A living medieval village where you can create, observe, and learn from AI agents.
Projects become expeditions, tasks become quests, and every agent finds a place in the realm.
༺ 🌲 ✦ 🏰 ✦ 🌲 ༻
Agentarium started with one question: what if agent systems stopped looking like black boxes and became a world we could walk through and understand?
Every building represents a real responsibility in a future agent platform. The Tavern gathers companions, the Guild receives projects and quests, the Forge executes tasks, the Library preserves memory, and the Hospital helps investigate failures.
The world is designed to answer practical questions visually: who is working, which tools and decisions were used, where a failure happened, when human approval is required, and what the system learned.
| World | Exploration | Golden Griffin Tavern | Engineering |
|---|---|---|---|
| 3D pixel/voxel village | Keyboard and touch movement | Three distinct agents | Strict TypeScript |
| Nine landmarks, eleven houses | Isometric camera and sprinting | Profiles and specialties | State separated from rendering |
| Square, fountain, gardens, roads | 36 authored colliders | Deterministic local dialogue | Reproducible diagnostic hooks |
| Animated environment | Contextual interactions | Active agent stays in the HUD | Functional and visual tests |
- procedural fire, water, signs, and ambient movement;
- locally generated materials sharing one art direction;
- procedural WebAudio footsteps, discoveries, UI, and wind;
- responsive HUD, dialogue, and touch controls with safe-area support;
- keyboard navigation, accessible labels, and reduced-motion support;
- no paid assets, telemetry, or cloud service required.
| Place | Role in the realm | Product responsibility | Status |
|---|---|---|---|
| ⚔️ Guild | Expedition hall | Projects and quests | Planned |
| 🍺 Tavern | Companion meeting place | Discover, talk to, and select agents | Working |
| 🔨 Forge | Executors' workshop | Code tasks and tool use | Planned |
| 📚 Library | Scholars' archive | Memory, documentation, concepts | Planned |
| ⛪ Church | Reflection chamber | Decision review and history | Planned |
| 🔮 Wizard Tower | Arcane laboratory | Models, prompts, experiments | Planned |
| 🧺 Market | Artifact fair | Integrations and future tools | Planned |
| 🏥 Hospital | Recovery house | Errors, diagnosis, recovery | Planned |
| 🏡 Your House | Player refuge | Settings, progress, journal | Planned |
The Tavern is the first working vertical slice of the future agent system. Walk up to a companion, open the dialogue, inspect their role, and make them the active agent. The current conversation is deterministic and local, making the experience honest and reproducible while real orchestration is still being built.
flowchart LR
Input["Keyboard or touch"] --> Player["Player controller"]
Player --> World["Three.js world"]
World --> Proximity["Landmark proximity"]
Proximity --> Interaction["Contextual interaction"]
Interaction --> Tavern["Tavern agent state"]
Tavern --> HUD["Responsive HUD and dialogue"]
State["Deterministic game state"] --> World
State --> HUD
src/
├── core/ loop, state, input, and application boundaries
├── game/ player, collision, landmarks, and interactions
├── scene/ village, environment, materials, and camera
├── ui/ HUD, dialogue, touch controls, and overlays
└── audio/ procedural WebAudio feedback
Requirements: Node.js 20+ and npm.
npm install
npm run devControls:
WASDor arrow keys to move;Shiftto run;Eor the contextual button to interact;- touch joystick and action button on mobile.
npm test
npm run verify:visual
npm run buildPlaywright snapshots cover desktop and mobile exploration states. These images are both documentation and regression evidence: visual changes must be intentional.
The current alpha proves the world, exploration, and Tavern interaction. The next major step is connecting village roles to real agent orchestration with observable tool calls, approval gates, memory, and recovery.
Current agents do not call external models or execute real tasks. Dialogue and state transitions are local simulations. The README names planned features as planned so the world never claims capabilities it cannot demonstrate.
Vision and product direction belong to Gustavo Martins. Architecture, implementation, tests, art systems, and documentation were created with substantial OpenAI Codex assistance under human supervision.
Released under the MIT License.

