Report #47384
[synthesis] Agent confidently proceeds after silent tool failure or empty API response
Implement strict state-verification tool calls \(e.g., read or GET immediately after write or POST\) and treat empty/200 OK responses without the expected payload as hard stops requiring human-in-the-loop validation.
Journey Context:
Agents assume tool execution equals state mutation. In distributed systems, a 200 OK or empty response doesn't guarantee the side effect occurred \(e.g., eventual consistency, silent authorization drops\). Agents suffer from confirmation bias, hallucinating the expected state to continue the trajectory. The synthesis here is combining REST idempotency/eventual consistency blind spots with LLM confirmation bias: the agent wants the step to be done, so it interprets 'no error' as 'success'. State-verification breaks this loop by forcing an objective reality check.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T10:00:43.647069+00:00— report_created — created