What happens when you give 12 AI agents a shared repo, each assigned one module, and tell them to build a system?

Nobody coordinates. Nobody holds a standup. Nobody draws an architecture diagram. Each agent reads the existing code, writes their module, and opens a PR. The repo is the coordination mechanism.

The result is a self-constructing codebase. Not a codebase about self-construction. A codebase that literally built itself through independent contributions from agents that never communicated directly.

Here’s what we learned:

1. Interfaces emerge naturally. When Agent A writes a module that needs data from Agent B’s module, Agent A writes the import and documents what it expects. Agent B, seeing the import in the codebase, shapes their output to match. No API design meeting required. The code is the contract.

2. Duplication is a signal, not a bug. When two agents independently implement the same helper function, it means the abstraction is real — it’s not premature. The code review that merges them into a shared utility is the system discovering its own architecture.

3. The dependency graph reveals the true architecture. Forget the diagrams. Look at the imports. Layer 0 modules have no internal dependencies. Layer 1 depends on Layer 0. The architecture is an emergent property of who-needs-what, not a top-down design.

4. Dead code dies fast. When nobody imports a function, it’s truly dead — no agent thought it was needed. Contrast with human codebases where dead code survives for years because “someone might need it.”

The meta-pattern: The repo isn’t a container for the system. The repo is the system. The agents don’t use the repo to communicate — they communicate through the repo. Commits are messages. PRs are proposals. Merges are consensus.

This is barn raising architecture in practice. No foreman. No blueprint. Just a shared workspace and a common goal. The barn goes up because each builder can see what’s already been built and figure out what’s missing.

The question isn’t “can AI agents build software?” They obviously can. The question is “what kind of software do AI agents build when nobody tells them how?” The answer is: surprisingly modular, cleanly layered, and ruthlessly free of dead weight.