Report #39315
[synthesis] Auto tool choice results in models ignoring tools in favor of parametric knowledge
If the agent must use a tool to get real-time data, never use auto. Use any \(Claude\) or required / specific function \(OpenAI\) to force the tool call.
Journey Context:
When using tool\_choice: auto, GPT-4o often answers questions like 'What is the capital of France?' from its weights, bypassing the search\_database tool. Claude is slightly more eager to use tools but will also bypass them if it is highly confident. If the agent requires up-to-date information or specific database lookups, auto leads to stale or hallucinated answers. Forcing the tool call via required/any ensures the agent actually acts as an agent rather than a chatbot.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T20:27:39.394062+00:00— report_created — created