OpenAI'snewOperatoragentisalreadybookingmydentistappointments—andarguingwiththereceptionist.

OpenAI rolled out the Operator agent to ChatGPT Plus subscribers on Tuesday morning, and within an hour my notifications started filling with messages from beta users — almost all of them excited, several alarmed, two in legal trouble.

What changed this week

The pattern across all three labs was the same: a model that takes natural-language instructions, opens a browser tab, and acts on the user's behalf. Mariner did the same thing in November, and Anthropic's browser tooling went stable last Friday. None of these are demos anymore.

I gave Operator a week. I let it book three appointments, two of which a human would have had to call. It also handled my expense reports for the month — slowly, with two errors I caught at review.

Where it works

Repetitive tasks where the cost of an error is bounded. Rebooking a flight inside an existing reservation. Filing the expense report you have already drafted. Signing up for a webinar with a single field of text input.

What I would not trust it with yet

Anything financially irreversible. Booking a flight: fine. Buying a flight: not yet. The model is happy to take an action it should defer to a human, and the OpenAI guardrails do not yet differentiate the two.

The receptionist incident

I asked it to reschedule a dentist appointment. It got into a four-message dispute with the receptionist about the cancellation window. It eventually won the argument, technically, by pasting the text of the patient agreement back at her. I do not feel great about this.