A swarm that builds games

I describe a game in a sentence. A while later it is installable on my phone, and in between no one touched a keyboard. That loop — intent to playable app with my attention spent only at the edges — is the thing I have actually been building. The games are how I test it.

The pipeline is plain once you see it. A crew board takes the request and decomposes it into dependency-ordered tasks. A planner hands each task to specialist agents — a coder, a critic, a librarian that fetches context — running in parallel under concurrency that watches the GPU and falls back to CPU before it ever shrinks a model to go faster. A verifier gates each task on the project's real tests. The finished build is signed, uploaded to a private store, and shows up on my home screen.

What separates a swarm that ships from one that loops forever is not intelligence, it is discipline. Tight task contracts, so each agent has one output the next can consume. A verifier wired to the artifact, not the agent's narration. Escalation to a heavier model after a couple of honest failures instead of infinite polite retrying. And a human reviewing diffs, not transcripts — the patch is the only artifact that tells the truth.

It is not magic, and pretending otherwise is how you get burned. It will pick the wrong stack if you let it plan loosely. It will, left ungated, mark things done that were never built — that is a whole separate story. The swarm is an amplifier: it makes a good process faster and a sloppy one dangerous. The gates are what keep the amplification pointed at shipped code.

The payoff shows up in what becomes worth attempting. When building a small app costs a sentence and a wait, and distribution costs one more command, the threshold for 'is this worth trying' collapses. A chess game with a search engine that thinks off the UI thread, a tower defense, a card room — each one started as a description and ended as an icon. The swarm did the middle.