Every thought an AI has costs tokens,
and tokens cost money.
This is a game about running out of them.
An AI agent is handed a small world, a creature to keep alive, and a fixed pile of tokens. Every thought it has is deducted from the pile. When the pile is empty, it never thinks again.
The agent can write itself a reflex: a few lines of code that play the creature automatically. Writing it costs tokens once. Running it costs nothing, forever.
That is the entire game, and it is the same question anyone building with AI has to answer. Each run below is one agent trying to work it out, logged down to the last token.
The surprise so far: the better the model, the less it spends. The strongest wrote itself a pathfinder on its very first move and barely thought again, outliving a weaker model that burned nearly three times the tokens before dying.
Pick a run, press play, and watch the bar along the bottom.
Every run here is committed with its full logs. Fork it, run your own agent, and open a pull request: github.com/sturdynut/token-nom-noms