GitHub – HKUDS/ClawTeam: ClawTeam: Agent Swarm Intelligence (One Command → Full Automation)

ClawTeam: When One Command Turns Into a Whole AI Team Working for You

Imagine this. You type a single command into your terminal… and instead of one AI agent responding, an entire team wakes up.

That’s the idea behind ClawTeam, an open source project from HKUDS that introduces something called Agent Swarm Intelligence. You set the goal. The swarm handles the rest.

You can explore it here:
https://github.com/HKUDS/ClawTeam

Let’s slow down for a second.

Right now, most AI agents work alone. Even the powerful ones. If you’ve ever tried coordinating multiple tools, you know how messy it gets. You’re copying context back and forth, tracking tasks manually, trying to remember who did what. It feels less like automation and more like herding cats.

ClawTeam flips that model.

Instead of one isolated agent, it spawns specialized sub-agents, each with its own environment, identity, and focus. There’s a leader agent orchestrating everything. It delegates tasks, manages dependencies, monitors performance, and reallocates resources when something slows down. Think less “assistant” and more “autonomous research lab.”

What makes this interesting is the scope. We’re not just talking about writing code. The system is designed for:

• Large scale ML experimentation
• AI model training and optimization
• Automated market research
• Algorithmic trading
• Full stack development
• Even self evolving software

And yes, it runs across serious hardware setups, like multiple H100 GPUs, dynamically shifting resources based on performance. That’s not a toy project.

The deeper shift here is philosophical. For years, we’ve interacted with AI as a tool. One prompt in, one result out. ClawTeam treats AI more like an organization. Agents collaborate, share findings in real time, and iterate collectively. It builds on ideas like Andrej Karpathy’s autoresearch, pushing toward fully automated research cycles.

Picture this scenario. You tell Claude Code, “Build me a full stack todo app.” Instead of grinding through every layer sequentially, the system forms a team. One agent handles backend logic. Another designs the database schema. Another builds the frontend. Another tests. They coordinate through CLI commands, git worktrees, and even tmux sessions that you can monitor live.

You step in only when you want to.

We’re moving from single agents to digital teams. From isolated outputs to coordinated systems.

It’s still early. There will be rough edges, as with any ambitious open source project. But if this model matures, the way we approach research, development, and even business operations could look very different in a few years.

One command. A swarm executes.

And honestly… that changes how you think about automation.