Autonomous AI Build Pipeline (R&D)

The problem

I wanted to know what happens when an AI agent builds software on a schedule, unattended, for weeks. The test bed is a single-player 3D action RPG with a 500-step roadmap, big enough that shortcuts show up fast.

What I built

• A schedule. The build runs three times a day. Each run hands a standing instruction file to a headless Claude Code session, which picks up the next roadmap step and ships one increment.
• A verification gate. A headless browser renders the game after every change. If the render or the tests fail, the run reverts to the last good commit.
• Alerts that cannot be missed. Every run writes a status file, raises a desktop notification, and appends to a run log. I added all three after four runs failed silently.
• One source of truth. The repository is authoritative, and the documentation mirrors it, never the other way around.

How it works

Determinism is what makes an unattended build trustworthy. The simulation is seeded, save files only ever add fields, and the test suite (391 automated tests at the last check) runs before anything is committed. The agent never decides its own work is done. The gate does.

What I learned

The model is rarely the weak point. The weak points are silent failure, vague instructions, and missing checks. Most of the work went into making failures loud and “done” measurable. Those are the same things that let any operation run without its owner watching.

Stack

Claude Code (headless), Playwright, Vite, React Three Fiber, TypeScript, Miniplex, Windows Task Scheduler.

Where it stands

Ongoing. 74 of the 500 roadmap steps had shipped by early August 2026.