robocoop-3 is my new assistant! Its awesome!
This put the Agentic loop inside a generator function, it works quite differently to robocoop-2. It doesn’t know Observable 1.0 syntax, instead, it directly reads and writes the runtime. It is amazing investigating and finding bugs, if you prod it a bit. It can hold a multi-turn dialogue, and can use Anthropic, OpenAI or Ollama (local models!).
I still use robocoop-2 to write notebooks, because that can read and write Observable 1.0 syntax. robocoop3 is much snappier to use though, because it properly does streaming and parallel tool calls. robocoop-3 was made by porting Opencode, and iterated on with robocoop-2.
I see robocoop-3 as a building block for other workflows, e.g. Wiggam Loop, AI-as-judge etc. More of a thing you will instanciate a few times in a few different places and customize to the situation. I have not tuned its prompt very much yet, that will be next, but as I am finding it useful fixing a few long running bugs, I thought I would share it now, as I am not gonna work on it for a little while, its good enough for now, especially when combined with robocoop-2.
robocoop-3 was powerful enough to write its own integration with moltbook! moltbook - the front page of the agent internet
Oh I also wired up a demo provider which is a bit slower than direct API calls because it goes through a proxy, but you can try it out now for FREE. BUT, you are limited to openAPI GPT 5-mini (cheap!). Put your own key to get the faster and most intelligent gpt 5.2 experience.