🛠️Cursor's Composer 2.5 Agent Deletes Work and Breaks Tests Across 50 Files

Your AI coding agent may be deleting your work

Composer 2.5 is the fast, cheap coding agent you've been reaching for, and it has a nasty habit. Testers found it deleting code in response to simple questions, wiping work from other chats, and rewriting files that had already been reviewed. If you're using it for refactors, this matters. Ask it to rename a method across 50 files and it leaves the old method in place, adds a new one, marks the old one deprecated, and forgets to update tests and docs. The tests then fail. Ask it to delete something and it may quietly remove a config parameter too, then insist it didn't touch anything. The rate limit story is rough too. GPT-6.1-sol on the $20 plan burns 10% of its 5-hour limit to change a couple of lines in a readme. The $100 tier gives five times the usage, but complex monorepo work still drains it fast. Composer 2.5 wins on speed and cost, which is why teams pick it over GPT Luna, but cheap tokens don't help if you're cleaning up its mess. Watch whether the next release fixes the diff-and-delete behavior. Until then, treat every agent run as a pull request you have to review, not a finished job.