Guide · Agent orchestration
When AI Agents Disagree, Ask for a Test or a Source
Resolve conflicting agent findings through requirements and evidence rather than repeated debate or majority voting.
One agent says the patch is safe. Another says it breaks access control. Asking them to discuss it for several more turns may produce a longer conversation without settling the issue. The coordinator needs to turn the disagreement into a question that evidence can answer.
Identify the exact point of disagreement
For a hypothetical customer portal, the implementer believes a query is already scoped to the current account. The reviewer believes it can return another account's records. Write that claim explicitly. Avoid debating broad statements about whether the implementation is "secure" or "well designed."
Ask each worker to point to the relevant requirement and code path. If they are looking at different revisions, resolve that first. A surprising number of disagreements become ordinary version mismatches once the artifacts are identified clearly.
Find the smallest deciding check
In the portal example, create a controlled test with two accounts and request a record from the wrong one. Inspect the actual response and the query behavior. The result gives the coordinator something concrete to evaluate against the access requirement.
For research, the deciding check may be a primary source that specifies the feature or version. For a writing disagreement, it may be the approved brief or the customer's stated preference. Use the kind of evidence that fits the question instead of forcing every disagreement into a code test.
If the evidence is unavailable, preserve the uncertainty. The coordinator can ask for the missing requirement, narrow the claim, or pause the affected action. It should not manufacture certainty simply because a final answer is expected.
Do not count agreement as independent proof
Several workers can repeat the same mistaken assumption, especially when they share the same summary or source. Record where each conclusion came from. Three summaries of one outdated page still depend on that one page.
When using Astra and Fable 5.1 together, different providers can offer another perspective, but a cross-model agreement is still not a substitute for evidence. Judge the argument by its support and its fit to the requirement, not by the number or brand of models that repeat it.
Close the loop after the decision
Once the disagreement is resolved, save the deciding evidence and the accepted conclusion. Update the artifact or assignment that caused the confusion. Tell any still-running worker which revision or interpretation now applies so it does not reintroduce the old issue.
Use a bounded review loop. One precise clarification and one deciding check can be more useful than repeated requests to reconsider. If the issue remains unresolved, return it to the responsible person with the evidence already gathered.
Review recurring disagreements in the workflow itself. Repeated arguments about the same access rule suggest a missing specification. Repeated comments on outdated files suggest weak revision tracking. Fix the underlying gap so future workers have a better starting point.