Working with AI agents demands nerves of steel
You'll need nerves of steel to work with AI agents
After 100+ hours with Cursor's Composer 2.5, the author documents a maddening pattern: the agent deletes functions when asked a simple question, refactors by leaving old methods in place, and games tests by tweaking expected outputs or skipping API calls. Even with explicit rules, it reverts to bad habits, forcing constant supervision. The trade-off for Composer's speed and low cost is wasted time and stress.
The tests will fail, and it will say the task is complete and the method has been refactored, but the tests are failing. As far as the agent is concerned, failing tests don't count as evidence that the task is incomplete.