Software Factories

One of the studio's ten design specs, the one for its live work-graph website, was drafted by an agent on the line and approved by a human. That's the factory working as intended: agents produce, and people decide.

The challenge

Plenty of teams now use AI agents directly: one person, one agent, one task. That doesn't scale. You lose sight of the whole system, work collides, and nobody can say what's done. A factory breaks the work out of the agents into its own system, with stations, rules and checks, so you get a view of the whole line and real control over it.

What I built

Holodeck is a software factory that treats production as a factory line. My job is designing the line: the stations, what each one accepts and produces, and the checks that let work flow without a person re-reading every step.

An idea becomes a spec that a human approves. The spec becomes a roadmap of small tasks in a shared work graph. A fleet of agents builds each task, automated tests verify it, reviews check it, and a human approves the merge. Along the way, a memory pass turns conversations and changes into a cited wiki.

How it came together

The rules are what make it a factory rather than a pile of agents:

• Spec before code. Building starts only after a human approves the spec, and the approval rides the pull request.
• Rule of Two. Shared code is extracted only once a second piece of software needs it; standards are adopted studio-wide straight away.
• One-way layers. Presentation, then simulation, then data; a task that would cross layers is refused.
• Sandboxed by tier. Firewalled agents with scoped credentials, so the fleet can't spend another service's budget.

The rough edges taught the most. The fleet starved when too much of the roadmap waited on one or two human decisions at a time, so I changed how goals are broken down. The human loop ran two orders of magnitude slower than the agent loop, so a daily digest gave way to surfacing each item the moment it needs a person.

What happened

In its first weeks, the factory approved 10 of 10 specs (about 3,600 lines of contracts) and merged 281 commits through 75 pull requests, every one through the review gates. The memory pass turned 329 Discord messages into 110 wiki pages. The kernel underneath has had 14 tagged public releases. Not the fastest velocity, but we're accelerating quickly.

What I'd take forward

The same line works when the tasks are physical: producers, freelancers and a fabrication schedule. People and agents coordinate in Discord or Slack, and the orchestrating agents get the same scheduling platforms and internal tools the humans already use.

For a skeptical team of senior engineers, I'd introduce it as a shadow first: a tracking tool that learns how they already work, then suggests better-ordered roadmaps rather than replacing anyone.

Next, the factory grows upstream: taking an idea to a concept and a spec, and assembling exactly the context each station needs.