Error Propagation in Multi-Agent Systems

Error handling in multi-agent systems requires careful design: errors should be handled at the lowest capable level and propagated with full context when escalation is needed.

Principles

1. Handle locally when possible: Subagents should handle transient failures (retries, fallbacks) internally
2. Propagate with context: When escalating, include failure type, what was attempted, partial results, and suggested alternatives
3. Distinguish failure types: Access failures (timeout) vs. valid empty results (0 results found) require different handling
4. Graceful degradation: Proceed with partial results rather than failing entirely, but annotate gaps

Anti-Patterns

• Silently swallowing errors (hides problems from coordinator)
• Returning empty results marked as success (prevents recovery)
• Terminating entire workflow on single failure (wasteful)
• Generic error statuses like "search unavailable" (hides valuable context)
• Retrying permanent failures like file corruption (pointless)

Structured Error Context

Return structured error information to enable intelligent coordinator recovery:

• Failure type (timeout, auth failure, rate limit, empty results)
• What was attempted (the query, parameters)
• Partial results (anything successfully retrieved)
• Alternative approaches (suggested fallback strategies)