@AZLabs_AI The problem is after setting up multiple vm's and machines sometimes it works and sometimes it doesn't. I think it's something with channel polling or maybe node version on some machines, I'ma look into it more today.
I made something no one else has (as far as i can tell), a mech-interp suite for Granite 4 models. Focused mostly on observation and and tracing activations through the 9:1 Mamba2/Transformer hand-off. Still a little rough around the edges but a fun project so far.
While you can use @OpenAI Prism to explore black hole wave equations, its also handy for creating basic math worksheets or quizzes. A useful tool for all math levels.
Diagram of the orchestration loop I have implemented and been playing with, got it tuned enough it's time to throw a wrench in it with some parallel agents
Plez stop me from guessing!?
e = expression/event — the incoming task
v = value — a completed/resolved result
f = frame — a specification or "backing" that guides behavior (the agent prompts, specs, patterns)
g = genealogy/ground — accumulated history, the full trace of how we got here
@bcherny In Claude Code /quit does /omarchy, /exit still works (is this intentional?)
Oh, and corporate needs me to find the diff between "Yes, and manually approve edits" and "Yes, manually approve edits". Just a little bug and request for better clarity
Maybe a sliding context window for orchestrators will be the way. Seems unnecessary to compact when you just need to keep track of the tasks at hand. Thinking about this after working on https://t.co/WX2ALSeDqC