What's your experience with coding AI agents

[ExtReMLapin]

Hey,

Been using Gemini and it's 1M context window for months to debug massive code bases.

It seems that they recently nerfed it and turned it into a rag instead of keeping the 1M context window, which makes it pretty much braindead.

I see that you're using Claude Code (CLI ? Web interface ? Something else)

Is it your favorite pick, what else did you try ?

[lvca]

After trying everything, the best choice today is Claude Code by far. It's the only one able to understand the code base and fix bugs, sometimes 100% autonomously. On designing new features is also REALLY good. Claude Code + Opus 4.5 is the best choice, nothing comes close. Tried Gemini CLI (I have a pro account), the class it made didn't even compile :-D

[ExtReMLapin]

Thanks for the feedback.
I used to go for the "shove the whole code base into the context", but as Gemini was the only able to do that and they nerfed it, i ended up testing claude code and the two PR it built aren't bad at all imo

[ExtReMLapin]

While claude code is working really well (Well it does the job at fixing issues, but I don't have the capacities of judging the java code), it's pretty much : 1 or 2 PR a day before exhausting the usage quota

Gemini has it's own coding agent i'll try to use

[include]

claude code (cli) user here too. my two cents are - keep a lean CLAUDE.md on your project and a leaner one on your ~/.claude. Make sure you have just the necessary MCPs enabled and eventually use a MCP proxy finder (Docker desktop has this feature). Also, under your project folder, keep also a short list of skills with specialists on the tech stack you are working on. regarding models Inam always mixing opus/sonnet 4.6. planning/execution