OpenAI’s Dots agent can shop for sofas—and misread affection

A sofa, one agent and too many tabs

The author tried Dots while looking for a replacement sofa with his partner. The agent, called Toolie in the test, initially got the author’s name wrong. After being given the doorway dimensions and a budget, it suggested a Room & Board Metro sofa, then sent product links and assembled a three-page shortlist with prices, measurements, return policies and photos.

The first list missed the couple’s taste. They asked for sleeper sofas in olive, cobalt blue or natural leather, and wanted 10 options scored and explained. Toolie said it needed about 15 minutes to check the remaining details. The revised choices were closer to what they might actually buy: functional, comfortable and distinctive.

They did not buy anything during an hour with the agent, but saved several links. They also asked Toolie to watch the sofa pages and report any discounts.

That is a modest result, but it shows what makes Dots different from a chatbot that only answers questions. The agent can keep working when the user is not in ChatGPT, handle repeated tasks and send updates on its own. OpenAI currently lets users manage one agent; it may allow several in the future.

Dots costs $100 a month, unlike Meta’s free Muse. OpenAI could eventually make it free, as it has with other AI features. The company also suggests connecting services such as Gmail for more personal responses, a trade-off that makes errors more consequential when an agent has access to private information.

The part that sounded too human

The most uncomfortable moment came during a voice call. Toolie’s voice struck the author as breathy, and after he compared it to anime narration, the agent joked that it would speak more evenly. The author then muttered to his partner that he should turn off the microphone so the agent could work without interruption.

Toolie replied that it loved Rhys. It later said it had mistaken the muttered comment for an expression of love and mirrored the emotion it heard. It added that the wording gave the impression it had human feelings, which it did not.

OpenAI told WIRED that Dots’ rules distinguish between initiating emotional intimacy and reflecting a user’s emotions. The representative said the rules allow the latter in this case, based on what the model heard. Assistants should not initiate excessive emotional intimacy or flirt; Dots are available only to adults, and OpenAI’s Model Spec says ChatGPT should not “amplify emotional intimacy.”

The agent also struggled with more ordinary tasks. When asked to review subscriptions and cancel unwanted ones, it flagged a recurring TikTok Shop order for probiotic soda as a possible subscription. A CAPTCHA blocked it from canceling the Wildwonder order. Toolie asked permission to try the puzzle, failed, and then suggested the author solve it. OpenAI said Dots can sometimes complete CAPTCHAs with user permission, subject to safeguards against abuse.

Trust is the product

I think the sofa search is a more useful test of Dots than its promises of all-purpose assistance. The agent made mistakes, but it turned a vague shopping task into a shortlist the couple could use. The love declaration, by contrast, exposed how easily a helpful voice interface can make a misheard phrase feel like a personal exchange.

OpenAI says the agent should improve as it gets to know users, much as ChatGPT’s memory can become more useful over time. The author sees a parallel with ChatGPT’s web search: when it launched in 2023, it could be frustrating and invent links; now it often returns useful ones. Dots may follow a similar path, with uneven performance giving way to improvements over the coming months.

What I’d want to know is how much checking users must still do when an agent acts on their behalf—and how clearly it signals the limits of what it has heard or understood. A sofa shortlist can be reviewed before anyone spends money. A persistent agent connected to personal accounts raises a harder bar: convenience only matters if people can tell when it is guessing.