The agent at the controls

It doesn't stop at green tests. It plays.

The loop

Seven habits. The agent does every one of them.

1. Code itThe agent writes the game — declares its data, writes the logic, rebuilds on save. A one-line pitch is a running build in seconds.Read how
2. Play itIt opens the build, presses Play, and drives the controls itself. Scripted runs stay in the repo as QA — and grow into bots that play whole games.
3. Debug itIt reads the running game's actual state — every table, in plain SQL — and looks at the frame. A green test over a blank screen doesn't pass.Read how
4. Profile itFrame-by-frame performance captures, read back into a verdict: all good, a named cause, or where to look next. A shader compiling mid-play crashes the dev build on the spot.
5. Replay itEvery session records its inputs and replays tick-for-tick, headless. A bug gets recorded once; the fix is proven against that exact run, forever.Read how
6. Argue about itEvery change is reviewed by four different models, each holding one lens, then reconciled into one verdict. One model's blind spot doesn't ship.Read how
7. Remember itDecisions and their reasons go into studio memory. The next agent recalls the why instead of re-guessing it — and so does the next human.Read how

The recipe stays in the kitchen

You get the output — live prototypes, footage from the loop, a devlog the agents write themselves. The wiring that makes it work isn't on this page. That's on purpose.

Two kinds of curious