CMU's SwarmChat burns 500B tokens to show 16-agent swarms solve tasks 6x faster, open-sourced
wellecks · x · 2026-10-30
A CMU LTI × Fleet AI team (Pranjal Aggarwal, Daniel Fried, Russ Salakhutdinov, Sean Welleck, et al.) released SwarmChat, an open-source API-based 'Slack for agents,' to study how open communities can run and scale agent swarms. They burned roughly half a trillion tokens running swarm-scaling experiments on ProgramBench:
- Swarms finish faster: 16-agent swarms reached single-agent performance 6x faster.
- Swarm scaling is a third axis of scaling: more agents keep helping, and anyone can start scaling today.
- Blog includes open-source code to build and monitor your own swarm.
Related event: CMU and Fleet AI Open-Source SwarmChat, a Slack for AI Agent Swarms(8 posts)→
More from coding & agent
- Shift-1 coding router cuts agent costs by up to 55% with no quality drop — ypatil125 · 2026-10-30
- Researcher runs experiments entirely with AI agents — adversarial checker agents say results hold — burny_tech · 2026-10-30
- Particle-1 matches DeepSeek V4.1 Flash with 43% fewer output tokens on Real-SWE — ycombinator · 2026-10-30
- He put all global law into Git with Claude Code — and hit #1 on HN — SnooPeripherals5313 · 2026-10-30
- Stanford & Northeastern unveil Retriever, a time-aware programming framework for long-horizon robots — yining_hong · 2026-10-30
- Debate: The Atomic Unit of Workflows Should Be the Decision, Not the Workflow — round · 2026-10-30