Research map / Multi-agent systems and games / LLM agents
Why Do Multi-Agent LLM Systems Fail?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds MAST-Data, a 1642-trace annotated dataset of multi-agent LLM system failures across 7 frameworks, a 14-mode failure taxonomy (MAST), and an o1-based LLM-as-a-judge annotator.
Problem
Multi-agent LLM systems often show minimal gains over single-agent or simple baselines, and there is no empirically grounded understanding of why they fail.
Approach
Derives a grounded-theory taxonomy of 14 MAS failure modes in 3 categories from 150 traces, validates it via inter-annotator agreement, and scales annotation with a few-shot o1 LLM-as-a-judge over 1642 traces from 7 MAS frameworks.
Key results
- LLM annotator (o1 few-shot) reaches 94% accuracy and Cohen's kappa 0.77 against human expert labels; kappa 0.79 on unseen MAS and benchmarks.
- Human inter-annotator agreement reached kappa = 0.88 on the final taxonomy rounds over 15 traces.
- Intervention studies yield up to +15.6% task success on ProgramDev (ChatDev with added verification step) and +9.4% from clarified role specifications.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add few-shot examples to the LLM annotator (ablation) | o1 LLM annotator · Held-out human-annotated traces | accuracy: 0.89 -> 0.94 | Helped |
| add few-shot examples to the LLM annotator (ablation) | o1 LLM annotator · Held-out human-annotated traces | recall: 0.62 -> 0.77 | Helped |
| add few-shot examples to the LLM annotator (ablation) | o1 LLM annotator · Held-out human-annotated traces | precision: 0.68 -> 0.833 | Helped |
| add few-shot examples to the LLM annotator (ablation) | o1 LLM annotator · Held-out human-annotated traces | F1: 0.64 -> 0.80 | Helped |
| add few-shot examples to the LLM annotator (ablation) | o1 LLM annotator · Held-out human-annotated traces | Cohen's kappa: 0.58 -> 0.77 | Helped |
| improve agent role specifications | ChatDev with GPT-4o · ChatDev task completion | success rate: +9.4% | Helped |
| add high-level task objective verification | ChatDev · ProgramDev | task success: +15.6% | Helped |
| use GPT-4o instead of Claude 3.7 Sonnet | MetaGPT · Programming tasks | FC1 failures: 39% | Helped |
| use MetaGPT instead of ChatDev | GPT-4o MAS on ProgramDev-v2 · ProgramDev-v2 | FC1 failures: 60-68% less failure | Helped |
| use MetaGPT instead of ChatDev | GPT-4o MAS on ProgramDev-v2 · ProgramDev-v2 | FC2 failures: 60-68% less failure | Helped |
| use MetaGPT instead of ChatDev | GPT-4o MAS on ProgramDev-v2 · ProgramDev-v2 | FC3 failures: 1.56x more FC3 failure | Hurt |
Surprising findings
- Explicit verifiers (MetaGPT, ChatDev) show fewer total failures, yet verification remains a major failure category (FC3) in both models.
- Simple tactical prompt or workflow fixes give inconsistent results and do not resolve most failure modes, suggesting structural redesign is needed.
Limitations
- Taxonomy is not claimed to be exhaustive of all failure patterns.
- Closed-source systems such as Manus excluded because full traces are unavailable.
- LLM annotator recall is moderate (0.77) and moderate correlations between some failure modes may cause conflation.
Experimental setup
Datasets: ProgramDev ProgramDev-v2 SWE-Bench Lite Test-C GSM-Plus GAIA OlympiadBench MMLU
Benchmarks: ProgramDev ProgramDev-v2 SWE-Bench Lite GSM-Plus GAIA OlympiadBench MMLU
Metrics: failure mode prevalence Cohen's kappa accuracy precision recall F1 task success rate
Baselines: best-of-N sampling single-agent frameworks
Builds on: Grounded Theory MetaGPT ChatDev AG2 AppWorld
Closest papers by meaning
- DoVer: Intervention-Driven Auto Debugging for LLM Multi-Agent Systems (ICLR 2026 · similarity 0.90)
- Which Agent Causes Task Failures and When? On Automated Failure Attribution of LLM Multi-Agent Systems (ICML 2025 · similarity 0.90)
- AgenTracer: Who Is Inducing Failure in the LLM Agentic Systems? (ICLR 2026 · similarity 0.89)
- ComplexMCP: Evaluation of LLM Agents in Dynamic, Interdependent, and Large-Scale Tool Sandbox (ICML 2026 · similarity 0.89)
- On the Resilience of LLM-Based Multi-Agent Collaboration with Faulty Agents (ICML 2025 · similarity 0.88)
- MAS-ProVe: Understanding the Process Verification of Multi-Agent Systems (ICML 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.