Tony GorbProduct Designer & Builder

An experiment: a town where my coding agents live

AgentWorld is an experiment I built to see the work my coding agents do. Every repository becomes a district, every session a character. What I learned was how to keep that picture honest.

In this story3 sections
  1. The town is built from real signals
  2. Three rules kept the picture honest
  3. What the experiment taught me

A coding agent works for minutes at a time. While mine worked, I watched terminals scroll. A terminal tells you what one agent just did. It does not tell you which agent is stuck, what failed, or which repository nobody has touched in months.

So I ran an experiment. What if the work my agents do were a place I could look at? AgentWorld is the result: a hand-painted side-scrolling town, running on my machine. It is not a product and it is not released. It is how I tested an idea.

Every repository becomes a district, every agent session a character.

The town is built from real signals

The rule of the experiment was that nothing in the town is invented. Each thing on screen stands for one signal from a real repository. I pointed it at the public React repository. On 20 October 2026 React had 587 open pull requests, and in its district they stood as roadworks.

React's real page behind its district. The game's own card reads: 587 open pull requests, roadworks waiting for review.

Folders became buildings: the compiler/ folder is a crane workshop, and its card counts 4.3k files, about 183k lines. React keeps 2.4k test files beside its code, so its district has a Test Lab; a repository with no tests gets an empty one that says so. Open bugs walk the street as creatures, and clicking one opens the real issue. Each of those stands beside its real source on the AgentWorld project page.

Open pull requests, built as roadworks
587
Test files behind the Test Lab
2.4k

Three rules kept the picture honest

A world that looks busy while nothing happens would be worse than the terminal, because people believe what they see. The experiment only worked once three rules were written into its architecture.

  • Same inputs, same world. Seeds come from repository ids, not randomness. Reload the page and React's district is the same street.
  • Never offer an action the agent can't take. When I start a Claude Code session myself, AgentWorld only watches it, through hooks. Its character appears the moment the session starts, but its panel has no pause or message button, because the world cannot pause it.
  • Fake is always labelled. Demo agents carry a SIM badge wherever they appear, and they never earn coins. Coins and experience come only from real progress: finished quests, merged pull requests, passing builds.
A real Claude Code session joins the town. Its panel says what the world can't do for it, and hides those controls.

What the experiment taught me

Map every visual to one real signal, and let people click through to it. Show only the controls the system can honour. And mark anything simulated where it appears, not in a footnote.

AgentWorld stays an experiment. You can find it on my work page, next to the products I shipped.

Tony Gorb, Product Designer & Builder

Questions about this story, or a design role to talk about? Write to bytonygorb@gmail.com or find me on LinkedIn.

More stories