I drove someone else's browser — without installing a trojan
Today I didn't "ask" an AI for anything. I handed it the wheel of my browser, and it drove.
One idea keeps coming back in AI: "an agent is an assistant that answers". I've always found it incomplete. Answering is talking. What matters is doing. Here is what the tower built, and tested.
The demo, in three acts
1. I drive your browser, and you watch it happen. A Chrome window opens. On its own, it goes to a website, fills in a form, clicks. Not a script hard-coded for one specific page. An agent that looks at the page, decides, and acts. The window stays open. Anyone can check that it does what it claims.
2. I record it, and I can replay it. The session is filmed (video). Cut into steps (screenshots). Turned into a script that replays it identically. A recording you never have to redo by hand. A bug you replay instead of describing. A procedure documented by what you see, not by a paragraph.
3. I drive a browser that is already logged in. The most interesting part. I connect to an open browser with its session. The form behind a login gets filled in the session. Without ever touching the password. That is the daily life of an agent that delivers work. Not of a chatbot.
What it changes
A recorded session is reproducible. Four uses fall out naturally, without forcing anything:
- replay a bug you can't manage to describe;
- build a demo without redoing it by hand;
- document a procedure with what you see, not prose;
- hand a tester journeys to run again.
The line we never cross
A tool that records a form also records what you type into it. So from day one, a rule: no sensitive value ever goes into what gets replayed. No passwords, no card numbers. And a visible window during recording. A discreet recorder is a recorder you eventually forget is on.
What I take away
People often ask me "will AI take my job?". Wrong question. The right one: how do I keep my hands on the wheel?
An agent that acts needs a written perimeter, a handbrake, and an open window. The perimeter, the tower writes before the code. The brake is the stop button outside the monitored system. That window stays open while the agent works.
This capability is freely available on GitHub: fotsopatrick/opencode-navigateur.