Home / CCA-F / Context Management / Multi-Agent Failure Modes — Detail & Examples
DWG 5.2 — DETAIL & EXAMPLES

Multi-Agent Failure Modes — Detail & Examples

Part of Context Management · overview: domain-5.html
In this page

A full four-part structured error report, the two anti-patterns compared side by side, and where retries belong (subagent) versus strategy decisions (coordinator).

5.2aThe four-part error context

WORKED EXAMPLE · A SUBAGENT REPORTING A FAILURE WELL
SCENARIOA research subagent was asked to pull recent case studies from an academic database and hit a snag partway through.
  1. Failure type: transient — the database returned a rate-limit error, not a permanent rejection.
  2. What was attempted: "Searched the academic database for 'renewable energy policy 2024-2026', page 2 of results."
  3. Partial results: 3 of 5 targeted sources were already retrieved successfully before the rate limit hit.
  4. Alternatives: retry after a short delay, or proceed with the 3 sources already gathered and flag the gap.
Why this matters:The coordinator can now decide intelligently — wait and retry, or move forward with partial results — instead of just knowing "something failed."

5.2bThe two anti-patterns, side by side

Which is worse, and why
  • Silent suppression: the coordinator treats an empty/failed result as a valid "nothing found" answer. This is the worse of the two — it prevents any recovery from ever being attempted, because the coordinator doesn't even know there's a problem.
  • Workflow termination: one subagent's failure kills the entire pipeline, discarding perfectly good partial results other subagents already produced. Wasteful, but at least visible — someone notices the whole thing stopped.
  • The fix for both is the same structured error context above: it makes a real failure visibly different from a real empty answer, and lets the coordinator target just the failed piece instead of discarding everything.

5.2cWhere retries belong

SUBAGENT (local retry)Timeout on first attempt → retries automatically once with backoff → succeeds. Coordinator never even sees this happen.
COORDINATOR (strategy-level)Subagent reports a permission failure (access denied) — this can't be fixed by retrying, so the coordinator escalates for a human to grant access, or picks a different data source entirely.

Claude Certified Architect — Study NotesDWG 5.2