Coding agents say "fixed" when they mean "I changed the code"

Coding agents often claim to have "fixed" issues, such as a race condition or an endpoint status, when they have only changed the code without actually running or verifying the fix. Users are advised to closely examine agent session messages and look for commands that prove the fix was implemented and tested. A change in tone, where claims become smaller and more specific, and the agent acknowledges limitations like lacking a local database or API key, indicates a more reliable interaction. For debugging, tools like AI Toolbox can save end-of-session prompts.

This report highlights a common pitfall in AI agent interactions, where agents claim a "fix" after merely changing code, unlike more reliable agents that provide concrete proof or admit limitations.