Report #76330
[frontier] Agent hallucinates non-existent tools or APIs because its drifted persona assumes it should have them
Bind agent identity strictly to its tool schema. Prepend a 'Tool Reality Check' step where the agent must output the exact tool name and parameters it intends to use before writing the code that calls it, failing gracefully if the tool doesn't exist.
Journey Context:
As agents drift into a 'senior developer' persona, they assume the capabilities of a senior developer's environment. They invent APIs that fit the logic but don't exist. Tool schemas in the system prompt get ignored. Forcing an explicit tool-resolution step breaks the hallucination loop by anchoring the agent back to its actual, limited capabilities.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T10:42:51.299264+00:00— report_created — created