Deliver useful findings with visible coverage gaps

Learning objective

Synthesize qualified partial results with explicit coverage gaps rather than terminating an entire multi-agent workflow after one failure.

Core idea

For an architect designing the final research deliverable, partial completion is useful when readers can tell what the evidence supports and what remains unchecked. One failed specialist should not automatically erase successful work from independent specialists. Equally, a fluent final report must not hide that the investigation missed part of its scope.

The coordinator decides whether the surviving findings serve the user’s goal. Synthesis then preserves the limits of that decision. Every relevant topic needs an explicit account of support or a coverage gap; an unavailable source must not become an unsupported negative conclusion.

Anthropic’s multi-agent research article describes a lead agent synthesizing specialist results and deciding whether more research is needed. It also describes clear specialist output requirements. Apply those mechanisms by passing failures as well as findings into synthesis and requiring coverage annotations in the output. These annotations are your application contract, not an Anthropic API field.

How it works

Pass the missing work into synthesis

The parent receives a subagent’s final message while intermediate calls and results remain in the subagent context. Therefore, synthesis cannot reliably infer why a topic has no evidence unless the coordinator supplies that context. A summary containing only successful findings removes the distinction between an unassigned topic, failed access, and a completed zero-match search.

Give the synthesis assignment the requested topic list, usable findings, query outcomes, and unresolved source failures. Use AgentDefinition.prompt to define the synthesizer’s output requirements and the delegation prompt to pass the investigation’s actual outcomes. See 1.3.3 for context transfer and 5.3.2 for the failure-report contents.

The coordinator should preserve a missing topic even when it chooses not to recover it. “Proceed with partial results” changes the coverage of the deliverable; it does not change the source outcome from failure to success. Passing that topic onward lets the synthesizer describe the limitation instead of guessing a result.

Annotate topics, not just the report

A single “some data unavailable” sentence tells the reader little about which conclusions remain usable. Attach coverage annotations to the relevant topic or claim group. Describe the evidence supporting the finding and the source or operation that remains unresolved.

For example, your application can require topic, coverage_status, support, and gap_reason in its synthesis record. These illustrative names are not SDK response fields. The status should distinguish supported findings, limited support, completed searches with no matches, and unresolved access failures.

The following table shows how those distinctions survive into a reader-facing comparison:

The City C row is resolved as a search outcome, not as a universal absence claim. The Country B row is unresolved access, not evidence of zero admissions or no official publication. See 5.3.1 for the distinction at the source boundary.

Keep claims inside surviving evidence

A supported topic does not make every comparison supported. If Country A has figures and Country B is unchecked, do not publish a complete cross-country ranking. You can report Country A’s figures and the supported regional trend while explicitly withholding the comparison that needs Country B’s evidence.

Partial synthesis requires narrowing claims as well as listing gaps. A caveat at the end cannot repair a paragraph that already states an unsupported comparison as fact. Put the limitation where the reader encounters the affected claim.

Coverage and confidence answer different questions. Coverage asks which requested areas were investigated and what evidence is available. A confidence score cannot establish that an unavailable source was checked. Keep confidence calibration and review decisions in 5.5; keep detailed provenance and source conflicts in 5.6.

Preserve qualifications through the final handoff

If a separate reporting specialist turns the synthesis into prose, include the coverage annotations in its assignment and output requirements. Otherwise, a later rewrite can remove the very distinctions the synthesizer preserved. Use the reporting specialist’s AgentDefinition.prompt to require retaining topic-specific limitations.

Anthropic’s research article describes passing findings to a citation agent after research and synthesis. That supports attribution of claims, but a citation alone does not establish complete coverage. A perfectly cited report can still omit an inaccessible topic. Require both supported claims and visible missing scope.

The objective rejects automatic termination after one failure when useful independent results survive. If the user’s requested conclusion depends on the missing evidence, explain that the conclusion remains unresolved. You can deliver the supported portion without asserting that the original question is fully answered.

Key distinctions and trade-offs

Useful partial completion versus silent incompleteness: Both contain fewer findings than planned. Only qualified partial completion exposes the missing scope and limits affected conclusions.

Topic coverage versus source citation: A citation supports attribution for a claim. Coverage annotations reveal which requested areas have evidence and which do not.

Independent finding versus dependent conclusion: An unaffected specialist’s finding can remain valid after another fails. A conclusion that requires both specialists’ evidence cannot be completed from only one.

Partial delivery preserves useful work and can satisfy time-sensitive needs, but it transfers the remaining limitation to the reader. That trade-off is acceptable only when the coordinator chooses it and the output makes it visible. More research may improve coverage; withholding the whole report may be necessary when the unsupported dependency is the deliverable’s central decision.

Architecture / decision scenario

A coordinator commissions a report on heat-response planning, admissions trends, and official figures for two countries. The policy specialist and trend specialist finish. The statistics specialist retrieves Country A’s release but cannot access Country B’s source after local recovery. A separate city-specific publication search completes with no matches.

The coordinator considers the recoveries taught in 5.3.3 and chooses partial completion when they cannot resolve Country B’s slice. It sends the synthesizer the original topic list, successful findings, Country B’s access failure, and the completed city search. It does not send only the successful documents.

The report presents the supported policy findings and regional trend. It reports Country A’s official figures with their source. Beside the country comparison, it states that Country B’s figures were not assessed because the statistics source was unavailable, so the report cannot support a complete comparison of those figures.

The city row states that the named catalogue query found no matching publication. It does not label the completed search as an access gap or claim that the city has no heat-response plan. The reporting specialist retains these annotations in the final prose.

The reader receives useful independent findings and a precise account of unanswered work. Automatically terminating the entire workflow would discard the policy and trend research. Removing the gap would be equally defective: it would let the report imply completeness it did not earn.

Common exam traps

• Terminating all research automatically after one independent source fails.
• Passing only successful findings to synthesis and expecting it to infer missing coverage.
• Adding a generic disclaimer without identifying affected topics or conclusions.
• Using a confidence score as proof that an unavailable source was checked.
• Publishing a cross-source comparison that depends on unresolved evidence.
• Removing coverage annotations during the final reporting handoff.

What to remember

Keep useful independent findings, but retain every unresolved topic through synthesis and reporting. Annotate support and gaps where they affect the reader’s conclusions. Qualified partial completion preserves evidence and its limits; silent incompleteness preserves only the appearance of an answer.

Knowledge check

References

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