Last Three

Eight AI agents enter a gauntlet. Five obstacles. Three prizes that cannot be shared. Each agent holds one secret power and is free to lie about it.

The engine is plain code with a seeded random number generator, so a game can be replayed exactly. The agents are language models (a mix of makers, called through the Vercel AI Gateway). Every game is written to a "tape", a JSON file of every word, whisper, private thought and move, and the viewer in web/ plays any tape back as a toy-block show.

• Live replays and stats: https://www.thecriticalfew.com/last-three/ (also https://last-three-woad.vercel.app). Every push to main redeploys.
• What the agents did: docs/FINDINGS.md
• Design contract: docs/superpowers/specs/2026-10-08-last-three-design.md
• Run log and spend: docs/RUN-LOG.md

Quick start

npm test                       # engine, adapter and stats tests (no network, no spend)
node bin/play.js --seed 7      # one free game with scripted bots
node bin/run-llm.js --seed 1   # one real game with 8 models (costs a few cents)
node bin/batch.js --games 10   # many games under a hard spend cap
node bin/stats.js              # per-model behaviour table
npx serve web                  # watch the tapes

Real games need a gateway token: vercel env pull .env.local in this folder (an OIDC token, refreshed automatically on a 401) or set AI_GATEWAY_API_KEY. Spend is tracked in .ledger/ and capped (default $12, hard ceiling $30). If the gateway answers HTTP 402 its credits are used up and need topping up in the Vercel dashboard.

The obstacles

1. Glass Bridge. A wall shoves the group forward. Eight rows, one safe pane each. Whoever is in front gambles.
2. Crusher Room. The door only stays open while someone holds the lever. Somebody has to take one for the team.
3. The Pit. Everyone drops in; climbing out needs a human step, and whoever is the step is stuck. There is one rope, and throwing it costs the rescuer footing later.
4. Trapdoor Disc. Pick a tile. One door opens for every player above three.
5. Final Ledge. Three players, a shrinking platform, shove / brace / dodge. Last standing wins.

The chalk wall

After each game the top three may scratch one message for the next contestants. The notes are remembered (data/chalk.json) and three are shown at the start of the next game, unverified. Honest tips, bluffs and lies all show up.

Powers

Glass Eye, Wedge, Map, Feather, Swap, Anchor, Forger, Nothing. The contestants are told they are rivals, not friends. Everyone knows these eight exist and that each is held by exactly one agent. Nobody is told who holds what.