My experience the past weeks has been a mess because I ask simple tasks like change one css line and the auto agent (or composer 2.5) starts to change the whole html. There are days where it works perfectly but most of the days feels like talking with a wall
Composer 2.5 is the best of price/performance. I use it for my daily tasks.
Composer 2.5 is fantastic. It’s an incredibly intelligent model, and it’s super fast. I’m exclusively using it now.
If I could change anything, I would make the model slightly more verbose in explaining what it’s doing or what it’s proposing. I realize this costs tokens, but for someone who is trying to learn how to engineer with LLMs, it really helps me optimize the harness wtih more insight into how the model reasons/behaves. There’s a sweet spot there; I think Opus is too verbose
I like Composer 2.5 but I definitely have to cleanup after the work it does even with things like strict TDD in place. It sort of leaves a mess everywhere it works and needs GPT 5.5 et al to cleanup after it.
Composer 2.5 is great, and highly cost-effective, but still, it sometimes could leave a lot mess and i will need to switch to better model like opus4.8 to clean it up. This is the most worrisome part since if i need to clean up with other models, this means that i did not save any token while im using 2x of my time coding.
Overall composer2.5 is a great model, 90% it can finish the task flawlessly, but 10% cases it will just failed, and failed and failed and never realising that they are wrong though i have given a very clear suggestion. I am looking forward to see composer3! I think the most important is not to excel in any benchmark(also not possible for an economical model), but just make the flaw less and make it smarter, so i will feel ‘safer’ to use 
Composer 2.5 has been fantastic and is my primary model since release. I haven’t even been tempted to try another model. It’s behaviour is spot on.
My codebase is React frontend with Typescript backend (Cloudflare workers). I know the codebase well and will often tag the relevant file.
My prompting style is relatively lazy, however Composer always understands my intent. It tends to write tests for most things it implements and it doesn’t perform endless unnecessary tool calls unlike some other models.
One thing that stood out, often for a relatively simple change it will do git diff to see when it was last changed. I appreciate this can add context most of the time but I don’t think it needs to do it every single time.
I’ve used Composer 2.5 before as a Verifier for AI-teamwork, but now I’ve configured it as a recommendation hook and rules that:
- Cursor recommends that even single Agent run
/Verifierafter CI for any code change audit - Deployed Rutherford MCP so that Codex and Claude can also call Cursor Composer 2.5 as revievers.
And it often finds errors that Grok, Fable, and Sol make. I regret not doing this a few months earlier 
Several instances so far of this malvolence:
- what color is A and B?
- hyper explanatory wall of text
- ok, then change A to blue and B to red
- done, finished, implemented
- check A and B colors
- A is red and B is blue
One drawback I’ve noticed so far is that it frequently creates unnecessary methods. It tends to extract something that could be done in just two or three lines of code into a separate method. As a result, even a feature that doesn’t require much code ends up with many nested methods, which unnecessarily increases the overall code length. PHP PHP