5.3 Error propagation across multi-agent systems
What a propagated error must carry
{
"status": "error",
"failure_type": "timeout",
"attempted_query": "AI adoption in film production 2024-2026",
"partial_results": [
{ "claim": "…", "source": "https://…", "published": "2026-01-30" }
],
"alternatives": [
"Narrow to 2025-2026 and retry",
"Substitute the industry-press source set"
]
}
Failure type, what was attempted, partial results, potential alternatives — these four give the coordinator enough to decide: retry with a modified query, try an alternative approach, or proceed with partial results.
Anti-patterns
| Anti-pattern | Why it fails |
|---|---|
Generic status ("search unavailable") | Hides context from the coordinator; no informed decision possible |
| Silent suppression — returning empty results marked successful | Prevents any recovery and risks silently incomplete output |
| Terminating the whole workflow on a single subagent failure | Discards recovery strategies that would have worked |
| Retrying internally with backoff, then returning only a generic status | The retries are fine; the generic status throws away the context |
Local recovery, then propagate
Subagents implement local recovery for transient failures. Only errors they cannot resolve propagate — with what was attempted and partial results attached.
Access failure vs valid empty result
The coordinator must be able to distinguish "the source was unreachable" (a retry decision) from "the query ran and matched nothing" (a fact). See Domain 2.2.
Coverage annotations
Structure synthesis output so it marks which findings are well-supported and which topic areas have gaps due to unavailable sources. A report that silently omits an unreachable area is worse than one that names the hole.
Quiz
- Question 1
The web search subagent times out while researching a complex topic. Which error propagation approach best enables intelligent recovery by the coordinator?
- Question 2
A document-analysis subagent times out on one of the four documents it was given. It returns {"status": "analysis unavailable"} and nothing else. The coordinator responds by re-running all four documents from scratch. What should the subagent have returned?
- Question 3Select all that apply (2)
Two of the six source types a run was meant to cover were unreachable for its entire duration. The other four produced good material, and the final report reads as a complete survey with no indication that anything is missing. Which two statements are correct? Select two.
- Question 4
Your enrichment step queries a vendor database and returns an empty object in two different situations: the vendor is genuinely not in the database, and the database connection failed. Downstream, unknown vendors and outages both become "vendor: unknown" in the final record, and an afternoon-long outage produced 4,000 records that look like legitimate unknowns. What is the fix?