The main pain point of Cursor Composer/Agent for me right now is that it’s pretty shallow in finding the right context to use when you don’t specify it yourself. If Cursor could solve this problem, going deeper into a large codebase, exploring it and adding the right files automagically, I would gladly pay as much as the ChatGPT Pro sub. This is where I lose most of my time, just getting the right files in. I know that with todays models context length is still very limited and the models struggle when you fill it, even for Gemini models (the codebase I’m working in has hundred million tokens more or less).
For now, in my opinion, Cursor still has moat, but if OpenAI (ChatGPT Pro tier specifically, o1 pro, o3 pro mode in the future) releases an agent/operator that can interact with the local file system, git etc… I think it will be GG if something doesn’t change.
We actually have auto-context in beta right now which should help with exactly this! It automatically finds and includes relevant files based on your queries
You can enable it in Settings > Features > Auto Context (Beta)
This should save you a ton of time vs manually specifying files. Let me know if you run into any issues with it
For large codebases specifically, you might also want to check out our codebase indexing docs: Cursor – Codebase Indexing - this helps Cursor better understand relationships between files and find relevant context
Thank you for the answer, but I was writing with auto context in mind and codebase indexing as well. It is shallow search, the agent maybe does a few lookups for references and thats it, which is almost never enough. Again, with a simple codebase sure, but a large, complex, enterprise one, it’s not enough.
For large codebases where the limitation is the context window of the underlying model, the best practice here is often to keep a file / set of files that describe the structure of the project to the AI, so it does not have to scan your files and infer that structure each time you start fresh.
Many users get the Agent to keep this up to date itself, such as a memory.md
file that is can read and write too!
3 Likes
I saw that suggestion pop up a few times, I think that would be a nice feature that Cursor could do automatically perhaps? The less manual work the better. 
I think this is definitely on our workload for the future, but executing this in a good way is something we’d really want to get right!
1 Like