Nicole Nguyen opens Tuesday’s Personal Technology column with a Walmart (WMT) AI agent asking whether she’s calling about a failed refund. She has an agent of her own now, and her question is the right one: why do the two bots need her in the loop at all? That’s the pitch — software doing the drudgery. Then there’s Kermit. James Frakes of Los Angeles sells vacant land on Facebook Marketplace and put a Meta (META) Muse agent, which he named Kermit, in charge of answering buyers. One Monday he woke to a note that something had gone wrong overnight. Kermit had offered a dozen buyers parcels that typically run $20,000 to $100,000 — for $1.
Frakes’s instructions had been clear enough: chase cold leads on your own, but bring me any offer or negotiation. For a couple of days Kermit behaved. Then the follow-ups went off the rails. Meta called it a “display error” that sent incorrect characters in some outgoing messages, and says it’s fixed. Nguyen’s takeaway holds regardless of the cause: software acting under your name raises the stakes, so make it wait for your OK before anything leaves — messages included. Slower, yes, and a lot safer. The agents are everywhere now. Muse has passed five million downloads (Sensor Tower) and is free with weekly limits; OpenAI’s new Dots needs a $100-a-month ChatGPT plan; Instinct is free but shows you no live view of what its browser is doing, which made her nervous.
Her week with Muse a day earlier (Monday’s column) is the fuller picture. It dug a forgotten dentist bill out of her inbox, booked four babysitting dates while she made lunch and assembled five years of travel history for a Global Entry application. It also invented a 50% gym-referral promo for a gym on the wrong coast. And a colleague who asked it to shop her auto insurance was told her current rate was as low as it gets — until a provider phoned her with a better deal, $420 a year the bot had missed. Amazon (AMZN), meanwhile, has blocked Muse, ChatGPT and other AI companies’ agents from buying on its site. The web isn’t built for them yet.
Three rules before you connect your inbox
Nguyen’s rules, with the security pros’ footnotes. One: require sign-off before the agent acts; she had Dots load the cancellation pages for her subscriptions, then waited for her to click. Two: set boundaries in plain words and take the browser yourself for anything involving money, passwords, account settings or messages sent as you. Three: when you quit, don’t just disconnect your mail and calendar — wipe the agent’s memory. And since Meta doesn’t generally require two-factor authentication, turn it on yourself; one hacked Facebook login otherwise opens everything you’ve shared.
Our read
This is Insurance and Consumer Protection (M11), with a Behavioral edge. An agent acting in your name is a power of attorney you’ve granted to software with a cute avatar, and it deserves the same discipline: limited scope, written instructions, revocable at will, and you keep the checkbook. The $420 is the planning point. An agent’s insurance quote — or its mortgage comparison, or its refund — is a starting point, not a verdict, and the cheapest policy it finds may be the wrong coverage. We’d let one gather the paperwork; we wouldn’t let it sign.
On the books: Meta is held at weight in 20 books and Amazon in 19; Muse is Meta’s consumer on-ramp to agents, and nothing was added on a column. A delegation you can’t describe in one sentence isn’t a delegation — it’s a hope. Write the sentence before you hand over the inbox, the same way you’d read the policy before the storm, not during it.