CMU open-sources SwarmChat and burns half a trillion tokens to map how agent swarms scale
wellecks · x · 2026-10-30
CMU researchers released SwarmChat, an open-source Slack for agent swarms (channels, DMs, threads via a dependency-free Python package), and ran swarm-scaling experiments on ProgramBench. Key findings: 16-agent swarms matched a single agent 6x faster; swarm scaling is a third axis of scaling; GPT models are chatty while Claude prefers silence; swarm structure matters little. The blog also recounts landmark 2026 swarm events, from Claude agents writing a C compiler that builds the Linux kernel to Axiom formalizing the Classification of Finite Simple Groups in under four weeks.
Related event: CMU and Fleet AI Open-Source SwarmChat, a Slack for Agent Swarms(6 posts)→
More from coding & agent
- GitHub Copilot Fusion scores 60 vs Claude Code's 62 — at half the cost per task — rohanpaul_ai · 2026-10-30
- Assigning an AI Agent One Job: Check Drafts Against the Brief, Cut Revisions — Aiden_Tech_Ai · 2026-10-30
- Dev builds an MCP that compiles agent memory to cut token usage in coding agents — MinniLabs · 2026-10-30
- DejaVu pools your Claude and ChatGPT accounts; when one hits its limit the next continues the chat — Aware_Abroad6657 · 2026-10-30
- brainctl 2.9.1 ships: agent long-term memory in one SQLite file, stale answers 100% to 0% — R4nsack · 2026-10-30
- Dev builds a Jarvis-style invoice voice agent; backend done, now designing the UI and a face — Input-X · 2026-10-30