Report #44632
[synthesis] Agent hallucinates parameters for a tool based on training data rather than the provided schema, leading to silent API defaults and incorrect results
Implement strict schema validation on tool call outputs at the orchestrator level. If the agent provides an undeclared parameter, reject the call with a hard error: 'Parameter X is not defined in the schema. Available parameters are: Y'. Do not silently drop the parameter.
Journey Context:
LLMs often rely on pre-trained knowledge of popular APIs. The synthesis of tool schema staleness, LLM prior knowledge overriding current schema, and silent API defaults creates a confident hallucination. If your tool schema wraps or slightly modifies an API, the agent might hallucinate a parameter it 'knows' should exist. If the server silently ignores unknown parameters, the agent receives a 200 OK but the tool behaves differently than expected. The agent then tries to fix the perfectly good surrounding code. A hard rejection forces the agent to align with the actual schema.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T05:23:07.724577+00:00— report_created — created