This project is now considered obsolete.
At the time of writing this in May 2025, there are much easier solutions for building powerful workflows using agentic frameworks like smolagents or tinyAgent by having a thinking model like DeepSeek R1 become a "boss" agent to orchestrate other agents powered by more efficient specialized models.
This approach may not have the same token efficiency as static chains of thought, but it's hundreds of times faster to build up and iterate on (especially with the advent of powerful coding LLMs and IDEs like Cursor and Zed), more portable, and much more powerful.
As such, I am archiving this repo. The codebase might still be useful to someone as a stepping stone to build something entirely different, so feel free to fork it and build what you like.
You can find the old documentation for the project in its own repo here.
