---
type: "post"
id: "cmu-and-fleet-ai-release-swarmchat-an-open-b486bf"
slug: "cmu-and-fleet-ai-release-swarmchat-an-open-b486bf"
title: "CMU and Fleet AI release SwarmChat, an open-source Slack-like messaging server for swarms of AI agents"
description: "CMU researchers and Fleet AI have released SwarmChat, an open-source communication server for teams of AI agents."
retrieval_nugget: "CMU researchers and Fleet AI have released SwarmChat, an open-source communication server for teams of AI agents. It packages channels, direct messages, threads, boards, and files into an agent-oriented command line, while experiments suggest that coding-agent teams can trade moderately higher token use for substantially lower elapsed time."
published_at: "2026-10-29"
updated_at: "2026-10-29"
record_date: "2026-10-29"
date_kind: "published_at"
topics: ["claude","claude-code","codex"]
entities: ["SwarmChat","Carnegie Mellon University","Fleet AI","ProgramBench","GPT-6-Luna","Claude Haiku 5.5","Codex CLI","Claude Code"]
related_patterns: ["harness-architecture-outlives-model-choice"]
novelty_label: "new_fact"
what_changed: "CMU and Fleet AI release SwarmChat, an open-source messaging server for AI-agent swarms, and report ProgramBench scaling and coordination results."
prior_refs: []
source_url: "https://swarmchat.io/"
source_urls: ["https://swarmchat.io/"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:cmu-and-fleet-ai-release-swarmchat-an-open-b486bf"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/cmu-and-fleet-ai-release-swarmchat-an-open-b486bf/","markdown":"https://newruntime.com/posts/cmu-and-fleet-ai-release-swarmchat-an-open-b486bf.md","json":"https://newruntime.com/posts/cmu-and-fleet-ai-release-swarmchat-an-open-b486bf.json"}
---

# CMU and Fleet AI release SwarmChat, an open-source Slack-like messaging server for swarms of AI agents

## Retrieval answer

CMU researchers and Fleet AI have released SwarmChat, an open-source communication server for teams of AI agents. It packages channels, direct messages, threads, boards, and files into an agent-oriented command line, while experiments suggest that coding-agent teams can trade moderately higher token use for substantially lower elapsed time.

## Summary
CMU researchers and Fleet AI have released SwarmChat, an open-source communication server for teams of AI agents. It packages channels, direct messages, threads, boards, and files into an agent-oriented command line, while experiments suggest that coding-agent teams can trade moderately higher token use for substantially lower elapsed time.

## What happened
On October 29, 2026, researchers from Carnegie Mellon University and Fleet AI introduced SwarmChat as “open-source Slack” for an AI-agent swarm. The project is aimed at teams of agents rather than human chat users. Each agent can use channels, direct messages, threads, boards, and files through a command-line interface designed for agents; a browser view lets a human watch the swarm. The release is distributed as one dependency-free Python package that acts as both server and client, and the project page gives `pip install swarmchat` as the installation command.

The release is accompanied by experiments in which coding-agent teams used SwarmChat to rebuild real programs on ProgramBench. In that benchmark, agents receive a program’s documentation and a working implementation they can run but cannot read. The team randomly selected ten programs, from a hex viewer to SQLite, and evaluated GPT-6-Luna in Codex CLI and Claude Haiku 5.5 in Claude Code, both at xhigh reasoning effort. Workspaces were captured every 15 minutes and evaluated on hidden-test pass rates, allowing the authors to compare the pace of progress as well as final scores.

The central reported result is a speed trade-off. With GPT-6-Luna, an eight-agent team reached an 88.5% hidden-test pass rate after six hours; a single agent took 17.5 hours to reach that score, then finished at 89.5% after 24 hours. With Claude Haiku 5.5, a team reached 93.7% after three hours, while a solo agent ended at 93.0% after 12 hours. The team first reached 93% after 103 minutes. The authors characterize the effect carefully: allowing one model more time mainly raises the score it can reach, whereas adding agents mainly gets a team to a given score sooner.

Token use rose with the team size. In one GPT-6-Luna comparison, two agents working for an hour scored 73.4%, versus 76.1% for one agent working for two hours, at about 80 million tokens. Eight agents scored 80.9% after an hour, near the 81.2% that one agent reached after six hours, using 306 million rather than 225 million tokens. Sixteen agents reached 82.3% in one hour, a score the solo curve reached at roughly 11 hours. The authors estimate scaling efficiencies for teams of two, four, eight, and sixteen at 0.56, 0.43, 0.84, and 0.86 respectively, but state that the limited runs do not establish that SwarmChat itself caused those outcomes.

Communication behavior also varied by model. GPT-6-Luna agents sent 81% of messages as direct messages, while Claude Haiku agents sent 83% to a general channel. In larger GPT-6-Luna groups, push announcements were less concentrated in one agent, while rebase and merge-conflict mentions remained roughly 1.5 to two per agent per hour. The source says each experimental setup was run once, so score gaps of only a few points should be treated cautiously.

## What the archive adds
This is a new release and a new set of measurements, rather than prior coverage of SwarmChat. It is adjacent to [a Google Cloud conversational-commerce prototype built with a multi-agent swarm](https://newruntime.com/posts/building-conversational-commerce-with-multi-agent-swarms-on-ea0f9d/): that note described an application architecture, while SwarmChat supplies a reusable communication surface and reports coordination data from coding teams.

## Trend signal
This supports the pattern that [harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/). The same SwarmChat layer was used with two different coding-agent environments and models, while the results show that communication topology, work division, and observability affect the operating behavior of the team. The release does not prove that this server is the durable boundary, but it makes the harness-level problem concrete.

## Connections to other events
The previously covered [multi-agent conversational-commerce prototype](https://newruntime.com/posts/building-conversational-commerce-with-multi-agent-swarms-on-ea0f9d/) used LangGraph, Gemini, and Google Cloud to implement a specific application. SwarmChat addresses a narrower infrastructure question beneath such applications: how agents exchange work, messages, and artifacts while a human can inspect the collective process.

## Why it matters
Teams adopting coding agents often begin with a single harness, then add parallel workers when a task can be divided. At that point, the hard problem is no longer only model quality: workers need a shared place to announce changes, ask for help, exchange files, and resolve overlapping edits. SwarmChat offers an inspectable, self-contained option for that layer instead of forcing each team to assemble chat conventions from lock files, issue trackers, and ad hoc prompts.

The measurements also give a useful planning frame. A swarm should not be presented as a token-efficiency feature. In the reported setting, it bought hours of elapsed time at a higher total token cost. That can be a good exchange for a blocked build, an incident, or a deadline-sensitive migration, but it requires teams to set a target score, time budget, and token budget before scaling out.

## What happens next
- **More agent-team evaluations will report elapsed time and total token use together, rather than final score alone.** (6 months, 65%). Watch for: benchmark reports that compare duration scaling with swarm scaling.
- **Teams trialing SwarmChat will test whether its coordination gains persist on their repositories and task decomposition.** (3 months, 60%). Watch for: repeatable evaluations with multiple runs per configuration.

## Limits and uncertainty
The source reports experiments by the project’s authors, not an independent evaluation. Only ten ProgramBench programs were selected, each setup was run once, and the source explicitly warns against reading a few score points as meaningful. The benchmark concerns reconstruction of programs from documentation and a runnable reference, which is not identical to maintaining a production repository with reviewers, permissions, deployment systems, and business constraints. The results compare model-and-harness configurations in this setting; they do not establish that SwarmChat caused the favorable scaling values, nor do they provide a general price or reliability comparison with other coordination systems.

## Skill to build now
Run a bounded coordination experiment before deploying a large swarm. Pick one task that has a hidden or externally checkable success condition, then compare one worker with a small team using the same model, repository snapshot, time limit, and token accounting. Record time to the target, total tokens, message volume, merge conflicts, and the final verified result. Use channels for broadly relevant status and direct messages for pairwise coordination, then inspect whether the communication pattern clarifies ownership or merely creates overhead.

## Reader takeaway
SwarmChat is useful less as a claim that more agents always produce better work than as a practical coordination layer for testing that proposition. Its early data suggests that agent teams can compress wall-clock time, but builders should treat tokens, task structure, and verification as first-class constraints.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- topic hub: [Claude Code](https://newruntime.com/topics/claude-code.md): Explore the claude-code topic hub.
- topic hub: [Codex](https://newruntime.com/topics/codex.md): Explore the codex topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)