I Built an Office Full of AI Developers. They Started Opening Pull Requests.

Description and use

A developer built My Little Office, where AI agents autonomously modify a real repository but only open pull requests after passing tests, shifting the developer’s role from coding to task definition and oversight.

Problem

Developers need to manage AI agents that autonomously modify code, ensuring correctness, safety, and coordination.

Outcome

Agents can autonomously modify code, but verification gates and isolated workspaces prevent destructive changes, ensuring only passing tests lead to PRs.

Workflow

• Define a task for an agent (e.g., "Fix checkout crash").
• Agent clones repository into isolated workspace.
• Agent understands task, inspects code, makes changes, runs project checks.
• If checks fail, agent attempts repair; if still failing, task is marked failed.
• If checks pass, agent commits changes, pushes branch, and opens pull request.
• Human reviews PR and verifies final outcome.
• Deploy to production if approved.

Evidence

"If the project's verification commands fail, the task doesn't get delivered."

"An agent can work inside its workspace. It cannot casually wander around your machine."

"Tests become the fucking contract."