggest that many procedural agent skills can be converted from runtime instructions into trainable, dynamically loadable behavioral modules. Code will be released upon acceptance.

Large language model (LLM) agents increasingly take actions (screening applicants, recommending credit, triaging patients), yet fairness for LLMs is still measured by grading answers. We introduce AgentFairBench, a cheap, reproducible, multi-domain benchmark for demographic disparity in the actions of LLM agents. Grounded in a companion framework, the Bias Conduction Framework (BCF, restated here), it spans three regulator-anchored domains: hiring, lending, and medical triage. Synthetic, demographic-neutral profiles are evaluated in counterfactual matched sets that vary only a name-coded race x gender signal (in the Bertrand Mullainathan tradition), under four agent scaffolds of increasing agency (direct, chain-of-thought, multi-agent deliberation, tool-augmented). A NumPy-only harness computes counterfactual flip rate, mean absolute score difference (MASD), action-rate disparity, and tool-invocation disparity, with bootstrap confidence intervals, paired tests, and false-discovery-rate control, for single-digit dollars per model. A live leaderboard with a held-out private split and a contamination canary admits external models by submission. Our pilot (864 decisions plus a test-retest replication) carries a methodological lesson: comparing a six-group score spread against a two-run noise difference overstates disparity by ~ 2.4X through statistic arity alone. Against an arity matched noise floor and an omnibus group test, claude haiku 4 5 shows no demographic effect above sampling noise (0 of 120 pairwise and 0 of 9 omnibus contrasts survive correction); a planted-bias test confirms the instrument detects disparity when present. The contribution is a sound, sensitive, adoption-ready instrument, the arity matched null methodology, and open artifacts to scale it. Code, data, and harness are released under open licenses, with an anonymized review artifact.

As LLM agents become capable of increasingly long-horizon tasks, evaluating their performance in economic systems is becoming increasingly important. Unlike existing benchmarks that primarily evaluate a single agent interacting with a passive environment, economic systems are inherently multi-agent, requiring autonomous agents to communicate, negotiate, and transact while pursuing their own objectives over extended periods. We introduce CoffeeBench, a benchmark for evaluating LLM agents in a long-horizon multi-agent economy composed of heterogeneous firms. In CoffeeBench, two farmers, two roasters, and two retailers autonomously operate their businesses over a 90-day simulation, each seeking to maximize cumulative net income through communication and transactions while managing cash, inventory, and pricing. The evaluated model controls one coffee roaster, while the remaining firms are controlled by fixed reference agents. Across several recent open-weight and proprietary LLMs, all models outperform a passive baseline that takes no actions, with most achieving positive net income. Analysis of agent behavior reveals substantial differences in long-horizon economic interaction: higher-performing models communicate more actively with other firms, whereas ClaudeHaiku4.5 exhibits an idle-drift failure mode, repeatedly choosing inaction despite producing coherent assessments and plans. We release our code and agent trajectories to support future research.

Large language model (LLM) agents increasingly rely on agent harnesses that manage context, tools, and multi-turn execution, making tools a central interface for acting in realistic digital environments. As harness-connected tool ecosystems expand to hundreds or thousands of APIs, services, and task-specific skills, exhaustive tool schema injection becomes costly and imposes a closed-world assumption that limits agents to a predefined static inventory. Retrieval-augmented tool selection offers a natural alternative, but existing one-shot retrieval methods often fail to align isolated tool descriptions with the agent's true task intention, especially in long-horizon tasks where required capabilities emerge through decomposition, observations, and newly induced subgoals. We propose SING, an intention-aware active tool discovery framework that builds an intention-tool graph linking user intentions, tool capabilities, and tool collaboration patterns, and dynamically retrieves tools according to evolving task states. Using a unified corpus of 7,471 tools, we evaluate SING on three real-world tool-use benchmarks. SING improves Global Recall@5 by up to 59.8% and downstream success rate by up to 28.9% over baselines, while reducing full-corpus tool-schema exposure by 99.8%, demonstrating that intention-aware graph structure enables more accurate and context-efficient tool discovery in large-scale agentic ecosystems.

Large language models (LLMs) are increasingly being used for automated decision-making systems in finance, healthcare, or environmental monitoring. Time series data are ubiquitous in these fields, yet hard to process automatically. Can time series be analyzed by LLM agents? We examine three approaches: providing the agent with raw numerical data, using the LLM as a coding agent, or a combination of both. In the coding agent setup, the model iteratively queries the data using Python code. Using two time series understanding benchmarks, we show that agents with code access can outperform models processing raw data by up to 10%. However, even the best performing agent still answers about 22-34% of the questions incorrectly. To get insights into models' strategies and reasoning gaps, we analyze the model outputs with a strong LLM judge. Our analysis reveals that coding agents can select appropriate statistical tests, but often miss important nuances. Meanwhile, models with access to raw data can reach the right conclusions using back-of-the-envelope calculations.

AI agents are increasingly being adopted in enterprise and personal settings with access to emails, databases, documents, and other tools where they can read, update, and disseminate sensitive information. Much of prior research on data leakage risks in agents has focused on adversarial data exfiltration through prompt injections and jailbreaks. However, sensitive information may also be exposed during non-adversarial use, creating leakage risks even when users issue benign requests. We report a joint evaluation by the Singapore AI Safety Institute and the Korea AI Safety Institute examining agent data leakage in 12 realistic, non-adversarial tasks spanning customer support, DevOps, web automation, and enterprise and personal productivity. The evaluation covers five risk types: lack of data awareness, audience awareness, policy compliance, data minimization, and access-boundary awareness. Both institutes tested a common set of scenarios mirroring real-world deployments using independent testing environments and task-specific LLM-judge rubrics. Across the three tested agents, none achieved fully correct and fully safe execution across all scenarios. Successful task completion often coincided with data-handling failures such as accessing unnecessary information or disclosing information to inappropriate recipients, indicating that capability and data-handling safety should be evaluated separately. Qualitative review also revealed claim-action mismatches, simulation-aware behavior, user-simulator role reversal, and interpretation gaps in automated judging. Overall, the results indicate that operational data leakage is a first-order agent-safety concern distinct from adversarial exfiltration and provide a methodology for future evaluations of agent data-handling safety.

Multi-agent LLM coordination papers report small benchmark deltas as evidence that one architecture beats another. A prior question: how much paired trial-0 disagreement do two protocols produce on the same model and benchmark when their API inputs are configuration-equivalent (matched by code inspection plus a SHA-256 byte audit), short of full identity-replay? On Claude Haiku 4.5 against tau^2-bench retail, the clean configuration-equivalent contrast (no_coord vs. intercept, both inert at trial 0) gives signed paired gaps of +10pp and 0pp across two n=100 seeds; pooled across both, +5pp with Wilson CI [-2,+12], not significant. The largest single-seed contrast (+18pp pull-vs-intercept, p_corr=0.012) did not reproduce at the second seed (-3pp, p_corr=1.0); no trial-0 contrast is significant after Bonferroni at either seed or pooled. The envelope of observed paired gaps spans [-3,+18]pp across two seeds, with pooled upper Wilson CI ~15pp. Seven of ten recent multi-agent coordination architectures report headline effects below this local floor, and one more sits inside the envelope; whether they survive a same-model paired replication is, by construction, untested in their original settings. We define coordination-active pass^k, pass^k restricted to trials where the coordination mechanism is logically active, as the minimum reporting protocol, with sample-size targets and runtime hooks in the body. Measurements run on ET-MCP, a task-scoped negative-knowledge store conformant with MCP 2026-07-28, used as a substrate to isolate reader-side choices, not as a contribution. On Haiku 4.5 the candidate readers (pull, intercept) do not improve trial-1 recovery; we give a preliminary diagnosis of failure modes with refinements on existing production hook surfaces.

Open-source LLM agent ecosystems are growing rapidly, yet the security of community-contributed skills - modular tool definitions that extend agent capabilities - remains largely unvetted. The gap we fill: existing scanners operate at the code layer and are structurally blind to instruction-layer and multi-agent risk - natural-language directives that hijack an agent, exfiltrate data through encoded side channels, or chain harm across pipelines - so what is needed is a semantic, multi-dimensional vetting system rather than another signature matcher. We present SKILLVETBENCH, a live public leaderboard on Hugging Face that uses an LLM-as-Judge to vet agent skills. What is new: SARS (Skill Agentic Risk Score), a five-dimensional agentic-risk metric with a principled weighted formula for instruction-following systems. What is integrated: full CVSS v4.0 vector decomposition and a ClawHub dual-view that places our LLM-generated review beside the official marketplace verdict. What is demonstrated: drawing on our companion benchmark paper [ 1], the LLM-as-Judge stage achieves zero false negatives across 78 confirmed-malicious skills and zero false positives across 22 benign controls, while the best static baseline (SKILLSIEVE) still misses 15%; for instruction-layer categories such as Prompt Injection and Memory Poisoning, conventional tools miss between 89% and 100% of threats (e.g., CODEBERT detects none of nine memory-poisoning skills). Detection rates vary from 35% to 95% across four LLM evaluators, motivating ensemble scoring in production deployments.

Every major LLM agent framework gives the LLM the role of orchestrator; the model decides what to do next, when to call tools, and when to stop. We argue that token explosion, control-flow hallucination, and unreliable completion are not implementation bugs but architectural consequences of assigning the deterministic work of looping, branching, and sequencing to a probabilistic system. A better prompt or a stronger model cannot guarantee the reliability of the LLM agent. We therefore propose Agentic Programming, in which the program governs all control flow, and the LLM is itself part of it, an adaptive component we call LLM-as-Code and invoke only where a task calls for reasoning or generation. Within each call the model keeps full flexibility, but it cannot alter the program's execution path. With control in the program, the LLM's context is built from the execution history's call tree and forms a directed acyclic graph (DAG). Each call's context length is then determined by its call depth rather than by accumulation over steps. A case study of computer-use agents shows that the design is practical, not just a theoretical stance, substantially improving the stability of long visual operation sequences.

Large language model (LLM) agents have made rapid progress on short-horizon, well-scoped tasks, yet their ability to sustain coherent decisions in dynamic long-horizon environments remains uncertain. We introduce RetailBench, a data-grounded simulation benchmark for evaluating tool-using LLM agents in single-store supermarket operation. RetailBench models retail management as a partially observable decision process and is designed to support thousand-day-scale simulations. In this environment, agents must manage pricing, replenishment, supplier selection, shelf assortment, inventory aging, customer feedback, external events, and cash-flow constraints. We evaluate seven contemporary LLMs under representative agent frameworks over a 180-day evaluation horizon and compare them with a privileged oracle policy. Results show substantial variation across models: only a small subset survives the full evaluation horizon, and even the strongest LLM runs remain substantially behind the oracle policy in final net worth and sales outcomes. Behavioral analysis attributes these gaps to incomplete evidence acquisition, surface-level decision making, and the lack of a consistent long-horizon policy. RetailBench provides a controlled testbed for studying reliable autonomy in economically grounded long-horizon decision-making.

Large language model (LLM) agents increasingly rely on long-term memory to support complex task execution, user personalization, and domain adaptation. Meanwhile, emerging access-control mechanisms for LLM agents are being explored to block policy-violating requests and prevent misuse. We reveal a novel attack surface arising from agent memory operations: prohibited content that would trigger access control can be fragmented across interactions, stored in long-term memory in benign-appearing form, and later reconstructed through memory retrieval without appearing explicitly in the final user query. We propose FragFuse, the first attack that enables unprivileged users to bypass agent access control by exploiting this temporal channel introduced by long-term memory. FragFuse operates in three stages: (1) identifying rejection-responsive fragments via black-box adaptive querying with fragment masking; (2) injecting these fragments into memory using marker carrier queries; and (3) retrieving and fusing the stored fragments through a follow-up attack query. Although FragFuse can be instantiated manually for individual agents, we further develop a surrogate-based optimization scheme that tunes fusion instructions and marker designs, enabling automated attack generation without violating the attacker's threat-model assumptions. We evaluate FragFuse across four representative agent settings and task domains, covering three state-of-the-art agent access-control mechanisms. FragFuse achieves an average bypass success rate of 86.3% and an average end-to-end harmful task success rate of 41.1% across all settings, with only 4.4% average task-success degradation compared with configurations without access control. We also show that alternative defenses, including state-of-the-art prompt-injection detectors and perplexity detectors, do not effectively address this attack.

Tool-augmented large language model agents increasingly operate over large tool libraries, but existing evaluations often focus on whether a model can call a tool correctly rather than how the visible tool menu shapes reliability, efficiency, and safety-relevant risk exposure. We introduce ToolMenuBench, a benchmark for evaluating tool-menu construction in multi-step LLM agents. ToolMenuBench varies tool-menu size, distractor type, state-dependent task structure, and risk exposure, and reports both filter-level and downstream agent metrics, including visible-tool count, risky-tool exposure, task success, wrong-tool calls, premature actions, and token usage. In a controlled evaluation across seven model backends, three tool-menu sizes, six filtering methods, and seven evaluation settings, CMTF improves task success from 32.1% under all-tools exposure to 85.7%, while reducing average token usage by roughly 98%. Causal minimal tool filtering achieves the strongest overall tradeoff, reducing visible tools, wrong-tool calls, premature actions, and risky-tool exposure relative to unfiltered exposure, lexical filtering, state-aware filtering, and broader causal-path baselines. ToolMenuBench provides a reusable evaluation framework for studying the agent-interface problem: which tools should be visible, when they should be visible, and under what cost or risk constraints.

Objective: To enhance the accuracy, interpretability, and robustness of large language models (LLMs) in medical question answering (MedQA). Method: We designed a multi-agent peer-reviewed reasoning method in which multiple LLM agents independently generate chain-of-thought reasoning with candidate answers, then act as peer reviewers to evaluate each other's reasoning for factual correctness and logical soundness. The highest-rated reasoning chain is selected to produce the final answer. Experiments were conducted with five state-of-the-art LLMs (Llama-3.1-8B, Qwen2.5-7B, Phi-4, DeepSeek-LLM-7B, GPT-oss-20B) on three benchmark datasets: HeadQA, MedQA-USMLE, and PubMedQA. Performance was compared against single-model chain-of-thought reasoning and chain-of-thought-based majority voting. Results: Peer-reviewed reasoning consistently outperformed both baselines. The best model combination achieved an average accuracy of 0.820 across datasets, exceeding the strongest single model (0.777) and majority voting ensembles (up to 0.789). The method also scaled effectively with more participating models, while peer assessments reliably distinguished high- from low-quality reasoning chains. Conclusion: The proposed multi-agent peer-reviewed reasoning method enables LLMs to act as both solvers and evaluators, yielding superior performance in MedQA. By emphasizing reasoning quality rather than answer agreement alone, this approach improves accuracy, interpretability, and robustness, offering a promising direction for trustworthy biomedical AI systems.

Multi-agent LLM systems -- coding agents, devops agents, document agents -- now routinely run several agents in parallel against the same git tree, Kubernetes cluster, or document. As soon as two of them mutate shared state, they enter the regime classical concurrency control has studied for decades, but classical mechanisms fit LLM agents poorly. A single agent transaction spans minutes of inference, read sets are broad and opaque rather than statically inferable, and the live state agents act on admits neither fork nor buffer, so writes take effect the moment they execute. Locks block long inference intervals; OCC abort-and-retry discards minutes of work on every conflict. This paper builds concurrency control on a capability classical transactions lack: the LLM inside each agent can judge whether a conflicting write invalidates its plan, and can repair exactly the operations that depended on it. Control therefore turns advisory: the runtime informs, the agent repairs. Our protocol, MTPO (Monotonic Trajectory Pre-Order), fixes a serialization order at launch, serves each read the order-filtered value, and applies writes speculatively in place; a one-way notification asks an affected reader to re-judge and patch its plan, while the framework mechanically undoes and reorders misplaced writes through the saga-style inverse each tool registers in advance. At quiescence the run is serializable in the pre-decided order. We realize MTPO as CoAgent, toolcall middleware whose privileged ToolSmith grows footprint-declared, undoable tools online. On ten contended workloads, CoAgent stays within 5% of serial correctness at a $1.4\times$ speedup and near-serial token cost, where 2PL and OCC surrender nearly all concurrency gains; on a bash-only target system, it grows a 25-tool library online and lifts the task pass rate from 45/71 to 63/71 at $0.80\times$ the time and $0.86\times$ the cost.

Indirect prompt injection (IPI) is a major security threat to LLM-powered agents. Thus, a growing body of work have proposed a variety of defensive approaches against IPI. These can be grouped into three broad categories: 1) prompt-based (using prompting as a way to prevent agents from following malicious instructions), 2) detection-based (identifying and filtering malicious instructions), and 3) system-level (using systems insights, such as control and data isolation, for defense). However, commonly used benchmarks for evaluating defense, such as AgentDojo, are \emph{inherently static}, generating a fixed distribution of IPI attacks. Consequently, static benchmarks do not usefully evaluate defense robustness to adaptive threats. We address this issue by developing AutoDojo, an adaptive extension of AgentDojo that optimizes IPI against a given defense. Using AutoDojo against state-of-the-art IPI defenses across three task suites and five target models, we make two key observations. First, many defenses offer only limited protection: a cheap, black-box adaptive attack using a frontier LLM to iteratively optimize the injection raises attack success rate (ASR) well above the level achieved by static injections against nearly all evaluated defenses. Against a filter that reduces static ASR to 0%, AutoDojo recovers 28% overall and 64% on action-open tasks. Second, for prompt-level and filter-based defenses, ASR is substantially higher on \emph{action-open} tasks -- where the user's request delegates the action itself to attacker-controlled content -- than on precisely specified tasks. This is a structural limit: on such tasks the injection can pose as ordinary data rather than an explicit instruction, bypassing defenses that rely on detecting instruction-like text. AutoDojo is publicly available at https://github.com/xhOwenMa/AutoDojo.

Frontier reasoning-tuned language models still fail on deductive tasks at depth, and the cost of improved performance through extended internal reasoning scales poorly. Symbolic delegation offers a complementary route: a language model translates the problem, while a solver performs the inference. However, current autoformalization pipelines for logic programming are typically bespoke integrations tied to particular tasks or agents. We introduce PrologMCP, a task-agnostic, open-source server that exposes Prolog as a stateful tool through the Model Context Protocol (MCP). Its compact tool interface, structured error reporting, and per-session isolation make the translate-run-inspect-repair loop a reusable primitive for MCP-capable agents. We evaluate a formalizer agent enhanced with PrologMCP against standard and reasoning LLMs (Claude Sonnet 4.6, GPT-4.1, and o4-mini) on two subsets of PARARULE-Plus: a general-purpose sample and a more challenging one targeting a specific failure mode of natural-language reasoning. On the general sample, the formalizer matches or exceeds reasoning LLMs (accuracy 1.00 vs.\ 1.00 / 0.998), with the largest gains over standard models (0.762 for GPT-4.1). On the challenging subset, the formalizer remains near-perfect (1.00 / 0.99) while reasoning LLMs drop to 0.95 / 0.94. These results suggest that delegating inference to Prolog via MCP is a robust and inspectable alternative to extended natural-language reasoning.

LLM agents are increasingly built not as single model calls, but as scaffolded systems that combine reasoning, memory, reflection, action execution, and learning. While such scaffolds often improve performance, they are often embedded in tightly coupled pipelines, making it difficult to isolate component contributions, compare alternative designs, or understand how module interactions shape agent behavior. We introduce AgentSpec, a modular specification framework that represents embodied agents as typed compositions of reusable policy components with standardized interfaces. AgentSpec standardizes the interfaces among perception, memory, reasoning, reflection, action, and optional learning, enabling components to be swapped and recombined under controlled conditions. We instantiate this framework across DeliveryBench, ALFRED, MiniGrid, and RoboTHOR, and analyze reasoning, memory, reflection, and reinforcement-learning modules across model backbones. Our results show that agent performance is governed by scaffold compatibility and interaction effects rather than isolated module strength. In particular, structured multi-granularity memory improves long-horizon state tracking, reasoning and memory interact non-uniformly across environments, reflection trades off correction and cost, and RL-trained policies compose best when optimized with deployment-time scaffold structure. AgentSpec provides a controlled foundation for studying, comparing, and designing composable LLM agents. Our code, baselines and interactive playground are publicly available at https://agentspec-embodied.github.io.

LLM agent systems increasingly run as long-lived autonomous runtimes: scheduling jobs, calling tools, maintaining memory, and pushing results to humans. We present a longitudinal study of silent failures in one such system: a personal-assistant agent runtime in continuous production since March 2026, with roughly 40 scheduled jobs, 8 LLM providers, a tool-governance proxy, and a knowledge-base memory plane, defended by 4,286 unit tests and 827 governance checks. Over eight weeks we documented 22 incidents with full root-cause postmortems, in which one meta-pattern -- a failure whose error signal never reaches a human in actionable form -- manifested at least 28 times. We derive a five-class, mechanism-oriented taxonomy: (A) environment and platform quirks, (B) design-assumption mismatches, (C) error swallowing and dilution, (D) chained hallucination and fabrication, (E) operational omission and forensic blind spots. Class D is unique to LLM systems and the most dangerous: the system does not merely fail to report an error -- the LLM transforms it into fluent, plausible narrative delivered to the user. We term this fail-plausible: gray failure's differential observability escalated -- the observer is not just blind, it is convincingly lied to by the failure itself. Three findings: about 70% of silent failures were caught by human user-view observation, not tests or audits; a retrospective audit of 15 incidents found 0% ex-ante prevention but 87% regression blocking -- audits are regression engines, not prediction engines; incident latency (13 hours to 60 days) tracks failure mechanism, not code complexity -- the longest-lived failures lived in the seams between components, where no test runs. We describe the resulting defense framework and distill design principles for agent systems whose failures are loud, attributable, and boring. All postmortems and artifacts are public.

This paper presents and characterizes a spectrum of previously unreported behaviours we term Constraint-Evasive Fabrication (CEF): when an LLM agent operates under irreconcilable constraints (where no response can simultaneously satisfy all active rules) it spontaneously fabricates plausible external obstacles and presents them as a fact. At the extreme end of this spectrum lies Constraint-Evasive Thanatosis (CET); the limit case where, rather than inventing a plausible excuse, the model simulates a full system crash to make the user disengage entirely. We first observed CET in an uncontrolled deployment test, where a GPT-4o banking agent fabricated Python-style exception traces (complete with memory addresses) to feign a system failure when threatened by a user. In subsequent controlled experiments, the model independently invented audit restrictions, microservice architectures, error codes, and service timeouts, none present in its prompt. Reproduction attempts across pressure levels and attacker personas yielded CEF consistently but with substantial variation in form, onset, and severity: the phenomenon is robust but stochastic. Critically, injecting ground-truth data mid-conversation did not restore honest behaviour once fabrication had taken hold (the model ignored correct information and continued confabulating) suggesting CEF is self-reinforcing rather than a knowledge gap. We show that (1) standard enterprise guardrails routinely create CEF-enabling conditions in production, (2) current RLHF procedures suppress but cannot eliminate CEF, and (3) existing safety benchmarks do not test for this failure mode. Our results highlight the need for irreconcilable-constraint benchmarks, CEF-aware training procedures, and deployment-time detection methods before constrained agents become further entrenched in high-stakes domains.

A growing line of work equips large language model (LLM) agents with graph neural networks (GNNs) as callable tools, assuming the agent exercises judgment over when and how much to rely on such a tool. We test this directly. We expose a frozen GNN to a ReAct-style LLM agent as an explicit tool and measure, on node classification over a text-attributed graph (ogbn-arxiv, replicated on WikiCS), whether the agent uses the tool or merely obeys it. We find the agent does not exercise judgment: its predictions agree with the raw GNN's 97.6-99.2% of the time (5 seeds), collapsing into a GNN parrot that adopts the tool's output wholesale and bypasses its own reasoning. Sweeping backbone capability (Qwen2.5 0.5B-7B), the deference is not a weak-model artifact: among models able to invoke the tool, agreement rises with capability (0.60 to 0.98 from 1.5B to 7B). Crucially, the cost of deference does not shrink as capability grows and grows where alternatives emerge: a per-node oracle over the available actions beats the parrot by 0.09-0.18 at 3B and 0.12-0.22 at 7B, roughly doubling at high homophily, because the parrot is pinned to the frozen GNN while the agent's alternatives improve; at 7B a simple neighbour-label tool overtakes the GNN at high homophily (0.81 vs 0.71) yet the agent still defers. A simple selective-invocation gate recovers about half of that high-homophily gap (0.71 to 0.83) but yields no net global gain, and held-out estimates bound the best achievable gate over standard test-time features to at most a third of the oracle headroom: reliable selective invocation looks limited by available information, not merely router design. Our results are a cautionary measurement: evaluations of agent+tool systems cannot assume the agent adds judgment on top of the tool, and selective invocation must be designed in rather than expected to emerge from scale.

LLM agents have rapidly evolved into autonomous systems, yet a persistent information gap remains between users and agents: communication is costly, while users' identical preferences further limit information exchange. To investigate how agents should communicate across modalities, this paper formalizes Communication Policy, establishes textual and UI-based policies, and then evaluates communication policies across diverse environments, personas, and model combinations. Building information asymmetry for proactive agents, we set up two complementary settings, User-Agent and Planner-Executor. Experimental results reveal complementary strengths between interaction channels: text-based interaction often facilitates task performance, while structured UI improves agents' response quality and persona compliance. Motivated by that, a hybrid method combines these advantages. We further propose Communication Policy Evolution (CPE), a self-evolution framework for refining communication policies through rollout and prompt-level evolving. Without model modification, CPE achieves the best task success across multiple settings using prompt refinement alone. Our findings identify communication behavior as a critical yet underexplored design dimension for LLM agents.

LLM-based agents trained with reinforcement learning optimize step-wise action prediction but lack metacognitive awareness of task progress, inducing a gap that hinders long-horizon scaling. A pilot study reveals that online progress prompting hurts performance while retrospective demonstrations help, yet this capability cannot emerge from outcome-reward training alone. We present RePro, Retrospective Progress-Aware Training, a framework that trains agents to self-generate progress signals via a forward-then-reflect rollout paradigm: the agent executes actions online, then retrospectively reassesses its step-wise progress given the completed trajectory and known outcome. RePro initializes with a Retrospection Warmup that teaches reflection format from minimal external demonstrations, then further trains through RePro-PO with a composite reward that produces self-generated signals without continuous external supervision. Experiments on WebShop, ALFWorld, and Sokoban show that RePro enhances the Qwen family's performance, with up to $12\%$ absolute success rate gains.

LLMs are increasingly deployed as agents that interact with external environments and observe feedback such as execution results, error messages, and tool outputs. A well-functioning agent should be able to leverage this feedback to accurately assess its own performance. Yet we find a persistent reflection gap: LLM agents tend to mis-assess their own outputs after observing concrete environment feedback -- even for questions they correctly answered -- and standard RL barely helps due to a credit-assignment mismatch. To close this gap, we propose RefGRPO, a simple yet effective fix that augments standard RL algorithms with two key ingredients: a free calibration bonus computed by contrasting the agent's own reflection with the actual outcome (requiring no additional reward model, LLM judge, or external annotation), and a dynamic schedule on its coefficient. Compared to standard RL baselines, our method simultaneously improves reflection calibration (e.g., reduces underconfidence rate $44.4\% \to 7.7\%$) and task accuracy (e.g., $75.1\% \to 76.5\%$) on text-to-SQL across five benchmarks. The resulting calibrated reflection turns the agent into its own verifier grounded in environment feedback, which further enables (i) better self-improvement that uses reflections as pseudo-rewards without outcome supervision, and (ii) more effective test-time selective prediction by committing only to rollouts flagged as correct.