What you need to know
What the coordinator needs to recover
A coordinator can only make a good decision if the failure report says enough. Compare:
| Report | What the coordinator can do |
|---|---|
| "search unavailable" | Guess |
{"failure": "timeout", "attempted": "EU e-bike sales 2025 site:eurostat", "retries": 2, "partial": [3 results], "alternatives": ["try industry body reports", "narrow to Germany"]} | Retry differently, switch source, accept partial coverage, or report a gap |
Structured error context carries the failure type, what was attempted, any partial results, and possible alternatives.
Access failure or empty result?
- Access failure: the query did not run (timeout, permission denied). The coordinator must decide whether to retry or work around it.
- Valid empty result: the query ran and found nothing. That is information, not an error.
Reporting them the same way leads either to pointless retries or to hidden gaps.
Two anti-patterns
- Silent suppression: returning an empty result as success when the search failed. The report then states conclusions with a gap nobody knows about.
- Abort on any failure: terminating the whole workflow because one subagent failed throws away all the work that succeeded.
Recover locally first
Subagents should retry transient failures themselves (a timeout, a rate-limited call), and propagate only errors they cannot resolve, including what they tried and what they did get. This keeps the coordinator's context for decisions, not for routine retries.
Make gaps visible in the output
The final synthesis should carry coverage annotations: which findings are well supported (several sources), which rest on one source, and which topic areas have gaps because sources were unavailable.
In the exam
Scenario 3: a subagent times out, the report omits a topic with no explanation, a generic error status, or a workflow that stops on the first error.
Check your understanding
A web search subagent times out on a complex query. What should it send the coordinator?
Show answer
Answer: C. Structured error context lets the coordinator choose a recovery: retry differently, switch source, or record a gap.
Which TWO are anti-patterns in multi-agent error handling? Select TWO.
Show answer
Answer: A, C. Silent suppression hides gaps; aborting on any failure discards good work. Local retries and partial results are good practice.
A final report covers some topics thinly because two sources were unavailable. How should the synthesis show this?
Show answer
Answer: A. Readers need to know where the evidence is weak. Coverage annotations make gaps explicit in the output itself.
Apply it: Whitfield's error contract
Whitfield's coordinator receives "search unavailable" and the report never mentions the gap. Design the error handling.
- Write the subagent's local recovery rule.
- Write the error payload for a timeout that still returned three results.
- Write the coordinator's decision rule for that payload.
- Show the coverage annotation in the final report.
Show model answer
1. Local recovery. On timeout or rate limit, retry twice with backoff, then try one narrower variant of the query. Only then report upward.
2. Payload
{"status": "partial_failure", "failure_type": "timeout",
"attempted": ["EU e-bike sales 2025", "EU e-bike sales 2025 Germany"],
"retries": 2, "partial_results": [/* 3 findings with sources */],
"alternatives": ["industry association annual report", "national statistics offices"]}3. Coordinator rule. If alternatives exist and the subtopic is central to the question, delegate one alternative. If not, accept the partial results and mark the subtopic thin. Never treat partial failure as complete coverage.
4. Annotation. "Coverage: Regulation and supply chain are well supported (5+ independent sources). Market size rests on three sources from one search; two searches timed out and national statistics were not consulted. Treat market size figures as indicative."