A City-Builder Interface for Managing 78 AI Coding Agents

When you run dozens of AI agents at once, a sidebar of chat threads stops working. Jarren Rocks built a video game instead: every project is a cluster of hexagon tiles on a 3D planet, every agent is a little astronaut walking around on it, and the ones that need him raise a glowing speech bubble. He has 78 agents total, around 60 visible on screen, and the whole thing is a spatial dashboard sitting on top of Claude Code.

The problem: task lists don't scale to agent fleets

The video's premise is in its caption: "a sidebar full of different tasks is not the ideal state" for managing many agents. Once agents run long tasks in parallel, the human job shifts from writing prompts to triage. You need to see at a glance which agents are working, which are blocked, and which finished. A flat list buries that state. Jarren's answer is to borrow the oldest trick in strategy games: put everything on a map and let visual state carry the information.

How the world maps to work

The app is called Cosmo Builder (the name is visible in the macOS menu bar). Each hexagon tile is part of a project, and tiles cluster into named regions you can drag around as a unit. Labels visible in the frames include personal-site-2025, emra-app-builder, Videogen, random, and jarren blog posts. Buildings, solar panels, and supply crates fill out each tile, and small astronaut robots with glowing faces wander between them. A central space station with a slide sits at the edge of the colony; as he adds more projects, new agents walk out of it onto the map. It reads like a cross between Catan's hex grid and an idle colony sim, except the colonists are coding agents.

Status you can read from across the room

Agent state shows up as floating speech bubble icons over the map. A glowing question mark means an agent is blocked and waiting on input. A green hammer means it is actively working. Clicking a bubble opens a status card with the agent's name, its state ("Waiting on you" or "Working"), the model it runs on (the cards show an opus-5 tag), how long ago it last acted, a progress bar, and two actions: Open and Archive. In the demo, his personal site cluster has three agents, one flagged as needing help. Another agent on the Videogen tile is mid-task recutting a video. One gap he calls out on camera: there is no way yet to mark an agent for follow-up ("I got to mark this guy... maybe I have to add a capability for that").

Claude Code underneath, game on top

The map is a viewer, not the whole system. When he clicks Open on a blocked agent, it launches Claude Code, "because that's the easiest way to manage the conversation." One frame shows the underlying transcript for the blog posts agent: a long log of dev work with file diffs, bug descriptions, and commit summaries. The game layer only handles the high-level overview: what exists, what is running, what needs attention. That split keeps the hard parts (conversation history, tool use, code review) in a tool that already does them well. A macOS banner in one frame ("This web app was reloaded because it was using significant memory") reveals the whole thing runs as a web app.

Cosmetic depth as a feature

The planet itself is swappable. He shows the same colony on a green terraformed world, a gray moon ("this maybe makes more sense but it's maybe not as good looking"), and a red Mars variant. The agents have faces, walk with idle animations, and get talked to like pets ("You go, little guy"). That sounds frivolous, but it does real work: making the fleet legible and pleasant to look at means he actually checks it, the way people check a home screen widget but not a buried admin page.

Key Takeaways

• Managing many concurrent agents is a monitoring problem, and spatial interfaces surface state faster than lists.
• Map metaphors carry real information: hex clusters = projects, walking astronauts = agents, speech bubbles = status (question mark blocked, hammer working).
• Keep the conversation layer in an existing tool. The game opens Claude Code for actual agent interaction rather than rebuilding chat.
• Status cards need little: name, state, model, recency, progress, and Open/Archive actions cover triage.
• Charm is functional. An interface you enjoy looking at is one you actually monitor.

Resources

• Claude Code - Anthropic's terminal-based coding agent, the tool each in-game agent opens into.
• @jarrenrocks on TikTok - The creator's profile. Cosmo Builder itself is a personal project with no public release found.