Return conflicting values for coordinator review

Learning objective

Preserve incompatible credible values with attribution and let the coordinator decide how to reconcile them.

Core idea

When credible sources disagree, the analysis is not complete if it hides one value. Complete the analysis with both values, their supporting evidence, and an explicit description of the conflict. The coordinator then decides what the research question requires before sending instructions to synthesis.

For a multi-agent research architect, this is an authority boundary. A specialist reports what its evidence supports; the coordinator decides whether to investigate further, present alternatives, or select a value using a stated criterion. A fluent summary must not make that decision by omission.

How it works

Establish what is incompatible

First apply the date and comparability checks in 5.6.2. Different observation periods can explain different values. Different populations or definitions can also make apparent disagreements dissolve into distinct measures. Do not label every unequal number a factual contradiction.

If two credible records still make incompatible assertions about the same quantity and period, keep both. If they use different methods to answer the client’s broader question, preserve that distinction too: they may be competing indicators rather than contradictory measurements. The coordinator needs that distinction to interpret the gap.

For example, two fictional sources estimate active deployments at 120 and 165 for the same date and definition. That is a direct conflict. A count of installed systems and a count of shipped units are different measures. Neither pair should become an invented midpoint, but the explanation attached to each pair differs.

Return a conflict record rather than a winner

Extend the evidence contract with an application-defined conflict_id, references to the participating claim_id values, a conflict_reason, and an unresolved resolution_status. Each original claim retains its own value, unit, date, excerpt, source location, and methodological qualifier. These are your pipeline’s fields, not native Agent SDK conflict-resolution options.

Use AgentDefinition.prompt to require the specialist to return all credible competing values and identify the comparison that remains unresolved. Ask it to complete the useful analysis around the disagreement. A conflict does not require discarding unrelated supported findings or treating the whole assignment as a failed tool call.

Anthropic’s hallucination guidance recommends grounding claims in quotes, checking support, and permitting uncertainty. Apply those instructions to both sides. A record should explain what the source actually says and where, rather than asking the model to choose the number that feels more plausible. An unsupported assertion does not gain credibility merely by being placed beside a supported one.

Preserve the source’s own status words. If one estimate is preliminary and another is audited, those labels inform the coordinator’s choice. Do not strengthen either source during paraphrase, and do not substitute a model-generated confidence score for the missing comparison.

Make the coordinator’s decision explicit

Anthropic’s multi-agent research architecture gives the lead agent responsibility for synthesizing results and deciding whether more research is necessary. For this unit’s conflict-handling design, express the coordinator’s choice as an application-defined resolution linked to the conflict record. Include the chosen treatment, reason, and any follow-up request.

The coordinator can request targeted research into definitions or source methodology. It can instruct synthesis to show both estimates as unresolved. It can select one for the client’s specified measure when evidence supports that choice, while retaining the competing record and the selection reason. These are design choices, not a built-in three-value API enum.

Pass the decision together with the evidence to synthesis. A non-forked Agent SDK subagent does not receive the parent’s conversation; include the necessary decision in the Agent tool’s prompt string or provide an accessible artifact containing it. Sending two values without the coordinator’s treatment leaves synthesis to invent its own reconciliation.

The specialist reports the conflict, the coordinator chooses its treatment, and synthesis receives that decision with the evidence.

If the deadline prevents further research, an attributed unresolved comparison can be a valid result. State what remains unresolved and how it affects the answer. Do not turn time pressure into permission to average incompatible quantities or erase the less convenient source.

Key distinctions and trade-offs

Disagreement versus analysis failure: A completed analysis can identify unresolved disagreement. Failed access is a different outcome, covered in 5.3.1; it does not provide a competing measurement.

Selection versus suppression: A justified selection states which source fits the requested measure and why. Suppression removes the alternative without leaving a decision trail. The former supports review; the latter hides uncertainty.

Reconciliation versus averaging: Reconciliation can explain different methods or choose an appropriate measure. An arithmetic mean is justified only by an explicit analytical method appropriate to the data, not by a desire to remove disagreement. An arbitrary midpoint is not evidence from either source.

Keeping alternatives adds detail to the report, but supports a more precise answer. Additional research costs time; showing an unresolved comparison preserves useful findings when the evidence cannot settle the issue within the assignment.

Architecture / decision scenario

A coordinator asks whether a fictional service has at least 150 active enterprise deployments. One specialist finds a vendor statement reporting 165; another finds an independent registry reporting 120. Both describe the same reporting date, and both sources appear credible, but their inclusion rules are unclear.

The analyst returns both values, their excerpts and locations, and a conflict record stating that the population definitions need verification. It completes the remaining document analysis rather than returning only an error or silently preferring the vendor statement.

The coordinator requests methodology passages from both sources. The follow-up establishes that the vendor includes pilot deployments while the registry counts only production installations. Because the client asked about production, the coordinator instructs synthesis to use 120 for that threshold assessment, explain the inclusion difference, and retain 165 as the broader count. The answer no longer frames the pair as two measurements of an identical population.

If the methodology passages remain unavailable, the coordinator instead instructs synthesis to present both attributed estimates and say the threshold cannot be established from this evidence. Either path makes the decision visible before synthesis, while preserving the original records.

Common exam traps

• Selecting the more authoritative-sounding source without comparing definitions.
• Averaging incompatible values to create a single confident answer.
• Removing a credible competing value from the specialist’s output.
• Stopping all analysis because one quantity remains disputed.
• Passing an unresolved conflict to synthesis with no coordinator instruction.
• Treating different measures as a factual contradiction after methodology explains the gap.

What to remember

Return credible competing values intact, with attribution, dates, methods, and an explicit conflict annotation. Let the coordinator decide the treatment and pass that decision to synthesis. A transparent unresolved result is stronger than a hidden selection or an invented compromise.

Knowledge check

References

• Anthropic, Subagents in the SDK
• Anthropic, Reduce hallucinations
• Anthropic, How we built our multi-agent research system