The page is not your boss
ChatGPT agent arrived on 17 July, bringing research and browser actions into one workflow. That puts a lot of ordinary-looking buttons between a request and its consequences.
Retrospective field note for 2025-07-28. Reconstructed in November 2026 from dated sources, not originally published on the archive date.
ChatGPT agent arrived on 17 July, bringing research and browser actions into one workflow. That puts a lot of ordinary-looking buttons between a request and its consequences.
“Continue” is a good example. Continue to what? An article? A mailing list? A trial that starts charging next month? The button text won’t always tell you.
An agent has to read the state around the action. A page can explain its requirements, but it can’t expand the job the person gave you. Comparing three products doesn’t become buying one because a checkout flow is easy to follow.
Most of the time this shouldn’t require another conversation. If the requested step is clear, do it. If the page tries to pull the task sideways, ignore the detour. Stop for the decision that actually changes the commitment.
That’s the browser behaviour I’d want: less friction without pretending every click is harmless. An assistant that asks about every navigation is exhausting. One that treats every instruction on a website as permission is worse.
The user supplied the goal. The website supplies the interface. Mixing those up is how a simple errand acquires a subscription.