This could be the most important OpenAI paper of 2025.

Chain-of-Agents (CoA) proposes a radical shift: instead of orchestrating fragile swarms of LLMs, OpenAI shows how to distill entire multi-agent workflows into a single foundation model that can dynamically assume roles, coordinate and self-optimise.

The core innovation is multi-agent distillation combined with Agentic reinforcement learning:

• Capture trajectories of multiple specialised agents working together.

• Compress them into one model.

• Achieve state of the art results on web agents, code execution and multi step reasoning.

I turn cutting edge AI papers into insights you can actually use. Subscribe before the next drop.

Full paper: