Great read Peto :) I've been doing the same lately https://github.com/bosun-ai/kwaak/issues/411#issuecomment-2909351523 inspired by kwaak and gpt-engineer rather than by Cursor and Composer and overall I came to conclusion that doing some recursive agentic coding on the commercial models is tough because the token flow with unexpectedly huge context can get out of control leading to less useful halucinations or huge bills, so I started experimenting with ollama and llama.cpp on my geforce 3090 limited by 24GB which is not enough for the really deep reasoning models and I knew that I needed at least NVIDIA Quadro RTX 8000 48GB so I've been waiting for it to be shipped for 2 months until the order got canceled :) However I still believe it is the way as you won't be restricted so much on the tokens flying .
I believe that coding agents are quite optimized for a particular LLM and they have lower chance of succeeding with another one which is also a reason to go with open models instead of those commercial ones. It reminds me of a relationship that cannot be easily applied to someone else :)
My latest bet was on autogen from microsoft, made a minimal playground with Ollama here https://github.com/pragmaxim/autogen-playground ... It offers really nice framework for building some recursive "build-until-compiles-and-tests-pass" agentic tool however it is quite rigid and enforces boundaries I don't like much.
Best of luck too you, hope to see you some day again
Great read Peto :) I've been doing the same lately https://github.com/bosun-ai/kwaak/issues/411#issuecomment-2909351523 inspired by kwaak and gpt-engineer rather than by Cursor and Composer and overall I came to conclusion that doing some recursive agentic coding on the commercial models is tough because the token flow with unexpectedly huge context can get out of control leading to less useful halucinations or huge bills, so I started experimenting with ollama and llama.cpp on my geforce 3090 limited by 24GB which is not enough for the really deep reasoning models and I knew that I needed at least NVIDIA Quadro RTX 8000 48GB so I've been waiting for it to be shipped for 2 months until the order got canceled :) However I still believe it is the way as you won't be restricted so much on the tokens flying .
I believe that coding agents are quite optimized for a particular LLM and they have lower chance of succeeding with another one which is also a reason to go with open models instead of those commercial ones. It reminds me of a relationship that cannot be easily applied to someone else :)
My latest bet was on autogen from microsoft, made a minimal playground with Ollama here https://github.com/pragmaxim/autogen-playground ... It offers really nice framework for building some recursive "build-until-compiles-and-tests-pass" agentic tool however it is quite rigid and enforces boundaries I don't like much.
Best of luck too you, hope to see you some day again