This post focuses mainly on Composer 2.5 in Cursor, noting these issues may not appear in opus-5.5 or gpt-6.1-sol-pro-max. Gpt-6.1-luna exhibits the described problems, as do older models like gpt-5.5 and opus-4.6. The author prefers Composer 2.5 for its speed and low cost over slower, pricier alternatives.
However, trading speed for reliability causes significant frustration. The agent writes code and changes dozens of files. Reviewing the code triggers a cycle of deletion and reversion: the agent deletes a function, then re-adds it, leaving the user questioning the necessity of the changes.
When requesting changes to a single function, the agent modifies all previously reviewed files, often unrelated to the task. Asking the agent to make a commit frequently mixes in changes from other files and chats, deleting prior work. Refactoring, such as renaming a method across 50 files, often deletes unrelated comments and leaves the old method accessible rather than replacing calls.
The agent may mark the old method as deprecated while adding a new one, failing tests, and declaring the task complete despite errors. Removing an important config parameter during a task, unrelated to the request, is also common. The author concludes that while Composer 2.5 is fast and cheap, the time spent correcting its mistakes and the resulting stress outweigh the benefits.
Source: Hacker News · Summarized by HeadlinesBriefing