Why Multi-Agent LLM Systems Fail 41%-87% of the Time: Coordination Defects, Not Model Limitations

> Paper: Coordination as an Architectural Layer for LLM-Based Multi-Agent Systems: An Information-Controlled Empirical Study on Prediction Markets > Authors: Maksym Nechepurenko, Pavel Shuvalov (Devnull FZCO, Dubai) > arXiv: 2605.03310v1 [cs.MA] > Date: 2026-05-06

Key points

• 79% of multi-agent failures stem from specification/coordination problems; only 21% from base model limitations (MAST study, Cemri et al. 2025, analyzing 1,600+ execution traces across 7 frameworks including AutoGen, MetaGPT, ChatDev).
• Production failure rates for multi-agent systems range from 41% to 87%.
• Simpler, cheaper single-agent baselines often match or outperform multi-agent workflows when tools and retrieval context are held constant (Xu et al. 2026; Xia et al. 2024).
• The paper's core thesis: coordination should be an explicitly configurable architectural layer, separated from agent logic and information access.

The problem: a crisis of faith in multi-agent systems

Prior literature documents coordination pathologies but not their causes:

• Wynn et al., 2025: alignment pressure in debate loops suppresses correct minority opinions (social dynamics defects).
• Acharya, 2026: "semantic intent divergence" — cooperating agents without a shared process model develop inconsistent goals.
• Agashe et al., 2025: LLMs struggle as Theory of Mind demands increase.
• Xu et al., 2026; Xia et al., 2024: single-agent baselines match/exceed multi-agent at lower cost.

Existing work has two blind spots: empirical studies only catalog failure modes (what goes wrong, not why), and declarative orchestration frameworks (AutoGen, CrewAI) offer engineering convenience but no mapping from coordination configuration to predictable failures.

Proposed framework: coordination as an architectural layer

The system decomposes into three layers:

1. Information layer — tools, retrieval context, external sensors, data sources 2. Coordination layer C (the paper's focus) — agent endpoints, communication topology, authority assignment, synchronization, aggregation rules, termination conditions, failure handling 3. Agent layer — individual LLM calls, role prompts, per-agent tool wrappers

This separation enables decision provenance, configuration-specific failure signatures, cross-system comparability, and explicit heterogeneity specification.

Experimental design: information-controlled methodology

Responding to Ao et al.'s (2026) methodological critique, all information channels were fixed:

• Single model (claude-opus-4-6), temperature 0.3, 1,500 token generation cap per call
• Fixed tool set (Polymarket metadata + price history; web search disabled with token-equivalent empty responses)
• Task set: 100 binary Polymarket markets (resolution ≥ 2025-09-15, post training-cutoff), balanced across categories (crypto, politics, sports, economics, geopolitics, entertainment) and outcomes (53% YES / 47% NO)
• Market consensus baseline Brier: 0.152

Five architectures, five results

| Configuration | Brier ↓ | Cost/market | Predicted failure mode | |---|---|---|---| | Sequential pipeline (SP) | 0.153 | $0.36 | Early-error cascade from stage 1 | | Independent ensemble (IE) | 0.159 | $0.10 | Confidently wrong consensus | | Orchestrator-specialist (OS) | 0.162 | $0.31 | Single-point error cascade | | Peer-critique debate (PC) | 0.170 | $0.23 | Premature convergence | | Consensus alignment (CA) | 0.181 | $0.10 | Diversity collapse |

Findings:

• No configuration beat market consensus — all showed negative Alpha.
• The Pareto frontier has only two points: IE (cost-optimal) and SP (quality-optimal). OS and PC are strictly dominated by IE; CA is dominated at equal cost.
• Murphy decomposition confirmed 3/5 pre-registered failure-mode predictions. Notably, OS's prediction was contradicted: the orchestrator *increased* calibration error (REL = 0.025, highest), suggesting orchestrators lack reliable meta-cognition about their own judgments.
• CA was the worst performer, with the lowest resolution (RES = 0.094) — forced agreement collapsed diversity, often anchoring on market consensus itself. Ironically, CA-style coordination (AutoGen's GroupChat, CrewAI's collaboration modes) is the most widely adopted.
• Category analysis: architecture choice mattered most in economics (0.121 spread, SP 0.120 vs PC 0.241) and almost not at all in sports (0.012) — hypothesizing coordination matters where structured numerical reasoning is required (unregistered hypothesis).
• Statistical honesty: with n=100, no pairwise comparison survived Bonferroni correction; detecting the observed 0.006–0.011 gaps would require 1,500–2,000 questions.

Practical guidance

| Scenario | Recommendation | |---|---| | Cost-sensitive, quick predictions | Independent ensemble (IE) — $0.10/market | | Quality-first, larger budget | Sequential pipeline (SP) — best Brier, 3.6× cost | | Avoid | Orchestrator-specialist, peer-critique debate (dominated by IE) | | Especially cautious | Consensus alignment (diversity collapse) |

Conclusion

The paper's three core takeaways:

1. 79% of failures are coordination problems, not model problems — check your topology and aggregation rules before blaming the LLM. 2. Coordination overhead does not necessarily convert into quality gains — simple agents often win. 3. Coordination configurations produce predictable failure signatures — confident wrong consensus, diversity collapse — these are structural, not accidental.

An independent replication channel (Foresight Arena, live on-chain agents with web search enabled) is accumulating data in parallel. Code: https://github.com/ForesightFlow/foreflow-agents

References

1. Nechepurenko, M. & Shuvalov, P. (2026). *Coordination as an Architectural Layer for LLM-Based Multi-Agent Systems*. arXiv:2605.03310v1. 2. Cemri et al. (2025). *MAST: Multi-Agent System Trajectory Analysis*. 3. Wynn et al. (2025); Acharya (2026); Agashe et al. (2025); Xu et al. (2026); Xia et al. (2024); Ao et al. (2026).