Implement error propagation strategies across multi-agent systems

Report a failed subagent as "search unavailable" so the coordinator can retry.

It discards what was attempted, what partially succeeded and what else might work, so the coordinator can only retry blindly or give up.

Failure type, attempted query, partial results and alternatives — enough for a recovery decision.

Return an empty result when a subagent fails, so the pipeline keeps moving.

Suppressing the error is worse than failing: synthesis treats absence as evidence and reports a confident conclusion built on a gap nobody can see.

A failure must remain visible. Degrade the answer, never the record of what happened.

Abort the workflow when a subagent fails, so nothing is built on incomplete data.

One transient timeout discards every other agent's completed work, and the retry re-pays for all of it.

Recover locally where possible; otherwise continue with the gap annotated.

Treat a query with no matches as a failed search.

It triggers retries and fallbacks over a correct answer, and can escalate a question that was already resolved.

Empty is a successful result. Only inability to execute is an error.

Present the synthesis uniformly once the run completes.

Well-supported findings and areas where the sources never loaded read identically, so the reader cannot tell which conclusions to trust.

Annotate coverage — say which areas are well supported and where a source was unavailable.