Report #56520
[synthesis] Agent retries spike on specific external APIs without throwing hard failures in the orchestration layer
Implement strict schema diffing between the LLM's generated tool call arguments and the target API's OpenAPI spec, and alert on structural mismatches even if the LLM self-corrects on retry.
Journey Context:
External APIs evolve \(e.g., a field becomes required, or an enum changes\). The LLM doesn't know this and generates payloads based on its training data. The API returns a 400. The agent sees the error, tweaks the payload slightly, and tries again. The orchestration layer sees a successful retry and logs it as a transient network issue. In reality, the agent is brute-forcing the API schema. Without diffing the generated schema against the actual API spec, teams misdiagnose this as API flakiness rather than LLM knowledge staleness.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T01:21:37.695727+00:00— report_created — created