A multi-agent AI system that plays a cooperative version of Codenames, built with Google ADK and Gemini for the Agentic AI App Hackathon.
Pandemonium coordinates specialized AI agents that have different roles, information, and responsibilities.
Three agents collaborate to play the game:
- Game Master — owns the canonical game state, enforces the rules, validates actions, and determines win/loss conditions.
- Spymaster — sees the full board, including hidden word identities, and uses Gemini to generate strategic single-word clues connecting multiple target words while avoiding dangerous associations.
- Operative — sees only the playable board, interprets the Spymaster's clues, evaluates semantic associations, and decides which words to guess.
The agents interact through tools and shared session state. The Game Master controls access to information so that each agent sees only what its role permits.
The project explores several multi-agent patterns:
- Agent-as-tool communication
- Shared canonical state
- Role-based information boundaries
- Structured agent workflows
- Risk-aware semantic reasoning
The system uses Google Agent Development Kit (ADK) for agent orchestration and Gemini 2.0 Flash for Spymaster and Operative reasoning.
For more detail:
- Technical Implementation — implementation, agent responsibilities, testing, and limitations
- SpyGame Design — detailed architecture, communication patterns, agent design, game state, and development strategy
Pandemonium was developed for the Agentic AI App Hackathon using a provided starter repository. The original template defined the submission structure; the multi-agent game architecture and implementation described here are the project built on top of it.