# AutoSwarm: One Brain Connecting 12 AI Coding Agents

> Source: <https://byteiota.com/autoswarm-unified-ai-coding-agents/>
> Published: 2026-11-18 08:10:20+00:00

**Key takeaway:** AutoSwarm coordinates up to 12 coding agents through task units with four mandatory fields — assigned agent, expected output, parent objective and acceptance criteria — and scored 74.36%, 11.11 points above the Teamwork baseline, on long-horizon programming, planning and financial analysis tasks.

Codex is fast at boilerplate. Claude Code handles complex refactoring. Cursor knows the file tree. But none of them know what the others just did — and that context gap is quietly wrecking multi-agent workflows. HKU’s Data Intelligence Lab just open-sourced [AutoSwarm](https://github.com/HKUDS/AutoSwarm), a coordination layer that connects up to 12 coding agents into a single shared workspace. Lead researcher Chao Huang put it plainly: *“I kept switching between Codex, Claude Code, DeepSeek Harness, Muse, Workbuddy, Nanobot, and other agents — each good at different things, but none sharing the same context.”* AutoSwarm is his team’s answer.

## It’s Not Just “Run Agents in Parallel”

Most tools in this space focus on running the same task across agents for redundancy or cost arbitrage. AutoSwarm is doing something architecturally different. Its core unit is the **task unit**: a scoped contribution with four mandatory fields — assigned agent, expected output, parent objective, and acceptance criteria. Work doesn’t get handed off to the next agent until it passes the acceptance check. That matters because the classic failure mode in multi-agent systems is error amplification: garbage in, garbage multiplied.

The framework operates in two phases. Collaboration formation decomposes a task and assigns sub-tasks to agents by capability. Execution coordination then tracks a shared state — dependencies, parallel work, reviews, revisions, and blockers — across all participating agents simultaneously. One agent modifies a file; every other agent immediately has that updated context. No stale imports. No conflicting rewrites.

## Twelve Agents, One Setup

AutoSwarm supports Codex, Claude Code, OpenCode, Cursor, GitHub Copilot CLI, Meta Muse Code, CodeBuddy, Hermes, Nanobot, DeepSeek Harness, Grok Build, and Kimi Code. The install path is straightforward — it works with whichever agents you already have:

```
uv tool install --force .
autoswarm setup --host claude-code
```

AutoSwarm reads the installed `SKILL.md` and configures itself around your existing setup. If you prefer a GUI, [AutoSwarm Desk](https://alphasignal.ai/news/hku-s-autoswarm-lets-one-prompt-command-six-coding-agents-at-once) shipped on November 6 for Windows and macOS (arm64 and x64, native installers, integrated Python runtime, automated updates). One window for coordinating agents, reviewing approvals, and tracking progress.

## The Numbers Hold Up

The [arXiv paper](https://arxiv.org/abs/2610.33048) ran AutoSwarm across long-horizon programming tasks, interactive planning scenarios, and financial analysis. AutoSwarm scored 74.36%, beating Teamwork — the strongest prior multi-agent baseline — by 11.11 percentage points. It also outperforms AutoGen and MetaGPT. These aren’t trivial benchmarks: long-horizon programming tasks specifically test whether the framework can maintain coherence over extended multi-step work, which is exactly where fragmented agents break down.

The repository hit 2,600+ stars and was [trending on GitHub on November 16–17](https://dev.to/karleeov/five-open-source-developer-tools-to-try-in-november-2026-1e96), adding nearly 700 stars in a single day. AutoSwarm is MIT-licensed, Python 3.10+ required (3.12 recommended). If your workflow already includes two or more AI coding agents — and in 2026, most developers’ workflows do — AutoSwarm is worth a look before those agents spend another week working in ignorance of each other.
