EvoRouter: Cheaper Multi-Agent Routing That Learns
Running a multi-agent system means choosing models, a collaboration mode, how many agents, how many rounds and which skills, and the combinations explode. A reward for final success and cost alone cannot say which of those choices helped and which wasted compute.
This episode walks through EvoRouter (arXiv 2610.35690). An expert model periodically reads performance–cost feedback from many routing trajectories and distills conditional routing insights: where to scale, how much to scale, and which configurations work together. A privileged teacher combines those insights with execution feedback…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.