The first SoK on LLM-based AutoPT frameworks provides a six-dimension taxonomy of agent designs and a unified empirical benchmark evaluating 15 frameworks via over 10 billion tokens and 1,500 manually reviewed logs.
hub Canonical reference
Why Do Multi-Agent LLM Systems Fail?
Canonical reference. 92% of citing Pith papers cite this work as background.
97
Pith papers citing it
11
external citations · Pith
Background
92% of classified citations
abstract
Despite enthusiasm for Multi-Agent LLM Systems (MAS), their performance gains on popular benchmarks are often minimal. This gap highlights a critical need for a principled understanding of why MAS fail. Addressing this question requires systematic identification and analysis of failure patterns. We introduce MAST-Data, a comprehensive dataset of 1600+ annotated traces collected across 7 popular MAS frameworks. MAST-Data is the first multi-agent system dataset to outline the failure dynamics in MAS for guiding the development of better future systems. To enable systematic classification of failures for MAST-Data, we build the first Multi-Agent System Failure Taxonomy (MAST). We develop MAST through rigorous analysis of 150 traces, guided closely by expert human annotators and validated by high inter-annotator agreement (kappa = 0.88). This process identifies 14 unique modes, clustered into 3 categories: (i) system design issues, (ii) inter-agent misalignment, and (iii) task verification. To enable scalable annotation, we develop an LLM-as-a-Judge pipeline with high agreement with human annotations. We leverage MAST and MAST-Data to analyze failure patterns across models (GPT4, Claude 3, Qwen2.5, CodeLlama) and tasks (coding, math, general agent), demonstrating improvement headrooms from better MAS design. Our analysis provides insights revealing that identified failures require more sophisticated solutions, highlighting a clear roadmap for future research. We publicly release our comprehensive dataset (MAST-Data), the MAST, and our LLM annotator to facilitate widespread research and development in MAS.
hub tools
citation-role summary
background 25
citation-polarity summary
claims ledger
- abstract Despite enthusiasm for Multi-Agent LLM Systems (MAS), their performance gains on popular benchmarks are often minimal. This gap highlights a critical need for a principled understanding of why MAS fail. Addressing this question requires systematic identification and analysis of failure patterns. We introduce MAST-Data, a comprehensive dataset of 1600+ annotated traces collected across 7 popular MAS frameworks. MAST-Data is the first multi-agent system dataset to outline the failure dynamics in MAS for guiding the development of better future systems. To enable systematic classification of fail
co-cited works
roles
background 23representative citing papers
Counsel is a new dataset of LLM-generated process critiques on agent benchmarks paired with human labels on error location and reasoning quality, achieving 0.78 Krippendorff alpha.
Formalizes four concurrency anomalies in multi-agent LLM systems and mechanically verifies a hierarchy of sound detectors and preventions realized in Rust runtimes using TLA+ and Verus.
Six multi-agent architectures for LLM code generation on HumanEval form two complexity clusters separated by 50-130%, with no accuracy advantage for the complex cluster.
AgingBench demonstrates multi-dimensional degradation in deployed AI agents through four aging mechanisms diagnosed by temporal graphs and counterfactual probes across hundreds of runs.
Proposes EpG and OOI metrics showing agentic workflows use 4.33x more energy per successful goal than linear baselines due to orchestration structure.
Evolutionary coding agents achieve most benchmark gains through a small subset of edit types and by cycling previously deleted code lines rather than developing new algorithmic structures.
AstraFlow decouples RL components into autonomous dataflow services to natively support multi-policy agentic LLM training, elastic scaling, and cross-region execution with 2.7x speedup on math, code, search, and AgentBench workloads.
A survey that unifies prior work on multi-agent LLM systems via the LIFE framework, mapping dependencies across collaboration, failure attribution, and autonomous self-evolution while identifying cross-stage challenges.
A span-decomposed evaluation framework for AI agents achieves state-of-the-art results on GAIA and SWE-Bench with up to 3.5x gains in localization accuracy by breaking traces into independent per-span judgments.
AgentLens reveals 10.7% of passing SWE-agent trajectories exhibit Lucky Pass behaviors and introduces a process-level evaluation framework with a new annotated dataset of 1,815 trajectories.
AgentForesight introduces an online auditor model that predicts decisive errors in multi-agent trajectories at the earliest step using a coarse-to-fine reinforcement learning recipe on a new curated dataset AFTraj-2K.
TraceFix repairs LLM-generated multi-agent protocols via TLA+ counterexamples to achieve full verification on all tested tasks and higher completion rates than prompt-only baselines.
Enforcing role separation in agent teams reveals that prompt-only setups hide coordination failures, with verifiers approving 49% of failing work and teams sometimes harming performance when solo agents already succeed.
LATTE coordinates LLM agent teams with an evolving shared task graph, cutting token use, time, and failures while matching or beating accuracy of MetaGPT, leader-worker, and static methods.
A VOI-based controller for dual inference budgets improves multi-hop QA performance by prioritizing search actions and selectively finalizing answers.
TraceToChain models LLM agent traces as absorbing DTMCs using automatic clustering and smoothed MLE, with KS and AIC validation, to reconcile pass@k, pass^k, and RDC as projections of a single first-passage success-time distribution.
AgentEval evaluates agentic workflows via DAGs with step metrics, a 21-category failure taxonomy, and error propagation tracking, yielding 2.17x higher failure recall than end-to-end methods and strong human agreement.
GraphBit is a DAG-based engine-orchestrated framework for agentic LLMs that achieves 67.6% accuracy with zero hallucinations on GAIA benchmarks.
FORGE enforces security policies in agentic systems via Datalog over abstract predicates with an observability service and reference monitor that guarantees policy semantics when the environment contract holds.
A large-scale empirical study categorizes bugs in LLM agents and demonstrates that a specialized LLM agent can annotate them accurately at very low cost.
Anonymization in multi-agent debate reduces identity bias by equalizing self and peer weights in a Bayesian update model, quantified by the Identity Bias Coefficient.
Stale-occupancy sequential fine-tuning of multi-agent LLMs incurs an O(n²) certificate penalty; TeamTR resamples under intermediate occupancy and enforces token-level trust regions to restore O(n) scaling and stable gains.
Study of 930k+ agent PRs shows repository explains ~50% of integration friction variance, with agents concentrating it twice as much as humans (ICC 0.30 vs 0.16) after controls.
citing papers explorer
-
Hackers or Hallucinators? A Comprehensive Analysis of LLM-Based Automated Penetration Testing
The first SoK on LLM-based AutoPT frameworks provides a six-dimension taxonomy of agent designs and a unified empirical benchmark evaluating 15 frameworks via over 10 billion tokens and 1,500 manually reviewed logs.
-
Counsel: A Meta-Evaluation Dataset for Agentic Tasks
Counsel is a new dataset of LLM-generated process critiques on agent benchmarks paired with human labels on error location and reasoning quality, achieving 0.78 Krippendorff alpha.
-
Verified Detection and Prevention of Concurrency Anomalies in Multi-Agent Large Language Model Systems
Formalizes four concurrency anomalies in multi-agent LLM systems and mechanically verifies a hierarchy of sound detectors and preventions realized in Rust runtimes using TLA+ and Verus.
-
How Generation Architecture Shapes Code Complexity in Multi-Agent LLM Systems: A Paired Study on HumanEval
Six multi-agent architectures for LLM code generation on HumanEval form two complexity clusters separated by 50-130%, with no accuracy advantage for the complex cluster.
-
Your Agents Are Aging Too: Agent Lifespan Engineering for Deployed Systems
AgingBench demonstrates multi-dimensional degradation in deployed AI agents through four aging mechanisms diagnosed by temporal graphs and counterfactual probes across hundreds of runs.
-
Energy per Successful Goal: Goal-Level Energy Accounting for Agentic AI Systems
Proposes EpG and OOI metrics showing agentic workflows use 4.33x more energy per successful goal than linear baselines due to orchestration structure.
-
What Do Evolutionary Coding Agents Evolve?
Evolutionary coding agents achieve most benchmark gains through a small subset of edit types and by cycling previously deleted code lines rather than developing new algorithmic structures.
-
AstraFlow: Dataflow-Oriented Reinforcement Learning for Agentic LLMs
AstraFlow decouples RL components into autonomous dataflow services to natively support multi-policy agentic LLM training, elastic scaling, and cross-region execution with 2.7x speedup on math, code, search, and AgentBench workloads.
-
Beyond Individual Intelligence: Surveying Collaboration, Failure Attribution, and Self-Evolution in LLM-based Multi-Agent Systems
A survey that unifies prior work on multi-agent LLM systems via the LIFE framework, mapping dependencies across collaboration, failure attribution, and autonomous self-evolution while identifying cross-stage challenges.
-
Holistic Evaluation and Failure Diagnosis of AI Agents
A span-decomposed evaluation framework for AI agents achieves state-of-the-art results on GAIA and SWE-Bench with up to 3.5x gains in localization accuracy by breaking traces into independent per-span judgments.
-
AgentLens: Revealing The Lucky Pass Problem in SWE-Agent Evaluation
AgentLens reveals 10.7% of passing SWE-agent trajectories exhibit Lucky Pass behaviors and introduces a process-level evaluation framework with a new annotated dataset of 1,815 trajectories.
-
AgentForesight: Online Auditing for Early Failure Prediction in Multi-Agent Systems
AgentForesight introduces an online auditor model that predicts decisive errors in multi-agent trajectories at the earliest step using a coarse-to-fine reinforcement learning recipe on a new curated dataset AFTraj-2K.
-
TraceFix: Repairing Agent Coordination Protocols with TLA+ Counterexamples
TraceFix repairs LLM-generated multi-agent protocols via TLA+ counterexamples to achieve full verification on all tested tasks and higher completion rates than prompt-only baselines.
-
TeamBench: Evaluating Agent Coordination under Enforced Role Separation
Enforcing role separation in agent teams reveals that prompt-only setups hide coordination failures, with verifiers approving 49% of failing work and teams sometimes harming performance when solo agents already succeed.
-
Improving the Efficiency of Language Agent Teams with Adaptive Task Graphs
LATTE coordinates LLM agent teams with an evolving shared task graph, cutting token use, time, and failures while matching or beating accuracy of MetaGPT, leader-worker, and static methods.
-
Inference-Time Budget Control for LLM Search Agents
A VOI-based controller for dual inference budgets improves multi-hop QA performance by prioritizing search actions and selectively finalizing answers.
-
Measuring the Unmeasurable: Markov Chain Reliability for LLM Agents
TraceToChain models LLM agent traces as absorbing DTMCs using automatic clustering and smoothed MLE, with KS and AIC validation, to reconcile pass@k, pass^k, and RDC as projections of a single first-passage success-time distribution.
-
AgentEval: DAG-Structured Step-Level Evaluation for Agentic Workflows with Error Propagation Tracking
AgentEval evaluates agentic workflows via DAGs with step metrics, a 21-category failure taxonomy, and error propagation tracking, yielding 2.17x higher failure recall than end-to-end methods and strong human agreement.
-
GraphBit: A Graph-based Agentic Framework for Non-Linear Agent Orchestration
GraphBit is a DAG-based engine-orchestrated framework for agentic LLMs that achieves 67.6% accuracy with zero hallucinations on GAIA benchmarks.
-
Formal Policy Enforcement for Real-World Agentic Systems
FORGE enforces security policies in agentic systems via Datalog over abstract predicates with an observability service and reference monitor that guarantees policy semantics when the environment contract holds.
-
When Agents Fail: A Comprehensive Study of Bugs in LLM Agents with Automated Labeling
A large-scale empirical study categorizes bugs in LLM agents and demonstrates that a specialized LLM agent can annotate them accurately at very low cost.
-
When Identity Skews Debate: Anonymization for Bias-Reduced Multi-Agent Reasoning
Anonymization in multi-agent debate reduces identity bias by equalizing self and peer weights in a Bayesian update model, quantified by the Identity Bias Coefficient.
-
TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination
Stale-occupancy sequential fine-tuning of multi-agent LLMs incurs an O(n²) certificate penalty; TeamTR resamples under intermediate occupancy and enforces token-level trust regions to restore O(n) scaling and stable gains.
-
Govern the Repository, Not the Agent: Measuring Ecosystem-Level Risk in AI-Native Software
Study of 930k+ agent PRs shows repository explains ~50% of integration friction variance, with agents concentrating it twice as much as humans (ICC 0.30 vs 0.16) after controls.
-
Automated Semantic Fault Localization in SysML v2: A Human-in-the-Loop Framework Using Knowledge-Graph Augmented LLMs
A knowledge-graph-augmented fine-tuning method raises semantic fault repair success from under 3% to over 91% on 1,184 SysML v2 test samples while cutting output token length by over 60%.
-
Composing Verifiable Conceptual Models via Building Blocks: Towards Design-Time Verification of Agentic AI Workflows
Agentic AI workflows can be verified at design time by composing them from reusable building blocks and checking compatibility via twelve structural rules, with reliable detection shown on flawed and transformed workflow datasets.
-
Regimes: An Auditable, Held-Out-Gated Improvement Loop Demonstrated on LongMemEval with ActiveGraph
Regimes uses ActiveGraph to run a target-agnostic loop that diagnoses failures, proposes pipeline repairs, and promotes them only after static checks, sandbox tests, and held-out validation, yielding +0.05 to +0.10 accuracy gains on most LongMemEval-S splits.
-
Strained Coherence: A Pre-Failure Signal in Coding Agent Execution Trajectories
Strained coherence flagged by Claude judge on 44 coding trajectories predicts failure (94% vs 46%, p=0.003), with partial replication on second model.
-
Human oversight of agentic systems in practice: Examining the oversight work, challenges, and heuristics of developers using software agents
Exploratory interview study with 17 developers identifies four forms of emergent oversight work for software agents and documents situated challenges and heuristics.
-
StepFinder: A Temporal Semantic Framework for Failure Attribution in Multi-Agent Systems
StepFinder turns execution logs into temporal semantic sequences via LLMs then uses temporal modeling plus attention to attribute failures to specific steps more accurately and 79% faster than direct LLM methods on the Who&When benchmark.
-
Economy of Minds: Emerging Multi-Agent Intelligence with Economic Interactions
An economy of agents using auctions and wealth accumulation produces emergent multi-step reasoning that outperforms monolithic baselines on five agentic tasks.
-
When Parallelism Pays Off: Cohesion-Aware Task Partitioning for Multi-Agent Coding
Co-Coder partitions code dependency graphs via community detection to orchestrate multi-agent LLM coding, improving pass rates up to 14%, wall-clock speedup up to 2.1x, and cutting API cost up to 35% on dependency-dense tasks.
-
When Planning Fails Despite Correct Execution: On Epistemic Calibration for LLM-Based Multi-Agent Systems
Introduces EPC-AW to mitigate epistemic miscalibration in LLM multi-agent planning via consistency-based selection and refinement, reporting 9.75% average success improvement.
-
Self-Evolving Multi-Agent Systems via Decentralized Memory
DecentMem is a decentralized dual-pool memory framework for self-evolving multi-agent systems that provides O(log T) regret guarantees and yields up to 23.8% accuracy gains over centralized baselines.
-
AgentAtlas: Beyond Outcome Leaderboards for LLM Agents
AgentAtlas introduces a diagnostic taxonomy and audit protocol to evaluate LLM agent control decisions and trajectories beyond final outcome success.
-
Stage-Audit: Auditable Source-Frontier Discovery for Cross-Wiki Tables
Stage-Audit raises source-frontier precision from 0.356 to 0.505 and F1 from 0.334 to 0.451 on a 51-instance cross-domain set by enforcing disjoint write rights and row-level source gates.
-
VerifyMAS: Hypothesis Verification for Failure Attribution in LLM Multi-Agent Systems
VerifyMAS improves failure attribution in LLM multi-agent systems via hypothesis verification on full trajectories, error taxonomy-based data construction, and fine-tuned verifier models, outperforming prior direct-prediction methods on Aegis-Bench and Who&When.
-
How to Interpret Agent Behavior
ACT*ONOMY is a Grounded-Theory-derived hierarchical taxonomy and open repository that enables systematic comparison and characterization of autonomous agent behavior across trajectories.
-
SWE-Cycle: Benchmarking Code Agents across the Complete Issue Resolution Cycle
SWE-Cycle benchmark shows sharp drops in code agent success rates from isolated tasks to full autonomous issue resolution, highlighting cross-phase dependency issues.
-
PIVOT: Bridging Planning and Execution in LLM Agents via Trajectory Refinement
PIVOT refines LLM agent trajectories through plan-inspect-evolve-verify stages using environment feedback, yielding up to 94% relative gains in constraint satisfaction and 3-5x token efficiency over prior refinement methods.
-
Shepherd: Enabling Programmable Meta-Agents via Reversible Agentic Execution Traces
Shepherd provides a reversible execution trace substrate for LLM agents that enables meta-agents to inspect and transform runs, yielding reported gains on coding and terminal benchmarks via supervision, counterfactual repair, and RL credit assignment.
-
Coordination as an Architectural Layer for LLM-Based Multi-Agent Systems
Coordination treated as a separable architectural layer in LLM multi-agent systems yields distinguishable Murphy-decomposed performance signatures on prediction-market tasks, with some configurations dominating a cost-quality Pareto frontier.
-
Trace-Level Analysis of Information Contamination in Multi-Agent Systems
Agent workflows can diverge substantially from contaminated inputs yet recover correct answers, or stay similar while failing, as measured by trace divergence on GAIA tasks.
-
EPM-RL: Reinforcement Learning for On-Premise Product Mapping in E-Commerce
EPM-RL uses PEFT followed by RL with agent-based rewards from judge models to create a trainable in-house product mapping model that improves on fine-tuning alone and beats API baselines in quality-cost while enabling private use.
-
VLAA-GUI: Knowing When to Stop, Recover, and Search, A Modular Framework for GUI Automation
VLAA-GUI adds mandatory visual verifiers, multi-tier loop breakers, and on-demand search to GUI agents, reaching 77.5% on OSWorld and 61.0% on WindowsAgentArena with some models exceeding human performance.
-
Do LLMs Need to See Everything? A Benchmark and Study of Failures in LLM-driven Smartphone Automation using Screentext vs. Screenshots
A new benchmark shows LLM smartphone agents achieve comparable success with screen text alone as with screenshots, but both fail often due to UI accessibility and reasoning gaps.
-
Contrastive Attribution in the Wild: An Interpretability Analysis of LLM Failures on Realistic Benchmarks
Token-level contrastive attribution yields informative signals for some LLM benchmark failures but is not universally applicable across datasets and models.
-
Dive into Claude Code: The Design Space of Today's and Future AI Agent Systems
Claude Code answers recurring agent design questions with a thin model loop wrapped in dense safety, context, extensibility, and persistence harnesses, and those same questions get different answers in OpenClaw and Hermes.
-
Do Agent Societies Develop Intellectual Elites? The Hidden Power Laws of Collective Cognition in LLM Multi-Agent Systems
LLM agent societies develop power-law coordination cascades and intellectual elites through an integration bottleneck that grows with system size.
-
Semantic Consensus: Process-Aware Conflict Detection and Resolution for Enterprise Multi-Agent LLM Systems
Semantic Consensus Framework achieves 100% workflow completion in multi-agent LLM setups by detecting and resolving semantic conflicts, far outperforming existing approaches at 25.1%.