Conflicting agent answers need evidence-based resolution, not a vote on which response sounds more confident. First establish whether the agents are answering the same question about the same resource and policy version.
Conflict detection starts with comparable claims
A multi-agent conflict occurs when relevant claims cannot all be true under the same conditions. Differences in scope, observation time, or interpretation can look contradictory while describing separate facts.
For a proposed detector, represent consequential claims with subject, predicate, value, source, version, observation time, and uncertainty. Compare only claims whose context matches. Route unresolved material conflicts to the appropriate authority or human reviewer.
Do not assume that agreement among several agents establishes truth. They may share the same stale document or repeat one another’s unsupported inference. Independent evidence matters more than the number of generated confirmations.
Examine a fictional opening-hours disagreement
One agent says an office closes at 17:00 based on the regular schedule. Another says it closes at 13:00 based on a holiday exception. The claims conflict only if the intended date and location match.
Once that context is established, the approved exception policy determines authority. The final answer should identify the applicable hours, while preserving unresolved uncertainty if neither source can be verified. Averaging the times would be meaningless.
| Apparent conflict | Resolution check |
|---|---|
| Different dates | Align the question before comparing |
| Regular rule versus exception | Apply approved policy precedence |
| Old versus current source | Verify effective version |
| Caller allegation versus system fact | Preserve attribution and verification status |
| Two unsupported assertions | Do not manufacture certainty through agreement |
Build fixtures where differences are genuine contradictions and where they are merely different scopes. Evaluate false conflict alarms as well as missed material disagreements.
Keep resolution separate from execution
An agent may recommend the likely correct answer, but the execution service still checks permissions and prerequisites. A resolved factual dispute does not authorise a cancellation or disclosure by itself.
Structured claim comparison adds modelling work and may not capture every nuance. Prioritise the fields that change business outcomes, such as eligibility, dates, identity, and commitment status. Use human review for ambiguous consequential language rather than forcing a binary verdict.
This design is proposed. The voice integration case study describes related retrieval and workflow work. Pair it with uncertainty-preserving summaries.
Agent conflict FAQ
Should majority agreement decide the answer?
Not when agents may share evidence or repeat the same error. Check source authority and relevance.
Are different wordings always a conflict?
No. Compare meaning, subject, conditions, and time.
What if the conflict cannot be resolved?
State the uncertainty and use the approved clarification or review path before making a consequential claim or action.