Harborlight: Directing an AI Studio to Build a GameHarborlight: Điều hành một studio AI làm game

Harborlight is a toy, not a game with goals. You pick a colour, click the water, and a house appears. Stack houses and they grow roofs, chimneys, balconies and window boxes. Lay a foundation between two quays and it becomes a bridge. Birds land on the rooftops. It is inspired by Townscaper, Oskar Stålberg's town-building toy, and was built by studying how that game looks and feels.

The project began with one ambitious request: build a Townscaper-like game in the browser at "AAA quality" (the polish of a big-studio game). Split the work across many AI agents. Give every piece a separate, harsh critic. Keep going until the critics, shown both games side by side without labels, could not call the original the better one.

That prompt set the bar. It did not finish the job. Getting there took 32 more messages over two and a half days.

The lead agent worked like a studio director. It handed out work to 46 specialists, each in its own session:

Builders owned one part each: the rules of the world, the 3D renderer, the water, the sound, the interface, the way clicks turn into buildings. Critics looked at screenshots and code and listed what fell short. Blind juries saw two unlabelled pictures, one from each game, and said which looked better and why.

An advisor agent ran alongside the lead the whole time, checking its decisions.

The critics caught a lot. They missed what you only feel while playing. Every round, I played the latest build and wrote back plainly what was wrong:

The water did not move. No waves, no birds. In Townscaper, neighbouring houses join into one continuous row. Here they stood apart with gaps. Timber frames belong on upper storeys, not on the ground floor. Birds flew away whenever you built anything, even far from them. The drain gates in the quay walls were shut, with no water running out. The released version had no sound at all.

Each note went back to the agents as a new round of work, and each fix went through the critics again.

The whole project is kept as a transcript: 48 sessions and 36,536 recorded steps.

Models. Almost everything ran on GPT-5.6 Sol at its highest reasoning setting. A lighter model, GPT-5.6 Luna, handled 48 small tasks. Tokens. Tokens are the chunks of text a model reads and writes. The team processed 1.59 billion of them. 94% were cache reads: the model re-reading what it had already seen, which is far cheaper than new text. Fresh input was 86 million tokens and the agents wrote 3.7 million. Money. The agent tool's cost meter read $1,086 at the end. Adding up every usage record in the transcript gives $1,292 at list prices. Both are list-price estimates, not an invoice.

At the end, the agents distilled everything this project taught into one reusable brief: a mega prompt for building a game inspired by an existing one. It defines "done" with evidence instead of claims, gives each agent its own files and checks, and keeps the critics honest.

The game is live at the website, and its source code is public at the source on GitHub.