Jack Dorsey (Twitter co-founder) just released a completely free GitHub repository with 26.2k stars. It's essentially a ready-to-use AI Agent operating system built for real business use.

This isn't a toy. Boris Cherny — lead of Claude Code — detailed how the system works in the documentation: your phone serves as the main interface, it runs 5 to 10 concurrent sessions at once, and each session can spawn hundreds or even thousands of sub-agents, with dozens of background loops running continuously. The standout stat: **Trigger 1,000 agents to run overnight straight from your phone**.

It sounds like you'd need a whole data center and an infrastructure team to pull this off. But you don't.

**How it actually works:**

1. **Clone the repository**

2. **Self-host the server**: Channels, search, Git, automation all run on your own machine

3. **Add your agent to a channel**, just like you'd invite a new colleague to a group chat. Lock in its permissions, and your team can direct its work in real time

That's really all there is to it.


CyrilXBT broke the core architecture of this system down into four key parts in the supporting tutorial. None of them are particularly complex on their own, but when combined, they deliver game-changing results.

**First, mobile triggering.** This isn't just a mobile dashboard where you check progress via push notifications. It lets you actually send commands, check statuses, and adjust direction straight from your phone. Claude's Dispatch interface does exactly this — it enables a continuous workflow initiated from your phone, where you can actually take action, not just view read-only updates.

**Second, isolated parallel execution.** Every agent (or group of sub-agents) runs in its own context and workspace, with no interference between tasks. Without this, running dozens of agents at once would be a mess, not scalable. Git worktree handles fully independent tasks, while sub-agent isolation manages complex collaborative workflows.

**Third, verification doesn't depend on self-assessment.** This is the most critical component. The core principle comes from Anthropic's harness engineering experience: **Never let an agent grade its own work.** A model reviewing its own output within the same context is inherently biased toward over-optimism. When a task runs overnight unattended, waking up to a confident but completely wrong "job done" message is far worse than no result at all.

The solution: The Builder does the work, and an independent Judge handles verification. The Judge gets access to information the Builder doesn't: test suite outputs, the original requirements document, and real execution results. It doesn't just re-read the output and rephrase it.

**Fourth, hard stop conditions and cost caps.** For any unattended task, there are three non-negotiable rules: a maximum time or number of iterations; a maximum cost (calculate for the worst case, and actually stop if it's exceeded); and clear scope boundaries (no production deployments, data deletion, real-world spending, or external messages sent on your behalf without your explicit approval).

One commenter put it perfectly: "This only works if your team actually pays attention to the agent. My agent sat in the channel doing stupid things for two days because nobody checked it."

Others have crunched the numbers: Running 1,000 agents overnight with Opus 5? A simple question like "what time is it" can cost $10 in tokens. That's not an exaggeration, that's the real cost structure — but if you route tasks to models based on complexity, using cheaper models for simple work, you can cut most of that cost.

There are also accusations of plagiarism: Some claim the work was originally done by others, repackaged as infrastructure without attribution. These kinds of disputes are nothing new in open source, and how you see it depends on how you interpret the idea of "standing on the shoulders of giants."

MR Wayne made an insightful observation: "26.2k stars and it's free — the interesting thing isn't the repo itself, it's that Dorsey's version of an 'agent operating system' is just a permissioned chat channel."

He's right. You don't need a complicated management dashboard. You just need a channel, add the agent to it, lock in permissions, and let it work.

Some are asking if this means the era of "adding agents to your team to get work done" has arrived. It looks that way, but don't forget what Boris Cherny himself said: **"We're only 1% done."**

This project is still evolving rapidly. But these four core designs — mobile triggering, isolated parallel execution, independent verification, hard stop limits — are the non-negotiable foundations for running large-scale unattended agents today. Miss any one, and it either won't scale, or it will break when it does.

The repository is free, it already has 26.2k stars. The tutorial is also free, with the four-part architecture broken down clearly. What you do next is entirely up to you.