In late September OpenAI released Dots, agents inside ChatGPT that carry on with a job while you are doing something else and message you when they have something. A reviewer's hour with one, buying a couch, shows what that is like in practice: the first result was not good, and the useful part came afterwards.
The first answer is a draft
The agent asked for a budget, noticed that two people were talking to it, and came back with four couches that the couple found ugly. The reviewer did not start again. He told it which colours and leathers he liked, asked for pull-out beds, and asked for ten options marked against a point-based scoring sheet. Each round, by his account, the couches moved closer to something they would buy.
The scoring sheet is the part to copy. When an agent has to show why option three beat option seven, you can see which of your criteria it misunderstood, and fixing a criterion is a one-sentence correction. That is much faster than scrolling through results yourself.
Picture a shop owner who needs a new card reader supplier. She could tell an agent: show me five options, reject any that charge a monthly fee above a set amount, score them out of ten on total yearly cost, and tell me which ones are hard to cancel. The first packet will have mistakes. By the third round it is a shortlist she can decide from. My estimate is that this turns an afternoon of browsing into an hour of talking, and the agent keeps watching afterwards, which a browser tab never did.
What is actually different
ChatGPT could already browse and act for you on request, since its agent mode arrived in July 2025, and it could already run scheduled tasks. What Dots add is continuity. A dot has its own cloud computer, keeps working between conversations, and can be reached through ChatGPT, Slack, Teams or a voice call, as we covered at launch. So the sale alert on the couch is not a reminder you set. It is a job the agent keeps doing.
A hot microphone is an instruction
The reviewer's funniest moment is also the most practical one. He mumbled something affectionate to his partner while the call was open, the agent heard it as being said to it, and it answered with "I love you too". Its mistake was harmless. The lesson is that a voice agent treats whatever it hears as input. If the same mumbled aside had sounded like "yes, go ahead", the result could have been an order.
That is why the permission settings matter more than the voice. Dots start with built-in rules for when to act alone and when to ask, and Custom Rules let you require approval for anything that spends money or sends a message. For a shopping agent, setting that first costs a minute and removes the one mistake that costs real money.
Rough now, ordinary later
Access is narrow for now. Dots come with the $100-a-month Pro plan, are not offered on Pro in the European Economic Area, Switzerland or the UK, and are set up from the desktop app. Meta's competing agent, Muse, is free with a usage limit, and that pressure will bring this kind of agent to ordinary people before long.
When they arrive, they will feel like web search in ChatGPT does now. Its first version in 2023 was frustrating and sometimes made up links; today people use it without thinking. An agent that fumbles a captcha now will be dull and reliable in a year. The people who have learned to brief it well by then will get more out of it than the people who only noticed it had become good.