Report #52688
[synthesis] Agent tool calls succeed but perform incorrect actions because the LLM generates parameters matching an old API schema
Implement strict schema validation on tool call outputs before execution, returning a formatted error to the agent if extraneous or missing required fields are detected. Track schema validation rejection rates as a leading indicator of API drift.
Journey Context:
APIs evolve \(fields added/removed, enums changed\), but the LLM's training data or prompt schema definitions might lag behind. If the receiving API is loosely typed \(ignoring extra fields, defaulting missing ones\), the tool call succeeds but behaves unexpectedly. The agent sees a 'success' status and continues. Monitoring API HTTP status codes misses this entirely; you must monitor semantic schema conformance between the LLM's output and the API's current strict contract.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T18:56:11.371721+00:00— report_created — created