No. I guess I'm not really good in expressing my topic. I don't need a solution. This is just curiosity and experimental. The situation right now is like this. Almost every agent (including agent zero) sends information about its environment to the llm with every call. Its basically telling the llm about the harness of the agent. Like you are operating on kali linux and you can use this toolset. This is static information (appr. 10k Tokens) sent every time. Then there is of course the dynamic information like memory and the users task. For the static part I would like to know if someone trained (fine tuned) a model so that it knows this information intrinsic. So the model itself knows about the harness. It doesn't need to be fed with the static information with every call. I know that this is technically possible. The question is about the effort and the cost. Is it even worth it for some use cases?

Cursor with their own trained model Composer. They're claiming it to be trained for the abilities of Cursor. And this two together seems to be performing very well.