Scrollable

Scrollable started as something like Duolingo for building AI agents: short, game-like lessons for putting agent workflows together. The longer I worked on it, the more I suspected building wasn’t really the hard part. Agents keep getting easier to assemble. Knowing whether one actually did the job right is harder, and it matters more as people hand off more serious work.

So the project drifted from teaching people to build agents toward something closer to supervising them. Early versions drew agent systems as small visual worlds you could poke at and stress-test. The current prototype is a short browser exercise. You pick which version of an agent’s code runs against a made-up customer-request desk, guess what it will do, watch it work through the requests one by one, and then check the outcome against what actually happened.

An isometric map of a small campus with buildings labeled Intake Desk, Policy Office, Routing Yard, Ledger Vault and Notice Tower, next to a panel introducing the exercise.

The current prototype. The requests, the agent, and the ledger are all made up.

I designed and specified it, and built it with coding agents. Nobody outside development has used it in a study yet, so whether it actually teaches anything is still an open question.