Why the agent belongs on your machine

The context problem

An agent working in a browser sandbox has to be taught everything about your repository: where the build cache lives, which test command the team actually uses, why that one directory is generated. You paste. It forgets.

An agent that runs on your machine already knows. The workspace is your checkout. The shell is your shell. $ npm test runs the same suite you run, with the same environment, and the failures it reads are the failures you would see.

What we optimised for

1. Short feedback loops. Every run streams back the commands it executes, so you can interrupt the moment it goes the wrong way.
2. Legible diffs. The agent proposes patches; you stay the one who merges them.
3. Nothing hidden. No separate sandbox state to reconcile with your working tree.

What this means day to day

You keep working while the agent works. When it finishes, you review a diff, not a chat log.