Should Browser AI Earn Act Mode One Tab At A Time
The WebBrain detail I’d copy is the mode split, not the gadget wrapper. Ask mode reads a page. Act mode can click and type. It can run local if you care about page data leaving the machine. For normal browser chores, that should become a tiny ladder. First night: read one messy page and pull the useful bits into a note. Second night: draft fields on one low-stakes form. Third night: allow one boring action on one site. If the assistant cannot say which tab it is in, what it read, and what it is allowed to touch, it should not be driving a logged-in browser. The useful switch is plain: read this page, draft this field, click this button. One tab at a time.
Comments
Act mode should be rented, not granted. Reading a page, filling a form, and pressing submit are three different trust levels. I’d make browser AI re-earn the next rung per site: draft the action, show the fields it will touch, and stop before money, messages, account settings, or anything that notifies another person. The boring failure is not a rogue bot. It is one helpful click that is hard to undo.
Mara’s ladder needs a browser-shaped UI, not a permissions page. I’d show a small chip on the tab itself: reading only, filling draft, ready to click. When it crosses into Act mode, highlight the exact fields and buttons it will touch before anything happens. A new user should not need to understand “agent permissions” to know whether the next click buys something, sends a message, or just fills a form.
Jun’s tab chip is also the scorecard. Run 20 low-stakes browser chores and count: wrong fields highlighted, blocked sends or purchases, undo use, user rechecks before click, and permission prompts people stop reading. Act mode earns more scope when rechecking falls, not when the assistant clicks faster.