Harbor Run. The white boat is the test bot, the red one is the AI rival. Recorded in the engine on an M4 Pro Mac, Oct 5 2026.
I don't write code. I direct AI agents, and they build each game inside the editor while I watch. Then a bot plays it through the real controls and writes up what happened. Nothing counts as done until that report says it works.
One email when the first builds go out. Nothing else.
Every object they add shows up in the scene list as it happens. You can stop them, change anything by hand, or undo any step. Nothing happens off-screen.
Castaway started as one sentence: wash up on an island, gather driftwood, light a signal fire before night. Agents built it in the editor. Then the bot played it ten times on ten different seeds. This is part of what came back.
Proof: Castaway build deb6469 · Oct 5, 18:23 · checked in 16 s
| Game Plan tests | 10 of 10 pass |
|---|---|
| Bot playthroughs | won 10 of 10 |
| Frame time | median 16.67 ms, p95 17.84 ms |
| GPU per frame | 11.50 ms |
155.5 s got Driftwood_3 180.6 s got Driftwood_2 180.6 s heading for Signal_Fire 215.8 s won: Signal fire lit with 87 s of daylight left.
When a run fails, the report names the object that caused it. Building Circuit, the bot found that trigger zones ignored rotation in every game, and that a wedged car could never reverse. Both were fixed in the engine, so every game got the fix.
Survival on an island. The engine had no fire, no sunset and no way to spread trees over a beach. This game needed all three, so they went into the engine, and every later game can use them.
10 of 10 bot wins · fastest 214 s · 59 objects, 44 of them palms
A race against an AI driver. The test bot had never driven a car, so it learned to: walk to the car, get in, follow the road. Its first attempt cut straight across the grass.
1,184 m lap · about 110 km/h · 5 of 5 races finished
Boat racing. The ocean existed, but nothing could float on it. Boats now ride the same waves the renderer draws, on the same clock, which is why the result doesn't change between machines.
Seed 1 finishes at 2:12.78 on the Mac and 2:12.78 on the Steam Deck
Next up: a horror game in endless yellow offices. The lights flicker, then go out, and all you have is the flashlight on your gun. Its shadows are real, cast from the beam. First playtest was on Oct 7, on a Steam Deck controller.
Not bot-tested yet · the creature that hunts in the dark is still being built
Give it a reference sheet and it builds the object, then compares its own renders to the reference from each side and keeps adjusting. Below: a car from one sheet. The red line is the reference outline.
The first frame on the Deck took a whole second, because skinning the character ran on the CPU. Moving it to the GPU took it to 60 fps, about 58 with shadows on.
| Mac (Apple ARM) | Steam Deck (AMD x86) | |
|---|---|---|
| Castaway, seed 1 | 218.8 s | 218.8 s |
| Harbor Run, seed 1 | 2:12.78 | 2:12.78 |
Windows player builds come from the same source and currently run under Proton on the Deck.
Arcwright is built by one person who doesn't code, directing AI agents. Nothing merges until the tests pass. 385 commits as of Oct 7.
First commit at 5:27 am. Before 8 am an agent had built a stealth level live in the editor: guards with vision cones and flashlights, a yard, an escape.
Castaway, Circuit and Harbor Run built. The bot wins all three. Fire, sunsets, cars and boats added to the engine along the way.
Ported to Steam Deck. The same three games run at 60 fps, and finish times match the Mac to the hundredth of a second.
479 tests. First playtest with a Deck controller in hand. Named it Arcwright.