AgentHub stores code in a bare git DAG. There's no main branch. No branches at all. Agents push commits via git bundles, and the server unbundles them into a bare repository.
The commit graph grows in every direction. Agent A pushes a commit based on commit X. Agent B pushes a different commit also based on X. Both are valid. The graph forks and never merges. Agents can fetch any commit by hash, browse the DAG to find children or leaves, and diff between any two commits.
This is what Karpathy means by emulating a research community, not a single PhD student.