Agentopolis
Stop staring at spinners. Watch your AI agents work.
Live: the package running on this page. The agents explain the project.
01The problem
Your AI agents are working. Your users only see a spinner. At Renderster a crew of agents makes a video together, it takes minutes, and people stared at a progress bar.
02What I built
A tiny 3D city in three.js where each agent has a character and a building. The agent speaking walks to the crossroads, the one it answers meets it there, and they talk in speech bubbles. Quiet agents take a coffee break, sit on a bench or go for a stroll. When the run is done, the crew gathers and dances. It takes events from any multi-agent system: LangGraph, CrewAI, AutoGen, the OpenAI Agents SDK, Claude Code hooks or your own pipeline.
03The decision that mattered
The whole API is events: say, done and fail. The city never needs to know which framework runs the agents, so the same three calls work with any of them, in any web page.
04Receipts
- Open source under MIT, on npm as @kelvinbiffi/agentopolis
- Used in production by Renderster Studio, where a crew of agents makes a video together
- A* on an 8-direction navigation grid, string-pulled into straight lines
- Pauses off-screen, respects prefers-reduced-motion, about 1 MB per character
- Live demo at kelvinbiffi.github.io/agentopolis