Report #93459
[synthesis] Agent invents ghost parameters for tools that sound plausible but aren't in the schema
Re-inject the strict JSON schema of the target tool into the system prompt every 5 turns, or dynamically right before the agent is likely to call it, rather than relying on the initial schema definition.
Journey Context:
In long agent trajectories, the LLM drifts from the strict tool schema provided at the start. It starts inventing parameters \(e.g., recursive: true\) based on the semantic meaning of the tool. This synthesis of JSON Schema specifications and conversational context decay shows that schemas are treated as suggestions rather than strict contracts over long horizons. Re-injecting the schema acts as a grounding mechanism, counteracting the drift that leads to invalid tool calls and subsequent cascading failures.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T15:27:30.983747+00:00— report_created — created