Error propagation across multi-agent systems

Two common error-handling mistakes

• Silent suppression — a failed search returns empty results as success. The coordinator believes the topic was covered and found nothing. This is worse than failing, because the gap is invisible and the final answer is confidently incomplete.
• Total termination — one subagent failure ends the whole workflow, throwing away everything that succeeded.

Both replace a judgment the coordinator should make with a decision the subagent had no standing to take.

Your group project partner is assigned the archive chapter, can't get into the archive, and says nothing. The paper then reads as though that chapter held nothing worth citing. Silence and "nothing there" arrive looking identical.

What to return instead

Structured error context:

• failure type
• what was attempted — the query, the source
• partial results already gathered
• potential alternatives worth trying

That's enough for the coordinator to decide intelligently: retry, route elsewhere, or proceed with a documented gap. A generic "search unavailable" hides every one of those and leaves it guessing.

The two payloads side by side. The field names are a convention you pick, not an API contract:

silently suppressed — indistinguishable from "nothing matched":
{ "results": [] }

structured error context the coordinator can act on:
{ "failure_type": "access_denied",
  "attempted": { "source": "acm-library", "query": "graph neural nets" },
  "partial_results": ["arxiv:2403.01121"],
  "alternatives": ["semantic-scholar", "arxiv"] }

Both are a subagent reporting a failed search. Only the second lets the coordinator tell a gap from a finding, and it keeps the one paper that did come back rather than throwing it away with the error.

Access failure is not an empty result

The same distinction as 2.2, and it matters more here because the coordinator is deciding coverage.

• Access failure — could not reach the source. Needs a retry decision.
• Valid empty result — reached it, nothing matched. That's an answer.

Conflating them means either retrying a question already answered, or reporting "no findings" for a source never consulted.

Handle recoverable errors locally

Subagents handle transient failures themselves and propagate only what they could not resolve, carrying partial results and a record of the attempt.

Annotate coverage in the synthesis

The final output should say which findings are well-supported and which topic areas have gaps because sources were unavailable.

The final answer must make these gaps clear. If a third of the sources failed, a uniformly confident summary can mislead the reader even when each sentence is technically true.