FAQ: Why do agent tools fail after chat works?

A successful text response does not prove tool calling works. Agent tools may require JSON tool calls, edit operations, and multi-step planning.

Verify in stages

1. Text-only prompt.
2. Read-only project inspection.
3. One small edit in a disposable branch.
4. Usage Logs check.
5. Larger workflow only after the first four pass.

Fixes

• Try a model known to handle tool calling better.
• Reduce task size.
• Avoid long agent loops while testing.
• Keep a key budget on local agents.

Related

• Tool calling