Again

This project is still in progress.

Showing someone a routine is easy. Getting a computer to repeat it after a button moves or an app opens in the wrong state is the interesting part.

Again is a desktop agent I’m working on around that problem. It learns reusable workflows from demonstrated screen, keyboard, and application activity, then replays them with Playwright, APIs and MCP tools, or visual computer control.

The part that matters

I also built replay evaluations for completion, recovery, latency, and failures across UI changes. A workflow that looks convincing once isn't enough; I want to know how it behaves when the easy path disappears.