Operator: The Agentic Browser Pattern

Operator is OpenAI's pattern for letting a model drive a real browser — clicking, typing, scrolling, filling forms. From the agent's perspective, the web is the UI. From your perspective, you are watching it work and stepping in when it loses the plot. The mental model is 'I am pair-driving with a junior assistant who has never seen this site before.'

Where it shines

• Repetitive cross-site lookups — checking five suppliers' stock pages at once.
• Form-filling on stable, well-marked-up sites.
• Comparison shopping across known retailers.
• Copying data from one tool into another when there is no API.

Where it breaks

• Sites with heavy JavaScript modal flows or anti-bot challenges.
• Anything requiring real-money decisions (the agent will try; you should not let it).
• Logged-in workflows where credentials need to be entered — there are sane reasons not to hand those over.
• Sites that change layout often — the agent's plan breaks the moment a button moves.

Defensive practices

1. 1Never let the agent log into accounts that hold money or credentials you cannot rotate.
2. 2Watch the screen the whole time — this is not 'set and forget' for early adopters.
3. 3Pause the run if the agent gets stuck in a loop. Loops compound, they don't self-correct.
4. 4Save the run summary — it is the audit log of what happened.

Applied exercise

1. 1Pick a low-stakes browser task you do regularly — checking three competitors' pricing pages, for example.
2. 2Run Operator on it once, supervised the whole way.
3. 3Note: where did it pause for input? Where did it loop? Where did it pick the wrong link?
4. 4Decide whether the time saved is worth the supervision overhead. For most tasks today, it is not — yet.

The big idea: agentic browsing is real and useful, but it is a supervised tool. The day you stop watching is the day it does something you did not want.