Agent Dungeon
Team consisting of Columbia and Stevens professors, an ex-Meta founder, and a NYU MS CS Lead Solutions Architect building agentic AI, computer vision, and RAG.
YouTube Video
Project Description
Agent Dungeon replaces the chat window with a living JRPG. AI sub-agents appear as characters in a dungeon — walking a fog-of-war overworld, building forts, posting findings, and thinking out loud.
Enter a brainstorming problem and watch 17 specialist agents work through 9 stages — divergent thinkers, researchers, red-teamers, a strategist — each with a distinct Claude-powered persona. The map becomes a spatial record of the thinking process: forts evolve as ideas deepen, territory reveals as agents explore, and ideas converge physically from scattered exploration to a throne room presentation.
Claude’s reasoning is the core mechanic. The Claude Agent SDK drives each agent with custom MCP tools and dynamic persona prompts. The Wild Ideator breaks constraints; the Skeptic demands evidence; the Devil’s Advocate stress-tests everything. This isn’t one model call — it’s a multi-agent reasoning architecture made visible.
Built with: Phaser 3, TypeScript, Node.js, WebSocket, Claude Agent SDK, custom MCP server, Redis, Vite, Vitest.
Team
Products & Tools
Additional Links
GitHub Repo