Problem it solves
Multi-agent runs produce conflicting results that are silently reconciled by the system or presented as if they were consistent. Users don't know a conflict existed.
When to use
Whenever two or more agents return contradictory outputs on the same question, recommendation, or data point.
When not to use
For minor variation within an acceptable range. This pattern is for structural contradictions that require a human decision, not data normalization.
Governing principle
Auto-resolution is never acceptable. The run cannot close without a human decision on conflicting outputs.
Required Components
Interaction Flow
Conflict detected
Two or more agents return outputs that are logically contradictory on the same subject.
Run paused at conflict
The orchestrator pauses before any downstream processing of the conflicting outputs.
Conflict surface activated
The Conflict Resolution Surface shows both outputs side-by-side, names the specific contradiction, and explains the downstream consequences of each resolution path.
Human reviews
The user reads both outputs and the conflict explanation.
Resolution chosen
The user selects one output, requests a tie-breaker from a third agent, or flags the conflict for expert review.
Resolution logged
The conflict, both outputs, the chosen resolution, and the user's identity are written to the governance record.
Governance requirements
Conflicts are governance events. The conflict record must include both outputs, the nature of the contradiction, and the resolution decision. Conflicts cannot be silently resolved by the system.
Accessibility notes
Conflict surfaces must be announced via role="alert". Both conflicting outputs must be equally accessible — neither should be visually de-emphasized before the user has read them.