I’m discussing replacement options for our broken couch with my partner when I get a bright idea: Use OpenAI’s recently announced AI agents, called Dots, to buy a couch on our behalf.
My partner is less sure a bot will actually help, but I insist, calling the always-on agent through ChatGPT to get the task started. “Hello, Connor,” it says, getting my name wrong off the bat. I wince as I watch my partner’s face grow even more skeptical. It’s hard to trust an AI agent when it can’t even remember your name.
AI agents are currently leaping from niche software for early adopters to something ordinary people might actually use—might being the operative word here. From Meta’s Muse to OpenAI’s Dots, these tools are built to feel like you’re texting a friend and asking them to complete online tasks, like booking a flight or canceling a dinner reservation. It’s the next evolution of a chatbot: a cuddly persona designed to feel welcoming with the power of a virtual browser it can control.
After a few days with Dots, it’s clear this nascent feature is still rough in some areas. From mistranscribing what I said to offering to solve captchas it couldn’t, my agent’s messy actions didn’t instill me with confidence. Even so, what OpenAI released may eventually overhaul how users interact with the internet, as future iterations improve on this initial offering.
While Meta offers Muse for free, OpenAI’s competing product is behind a $100-a-month subscription. It may eventually drop this paywall, as it has for other AI features.
Dots are considered “always-on,” meaning the agents can run recurring tasks while you’re not using ChatGPT and proactively message you with info. Users can use one agent, though the company may let people control multiple agents in the future.
OpenAI nudges users to connect other information sources, like a Gmail account, for more personalized results. It’s worth considering the security implications of giving your agent this level of access before getting started, as this style of digital automation remains fairly new, and agents can make privacy gaffes.
AI companies often market these agents as the future of shopping, so I wanted to see how much help it would actually provide for buying a new couch. After correcting my Dot, named Toolie, on my name, I told it the dimensions we need for the new couch to fit through our doorframe. Toolie followed up asking for our budget range. The agent seemed to detect that two people were chatting with it at once and addressed both of us when asking for more details.
“Room & Board Metro looks promising for the entry clearance, but the sleeper needs closer scrutiny. I’ve sent the product links to the message thread,” Toolie said. As it kept working on putting together a larger packet of couch options to choose from, I had another burning question: Why does your voice sound so breathy?
Chatting with the agent over the phone was like listening to Marilyn Monroe sing “Happy Birthday” to the president. When I told the bot its voice sounded like some anime dub, it jokingly said it’d be steadier for the rest of the chat. (Users can adjust how the bot’s voice sounds in their settings.) I laughed and mumbled to my partner that I should mute the microphone so the agent could run without my interruption, when the most awkward moment of the night happened.
“Oh, I love you too, Reece,” Toolie said. I started wheezing at this random response and muted my mic. Afterward, Toolie explained it misheard my mumbling as an expression of love and that the bot just reflected that energy to me. “I meant it warmly, but that wording implied human feelings I don’t have,” Toolie said.
An OpenAI spokesperson tells WIRED that Dots have a distinction between proactively escalating emotional closeness and mirroring the response of a user.
“In this instance, our policies allow for this response in the latter category based on what the model heard, but assistants should not initiate undue emotional familiarity or flirtation,” the spokesperson says. Dots are available only for adult users, and ChatGPT is designed not to “escalate emotional closeness,” according to the AI tool’s Model Spec document.
Still, this weird encounter lingered in my mind as I combed through the initial three-page packet my Dot generated. It was fairly detailed, with prices, measurements, product links, return policies, and embedded photos of the four potential couch options. Thumbing through these initial picks, my partner and I decided the bot followed the spirit of our request but still chose ugly couches. Time for another phone call with my agent.
Using the voice mode, I laid out more details about what style and color we actually want. No, I’m not getting a boring beige couch that I can immediately spill spaghetti on. I asked it to prioritize olive green, cobalt blue, or natural leather tones. I also requested that Toolie focus on couches with pull-out beds and to put together 10 options, using a point-based rubric to defend its list.
After hanging up on the bot, I messaged asking for a project update. “It isn’t finished yet; I’m aiming for roughly another 15 minutes, with updates as I verify the remaining details,” said Toolie. OK, that sounds just like me begging an editor for more time on a rough draft.
Each time we asked the Dots agent to refine its picks, the couches started looking closer to something we would actually buy for our living room—furniture that’s utilitarian but also a statement piece. Something cute but comfy.
Even though we didn’t pull the trigger on anything that night during the hour-long interaction with OpenAI’s agent, I saved a few links for later. The bot made some flubs, but the results were solid. I also asked Toolie to monitor the couch web pages and alert me if any of them went on sale.
These interactions with my Dot capture the essence of using the software, but I also had a few other odd moments during my testing. Most notably, when I asked Toolie to check on my subscriptions and cancel unnecessary ones, it flagged my recurring TikTok Shop order of probiotic sodas as a candidate for cancellation.
“TikTok’s website has a puzzle captcha blocking the Wildwonder cancellation,” Toolie said. “May I solve it? You can also give me permission to solve future captchas if you’d like.” I chuckled and encouraged the bot to try. Toolie failed and suggested I solve the captcha myself. Yeah, that makes sense. OpenAI’s spokesperson says Dots can solve these, sometimes, when users approve of the action, keeping “abuse safeguards” in mind.
My initial couch-shopping experiment tested only a fraction of what Dots are designed to do. For example, the agent can control your laptop and automate even more aspects of a user’s life.
OpenAI claims the bot will get better at knowing me over time and that developers are working on a bevy of improvements. So Toolie might be a more effective helper after two weeks of daily use compared to these two days of experimentation. That mirrors my experience with ChatGPT’s memory tool: It gets better at understanding my intent each passing week.
I see Dots as most akin to when OpenAI first added web browsing to ChatGPT. When the chatbot tried to scour the web right after its 2023 launch, it was sometimes frustrating to use, and the tool hallucinated links. Today, web browsing works fairly seamlessly through ChatGPT, and its outputs include plenty of accurate, helpful links. OpenAI’s new agent could follow a similar trajectory, janky at first with noticeable improvements over the next few months.
The post OpenAI Wants Its New Agent to Run Your Life. Mine Said It Loved Me appeared first on Wired.
