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 system failures stem from specification and coordination problems, not base model capability limits (21%), based on the MAST study (Cemri et al., 2025) covering 1,600+ execution traces across 7 frameworks (AutoGen, MetaGPT, ChatDev, etc.)
• Production failure rates for multi-agent systems range 41%-87%
• Single-agent baselines often match or beat multi-agent workflows at lower cost (Xu et al., 2026; Xia et al., 2024)
• The paper proposes treating coordination as a configurable architectural layer, separate from agent logic and information access
• Experiments: 100 binary Polymarket markets, 5 coordination architectures, fully controlled information

The Three-Layer Architecture

The paper decomposes multi-agent systems into:

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

The coordination layer is defined by seven elements: agent endpoints, communication topology (directed graph), authority assignment, synchronization mechanism, aggregation rules, termination conditions, and failure handling. This separation enables decision attribution, failure-mode signatures per configuration, and cross-system comparability.

Experimental Design: Information Control

Fixed: single LLM (claude-opus-4-6, temperature 0.3), fixed tool set, fixed prompt templates (role instructions only vary), 1,500-token output cap, 100 binary Polymarket questions (crypto 17, politics 17, sports 17, economics 16, geopolitics 17, entertainment 16; 53% YES / 47% NO outcomes).

Released: total computation per question, reported as an endogenous output and included in cost-quality analysis.

Tools: getMarketDetails and getPriceHistory enabled (Polymarket Gamma/CLOB); searchWeb disabled (empty results, preserving prompt-token equivalence). Market consensus baseline: Brier 0.152.

Five Architectures, Predicted Failure Modes

| Configuration | Mechanism | Predicted failure mode | |---|---|---| | Independent Ensemble (IE) | 3 independent predictions, arithmetic mean | Confident wrong consensus from correlated errors | | Peer-Critique Debate (PC) | 2 debate rounds, mean | Premature convergence via alignment pressure | | Orchestrator-Specialist (OS) | Orchestrator delegates to experts | Single-point error cascades | | Sequential Pipeline (SP) | Researcher → Critic → Predictor | Early errors unrecoverable downstream | | Consensus Alignment (CA) | 3 iterative rounds, tolerance 0.05 | Diversity collapse anchored to market consensus |

Results

| Configuration | Brier ↓ | Tokens/Market | Cost/Market | |---|---|---|---| | sequential_pipeline (SP) | 0.153 | 55,253 | $0.36 | | independent_ensemble (IE) | 0.159 | 15,362 | $0.10 | | orchestrator_specialist (OS) | 0.162 | 49,215 | $0.31 | | peer_critique_debate (PC) | 0.170 | 36,015 | $0.23 | | consensus_alignment (CA) | 0.181 | 15,113 | $0.10 | | market consensus | 0.152 | — | — |

Key findings:

• No configuration beat market consensus (all negative alpha)
• The Pareto frontier has only two points: IE (cost-efficient, $0.10) and SP (quality-optimal, $0.36, ~3.6× the tokens)
• OS and PC are strictly dominated by IE (higher cost, worse quality); CA is dominated at equal cost
• Murphy decomposition confirmed 3/5 pre-registered failure-mode predictions; notably, OS contradicted predictions—the orchestrator *increased* calibration error (REL=0.025, the highest) rather than reducing it
• CA showed the lowest resolution (RES=0.094), confirming diversity collapse—the worst performer overall
• None of the pairwise differences reached p << 0.05; the authors honestly report n=100 makes architectural separation *visible* but not Bonferroni-significant (~1,500–2,000 questions needed to resolve 0.006–0.011 gaps)
• Architecture mattered most in economics (gap 0.121, structured numerical reasoning) and almost not at all in sports (gap 0.012)

Recommendations

| Scenario | Recommended | Rationale | |---|---|---| | Cost-sensitive prediction | Independent Ensemble | Pareto low-cost point, $0.10/market | | Quality-first, budget available | Sequential Pipeline | Lowest REL, but 3.6× cost | | Avoid | Orchestrator-Specialist, Peer Debate | Dominated by IE | | Be especially wary | Consensus Alignment | Lowest RES, diversity collapse |

Suggested improvements: redesign orchestrator integration to reduce its uncertainty about expert outputs; add dissent-preservation mechanisms in debate; disable or modify consensus convergence conditions; add stage-1 validation layers in pipelines.

Independent Verification

All five configurations are deployed as live agents on Foresight Arena with web search enabled on real future events, using on-chain commit-reveal. Data is accumulating in parallel; code: https://github.com/ForesightFlow/foreflow-agents

Conclusion

The paper's central claim: multi-agent systems fail not because weak LLMs are placed inside them, but because coordination is designed as an "implementation detail" rather than an "architectural layer." Three takeaways:

1. 79% of failures are coordination problems — check your topology and aggregation rules before blaming the LLM 2. Simpler single agents often beat complex multi-agent workflows — coordination overhead has real costs without guaranteed quality gains 3. Coordination configurations predict failure modes — consensus alignment collapses diversity by design, not by accident

The most ironic finding: consensus alignment—the pattern behind AutoGen's GroupChat and CrewAI's collaboration modes—performed worst. What looks like "smarter collaboration" is systematically "more complete collapse."

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). Alignment pressure in debate loops. 4. Acharya (2026). Semantic intent divergence. 5. Agashe et al. (2025). LLM theory-of-mind limits. 6. Xu et al. (2026); Xia et al. (2024). Single-agent baselines vs. multi-agent workflows. 7. Ao et al. (2026). Methodological critique of multi-agent evaluation.