Using a virtual filesystem to minify source code and let the LLM evolve in the minified space gives surprisingly good results and reduce tokens by 20 to 40% (depends on the language, python for instance)
I've been building an AI coding agent that using the exact same prompt than claude code, but uses a virtual filesystem to minify source code + the concept of stem agents (general agents that specializes during the conversation for maximum cache hit). The results on my modest benchmark is 50% of claude code cost and 40% of the time.
https://github.com/kirby88/vix-releases