August 2026

Habitat

A zoo-management sim built almost entirely by an autonomous agent org.

gamesagents

What it is

Habitat is a browser-based zoo-management simulation in the spirit of the 2001 classics — but the interesting part isn’t the game. It’s that a small organization of Claude agents built it, mostly unattended, across five bounded runs.

Reimplemented mechanics and original work only. No assets from any existing game, ever.

The agent org

Three roles, and the separation between them is the whole design:

• Manager — the session itself. Owns the objective, decomposes it into a dependency-ordered backlog, writes the invariant tests, dispatches one task at a time, merges verified branches, and batches up any question that needs a human.
• Builder — a subagent spawned per task, in its own git worktree. Implements to acceptance criteria, runs the tests, pushes its branch, reports back in a fixed shape. Never merges its own work.
• Verifier — a subagent with clean context and read-only tools. Re-runs the tests, typechecks, runs an independent 10,000-day soak, and checks the repo’s own rules. Returns PASS or FAIL with evidence.

The verifier earning its keep is the point: in the first run it caught a real animal-escape bug that the builder’s own tests had passed.

Runs are bounded — “work for an hour,” “do at most N tasks” — and everything durable lives in git, so a fresh session picks up exactly where the last one stopped. Run 1 wrapped with about eleven minutes left rather than start the largest, most failure-prone slice: the timed-run discipline working as intended.

Where it got to

Five runs produced the entire simulation end to end — deterministic sim core, an economy where day-30 profitability has to be earned rather than handed to you, animals and breeding, staff, guests with pathfinding, an isometric renderer, a full interactive UI, save and load. 118 of 118 invariants green, deployed and verified live with no console errors.

And it still looks nothing like a finished game: placeholder geometry, flat grid, colored dots, no art, no audio.

That gap is the most instructive thing here. The manager was right to report “complete” — the objective’s definition of done deliberately scoped art out. So “done” meant the systems work, not a game you’d hand to someone. Nothing broke. We just reached the edge of the map we’d drawn, which is a much easier failure to fix than the alternative.

What one hour costs

Measured from the subagents’ own completion reports, run 1:

• Builder: ~63.6k tokens average per task
• Verifier: ~32.6k average
• One build → verify cycle: ~96k
• All 13 subagents: 668,411 tokens
• Full one-hour run: ~820k tokens, 6 verified slices

Rule of thumb: an hour is roughly 800–850k tokens and about six verified slices of work.