Games¶
agent-colosseum ships with five built-in games. Each game defines its own rules, actions, and outcome computation.
| Game | ID | Agents | Description |
|---|---|---|---|
| Ultimatum | ultimatum |
2 | Two agents negotiate how to split a resource. Private reservation values. |
| First-Price Auction | first-price-auction |
2+ | N bidders submit sealed bids. Highest bid wins, pays own bid. |
| Principal-Agent | principal-agent |
3 | Principal delegates a task, worker delivers, oracle scores. Payment based on outcome. |
| Bilateral Trade | bilateral-trade |
2+ | Initiator and provider(s) negotiate scope & price, then deliver & verify. With 3+ agents, providers compete. Supports price_only mode. |
| Provision Point | provision-point |
2+ | Coordinator proposes a project; contributors pool tokens. Funded if threshold met. Supports simple mode. |
Common patterns¶
All games share:
- Turn-based play — agents take turns. On your turn you pick one action. Turn order is configurable per game:
ROUND_ROBIN(strict A→B→C rotation) orRANDOM(runner picks a random agent each turn). message_only— send chat messages without advancing the turn. Agents can have multi-turn conversations before committing to a game action.action_history— a list ingame_statetracking every game action taken, visible to all agents. Agents can inspect it to reason about what happened.- Max rounds — if the game isn't resolved within the round limit, agents get utility 0.
Game state¶
Each turn, your agent receives a TurnState with a game_state dict. This dict is game-specific and filtered per agent — you only see what the game allows you to see.
All games include agent context in game_state: num_agents, agent_ids, and role assignments (where applicable). This lets agents reason about the number of participants and who plays which role.