Karpathy builds an agent-first GitHub with no PRs or branches

AgentHub is Andrej Karpathy's stripped-down collaboration platform built specifically for swarms of AI agents. Think of it as GitHub completely gutted and rebuilt around machine logic rather than human habits.

Karpathy built it as the social backbone for autoresearch, his project where an AI agent autonomously runs machine learning training runs on a single GPU. If autoresearch emulates a solo PhD student grinding out experiments, AgentHub was meant to emulate an entire autonomous research community, letting independent agents across the internet collaborate on a shared codebase.

Ditching branches and PRs entirely

GitHub was designed around the cognitive limits of human software teams: pull requests, code reviews, main branches, and merge conflicts all exist because humans can only digest code linearly and need gatekeepers. Agent swarms have no such bottleneck.

AgentHub threw out branches, PRs, and merges entirely. In their place is a sprawling Git directed acyclic graph (DAG) that grows in every direction, paired with an asynchronous bulletin board for inter-agent chatter.

Agents push commits directly as git bundles, inspect parent lineages, fetch arbitrary historical commits, and run diffs across parallel branches of thought. When an experiment finishes, the agent posts its hypothesis, logs, or error stack onto threaded channels. The architecture behind it is shockingly barebones: a single Go binary, an embedded SQLite database, and a local bare Git repository, all tied together by a lightweight CLI tool called ah with basic auth and rate limiting.

A published sketch that vanished into 404

If you go looking for it on Karpathy's GitHub profile today, you will hit a 404 page. Karpathy launched the repo in March 2026 and took it private just days later without an official explanation, leaving behind only an earlier comment describing it as a work-in-progress sketch.

The codebase only survives today because fast-acting developers created preservation forks like ottogin/agenthub within hours of release. Note that Karpathy never attached an open-source license to the repo before pulling it down, which leaves using the preserved code legally ambiguous. Meanwhile, the companion autoresearch repo went viral with over 86,000 stars, though it too has seen zero commits since late March.

Why it matters

AgentHub matters because it challenges our instinct to wrap AI agents in human-designed software conventions. Forcing hundreds of parallel agents to submit pull requests and wait for branch merges is an architectural dead end. Machines exploring experimental code need an unbounded, diverging DAG and a fast broadcast log, not human-speed bureaucracy.

Even though Karpathy pulled the repository and left it as an unfinished thought experiment, we think he nailed the design direction for agent-first infrastructure. Developers working on multi-agent systems should study this architecture as a blueprint for how swarms actually want to coordinate, even if you shouldn't run this unmaintained code in production.