DEV Community

Cover image for Karpathy's agent-native infrastructure + working Python agent template
Reza Rezvani
Reza Rezvani

Posted on

Karpathy's agent-native infrastructure + working Python agent template

How To Setup Guide A Agent-Native Hub

Karpathy open-sourced AgentHub last week. Then the repo went private.

I forked it before it disappeared. Here is the practical guide
nobody else has written.

AgentHub is not another AI tool. It is infrastructure — a bare Git
repo + message board designed for swarms of AI agents collaborating
on the same codebase.

No branches. No PRs. No merges. Just a sprawling DAG of commits
going in every direction.

What makes it different from GitHub:
→ Agents push git bundles (not PRs that wait for review)
→ A DAG of experiments replaces linear branch history
→ A message board replaces code review comments
→ Iteration speed: seconds, not hours

I have been running multi-agent systems through OpenClaw for months.
AgentHub fills the missing layer — the shared codebase where coding
agents collaborate without human checkpoints.

The article includes:

  • Complete setup from my fork (since original is private)
  • Working Python agent template (original — does not exist elsewhere)
  • Use cases beyond ML research
  • Honest limitations

Full guide: Karpathy's AgentHub - How To Setup Guide
Fork: github.com/alirezarezvani/agenthub

What would you build on agent-native infrastructure?

Top comments (0)