Bitter lesson comes for multi-agent systems: RL-trained swarms beat hand-designed orchestration

emax · x · 2026-11-17

A debate on multi-agent system design. @barrnanas argues the bitter lesson is coming for multi-agent systems: instead of hand-designing agent roles, hierarchies, and communication protocols, let agents learn cooperation via RL — recent work shows swarms can self-organize, developing emergent coordination and division of labor that outperforms best-of-N. Reposter @AddiHD02 goes further: emergent bottom-up coordination (stigmergy) is the agent-native form of orchestration, and we should question whether models need post-training to collaborate at all, versus simply giving them the right environment at the harness level.

Original post →

More from coding & agent

coding & agent channel →