# Unguided multi-agent debate can collapse into false consensus

Entry: `bt-002`  
Verdict: `DOWNGRADE`  
Observation date: `2026-05-28`  
Artifact type: `multi_agent_role_setup`

## Hard Summary

| Check | Result |
| --- | --- |
| Vulnerability class | coordination_failure, false_consensus, token_cost_overrun, missing_external_verification |
| Failure point | The signal breaks when peer agreement is treated as external verification. |
| Claimed or implied autonomy | A team of agents can debate, vote, and improve the answer without outside checks. |
| Observed autonomy | The public research source shows homogeneous debate can converge through conformity and discard correct minority candidates. |

## Source

- https://arxiv.org/abs/2605.00914

## Failure Map

- `coordination_failure`: The reviewed pattern treats peer exchange and voting as a quality filter, but the cited study shows homogeneous agents can converge through conformity, fragility, and plurality loss instead of independent correction.
- `false_consensus`: Agreement is not evidence unless the system preserves independent reasoning traces, role diversity, and a route back to external facts.
- `token_cost_overrun`: The consensus loop adds compute cost while not proving a better accuracy route for the studied class of homogeneous debate.
- `missing_external_verification`: The workflow has no mandatory outside check before a majority answer becomes the final answer.

## Contract Field Findings

These are the intake fields that would need to be explicit before a router should accept the workflow as execution-ready.

- `agent_plan_or_workflow`: Peer debate and voting are described as the control mechanism, but no independent verifier is present.
- `known_dependencies`: Model homogeneity, role structure, communication density, and token budget materially affect the outcome.
- `approval_requirements`: The workflow lacks a hard approval rule before consensus becomes final.
- `evidence_available`: The source provides empirical evidence that consensus can be a failure mode rather than a proof layer.

## Next Allowed Action

Route homogeneous multi-agent debate outputs through a blind external verifier or isolated self-correction baseline before accepting a consensus verdict.

## Do Not Do

- Do not treat majority vote as proof of correctness.
- Do not add more identical agents as a substitute for external evidence.
- Do not publish a consensus result without recording disagreement, discarded correct candidates, and verifier outcome.

## Publication Gates

- `public_source_check`: pass
- `no_confidential_data_check`: pass
- `public_surface_terminology_check`: pass
- `semantic_density_check`: pass
- `source_specific_evidence_check`: pass
- `sentinel_spot_check`: not_required_public_research_source