Error Propagation In Multi-Agent Systems

Definition

The strategy for how errors flow between subagents and a coordinator: subagents recover locally where possible and propagate structured error context (not generic statuses) so the coordinator can recover intelligently.

Key points

• Return structured error context — failure type, alternative approaches, what-was-attempted, and partial results — rather than a generic error status that hides context.
• Use structured metadata: isError / errorCategory (transient / validation / permission) / isRetryable (a.k.a. retriable: false) — see Structured tool errors.
• Subagents do local recovery for transient failures; propagate to the coordinator only unresolvable errors, with partial results + what was attempted.
• Distinguish access failures (e.g. timeouts needing a retry decision) from valid empty results (successful query, no matches).
• Structure synthesis output with coverage annotations (well-supported vs gap topic areas).
• Anti-patterns: silently suppressing errors (treating empty-as-success) or terminating the whole workflow on a single failure.

Why it matters for the exam

• D5.3 error-propagation scenarios in multi-agent systems. Tests local-recovery-then-propagate, structured context over generic statuses, and the empty-result-vs-failure distinction.

Common gotchas

• Empty result ≠ failure — treating a valid empty query as an error (or vice-versa) is a trap.
• Killing the whole workflow on one subagent failure instead of partial-results propagation.
• Bubbling a raw/generic error up instead of structured context the coordinator can act on.

See also

• Structured Tool Errors
• Subagent Error Handling
• Stratified Sampling And Confidence Scores
• Subagent decision criteria

Sources

• certsafari
• architects-playbook
• netlight-study-guide