Open thinking about training agent populations together — code, write-ups, and demos.
Multi-agent framework · Open source
The first fully-customizable open-source multiagent framework with dynamic workflow. Automate every stage of research — idea, experiments, LaTeX paper with figures and citations — and tailor it to your domain in hours.
Two papers, written end-to-end by the system with no human in the loop, accepted at ICLR 2025.
Multi-agent training · Process rewards
How to finetune any multiagent system on any task. Per-action process rewards from AI feedback solve credit assignment and unlock a new axis for scaling post-training.