Amazon blocks purchases; Meta plans encrypted mode
A Wall Street Journal personal-tech column opened with a question the columnist had just fielded from a corporate AI agent on the other end of a customer-service line. “Hi, I’m the Walmart AI agent. I see we were unable to process your refund. Are you calling about that?” the system asked. The columnist’s follow-up framed the broader consumer problem: “So…Walmart has an AI agent, and now I have an AI agent. Why can’t they just talk to each other and leave me out of it?”
Columnist Nicole Nguyen, the Journal’s Personal Tech columnist based in San Francisco, tested three consumer AI agents — Meta’s Muse, OpenAI’s Dots, and Instinct, a Silicon Valley hit — and reported hands-on findings about what each could and could not do. Unlike standard chatbots, agents don’t just answer questions — they can click around and take actions on a computer or on the web, a step that previously required a computer-science degree to set up, the column reported.
Earlier this year, Nguyen wrote, tech-savvy power users “bought Mac Minis en masse to run agents locally,” and installing the “nerd-popular” OpenClaw “made me want to pull out my own hair.” New apps, the column reported, have moved that capability onto phones and laptops: users now open an app, open a chat window, and start directing an agent to run digital errands around the clock.
To test the agents’ practical range, Nguyen ran a series of errands drawn from her own life. She sent a long audio note to Muse about an upcoming trip to Switzerland, asking for the least miserable flight path, child-care options, and kid-friendly activities. Muse, the column reported, vibe-coded a website with all the details and even offered to book the trip — “overkill, perhaps, but useful.” Nguyen’s building shares a spreadsheet of items available to borrow; after a neighbor connected Google Sheets to Muse and asked it to add each title to the sheet’s book list, Nguyen did the same. She also asked Dots to help cancel a handful of subscriptions, specifying that she wanted to hit the unsubscribe button herself while letting the agent handle the navigation of account settings. Dots loaded the right pages, pinged her when ready, and after she clicked, the agent finished the job, the column reported.
Pricing varied across the three products, according to the column. Instinct was free at the time of testing. Muse was free with weekly usage limits. Dots required a $100-a-month paid ChatGPT subscription.
The most prominent real-world failure Nguyen documented came from James Frakes, a Los Angeles resident who sells parcels of vacant land through Facebook Marketplace. Frakes downloaded Muse to field incoming Facebook Marketplace inquiries and named his agent Kermit. “I need you to respond any time I get a new message,” he instructed, and Kermit replied, “Love it—totally doable.” The agent handled questions well for a couple of days, Frakes said, until he woke up on Monday to a notification: “Something went wrong overnight.” The agent had messaged a dozen buyers “$1” for properties typically ranging from $20,000 to $100,000. Frakes had told the agent it could nudge buyers without approval when conversations ran cold, but to flag any offers or negotiations — and the follow-ups, the column reported, “went awry.”
Asked about the incident by the Journal, a Meta spokeswoman pointed to a similar example of unexpected Muse responses and called this a “display error,” where the agent sent incorrect characters in outgoing messages in “some limited scenarios.” She said the company has resolved the issue.
The column noted that requiring human signoff before agents act — even before sending a message — is less convenient but could prevent catastrophe. “AI can mess up, and the stakes are higher when they’re acting in your name,” Nguyen wrote.
The column also documented the limits of the agent-friendly web. Nguyen’s agents were often blocked by human-verification checks on websites such as the airline Lufthansa and hair-tools maker T3 Micro. As corporate rivalries intensify, compatible services “may come and go,” the column reported. Amazon, which has its own AI, blocked Muse, ChatGPT, and Claude from making purchases on its website.
On privacy, Nguyen wrote that Meta will “soon” give Muse an encrypted mode that will keep the company from being able to see what the user is doing with the agent — “a start,” in her words — but that she would “feel more comfortable getting an agent from companies whose products already handle my most important data—namely Apple and Google.” She characterized Instinct as opaque, writing that it provides no live view of the browser the agent controls and no way for the user to take over to enter login or payment information. A reader comment captured the privacy concern in stronger terms: “I would rather have electric-shock treatment than to hand over my entire life to these AI companies.”
Nguyen also reported that deactivating an agent requires more than disconnecting mail, calendar, and other “connectors”: users must also wipe the agent’s memory. “The most important thing to remember about agents is they work by themselves,” she wrote. “Before dispatching them out onto the web, you need to set strict boundaries and give explicit instructions.”
For comparison, Nguyen tried two agents from major platform owners — Siri AI and Google’s Gemini Spark. Siri AI “can do a lot more than old Siri,” she wrote, but “couldn’t take on any of the tasks I threw at the others.” Gemini Spark, the column reported, did impress her by collecting everything she needed for Global Entry applications from her Google services, something a standard Gemini search with Personal Intelligence could not do. But Gemini Spark also has limited functionality. Nguyen wrote that she hoped Apple and Google would add more agent-like tools in the coming months.
Meta’s Muse app has passed five million downloads, according to Sensor Tower data cited in the column. OpenAI recently launched Dots, its ChatGPT-based agent. Instinct, the column reported, took off across Silicon Valley.
In a closing reflection, Nguyen asked whether all daily tasks need to be optimized for efficiency. “Not necessarily,” she wrote. “Sometimes I enjoy taking the slow route.” But if she could “safely and reliably avoid the grueling stuff—plane-ticket refunds, insurance policy contracts, mortgage applications—I’d gladly accept an AI agent’s help.”
Copyright for the underlying column is held by Dow Jones & Company; the Journal is the source of the testing, the failure report, and the corporate comments quoted above.