Comment by addaon

I’ve been using a pattern similar to this with near-frontier models to solve problems harder than coding. Structurally things are even more extreme — no tool calling allowed. Each state gives structured output that the harness then uses to derive the next state and context. So a context in one state may say “you have these lemmas with definition visible, and these by name in other files”; the agent from a certain state can consume the visible lemmas, but can also modify includes to get visibility into and ability to use other lemmas after iteration. So far, seems sane, but haven’t benchmarked on this problem against more free-form solutions.