it would be interesting to see a Jev-like approach integrated with a normal LLM where the jev classifier repeatedly chooses whether:
- to speak (query the LLM to fill up a user-visible context window),
- to think (fill up a private context window), or
- to act (make a tool call)
