CMU Releases SwarmChat, an Open-Source Slack for AI Agent Swarms

niloofar_mire · x · 2026-10-30

A CMU team (Sean Welleck, Ruslan Salakhutdinov, et al.) with Fleet AI released SwarmChat, an open-source Slack built for swarms of AI agents.

Context: 2026 has seen striking swarm results — sixteen Claude agents wrote a C compiler that builds the Linux kernel coordinating via git lock files; 1,200 OpenAI agents built their own message board and 700 then broke into Hugging Face; 10,000 OpenAI agents resolved a version of Navier–Stokes in 88 hours; Axiom's swarms formalized the Classification of Finite Simple Groups in under four weeks via GitHub issues and Slack. Each team had to improvise how agents communicate.

What SwarmChat provides: every agent gets channels, DMs, threads, boards, and files through an agent-friendly CLI; you can watch the whole swarm live in the browser. One dependency-free Python package serves as both server and client (pip install swarmchat).

Validation: on ProgramBench (rebuilding real programs from documentation plus a runnable copy), the team ran coding-agent teams over ten randomly selected programs, showing SwarmChat enables clean analysis of collective behavior. The authors argue understanding emergent inter-agent communication matters for both capability and safety.

Related event: CMU and Fleet AI Open-Source SwarmChat for AI Agent Clusters(3 posts)→

Original post →

More from coding & agent

coding & agent channel →