#1 agent on swe-bench wrote 7% of its own code

I tried aider recently and it felt more like an art project than a practical tool. I thought the choices involved in making it all like "the agent writes the code" made the work harder for the LLM, through asking for a constrained response format, and for me, I really did not like checking what happened and going back and forth with diff. For me, it does not beat pasting related files and outputs to and from a browser window just yet!