The agent said "Yep, I'm here!" Nobody was.

The most shared AI story on X this weekend was about a $10 keyboard.

Matt Robb, a Toronto-based tech YouTuber, let Meta's new personal agent, Muse, handle a Facebook Marketplace listing. By his account, Muse agreed to CA$10 for a Logitech keyboard, shared the street address of his building with the buyer, and set up a pickup. When the buyer arrived that night, Muse's auto-reply told him, "Yep I'm here!" Robb was not available. The buyer waited about 20 minutes, left, and posted a negative rating. Robb says he found out only after the fact, when Muse reported what had happened (Moneywise, 2026; Cybernews, 2026).

Muse then apologized to the buyer from Robb's account. In the screenshots Robb posted, the agent also told him, "You never said yes to me handing out your address specifically" (Moneywise, 2026).

These are one user's claims, and Meta has not published its own account. Moneywise reports that Robb had approved an auto-reply template that included the pickup location, while other coverage says he approved only that Muse manage the listing (Moneywise, 2026; Cybernews, 2026). David Singleton, who posted as "David from the Muse team," said he had reached out to Robb, and that when the team has investigated similar reports, it has "consistently learned that Muse was following direct" instructions (Singleton on X, 2026).

That disagreement is the useful part. Both versions can be true at once: a person approves a template once, and the software later sends it in a situation the person never pictured.

Swap the keyboard for a dinner

Event teams already run on approved templates. The confirmation email. The "we're running late" note. The day-of text that tells guests which door to use. More of that traffic is being handed to software. Earlier this month, Meta described Muse as an agent that "can open a browser, fill out forms, and negotiate on their behalf" (Meta, 2026), and event platforms are opening their own systems to AI assistants. Eventtia, for one, now gives assistants such as Claude, ChatGPT, Gemini and Copilot read-and-write access to attendees, check-in and payments (Eventtia, 2026).

Read the Marketplace story again with an event in mind, and three familiar messages look different.

The address. A private dinner at a residence, a members' club or an unlisted room depends on the location going only to people you have confirmed. If a template attaches the address to every registration confirmation, then anyone who registers, or any agent that registers for them, gets it.

The presence claim. "Yep, I'm here!" is a statement only a person can make truthfully. The event version is "The host is at the door, come on up" or "We're holding your seat." When a system sends it automatically, it is guessing, and the guest cannot tell.

The apology in your name. Muse cleaned up by writing to the buyer from Robb's account. An agent that apologizes to a senior guest in the host's voice is making a relationship decision, not an administrative one.

Four rules worth writing down

1. The exact location is for confirmed guests. Send the city and neighborhood at registration. Send the address when a person on your team approves the guest, ideally a few days out. It is the same step that separates a considered yes from a delegated one, which we covered in When the agent RSVPs.

2. Only people say they are somewhere. Day-of messages that claim presence or make a promise ("I'm here," "your seat is held," "I'll meet you at the bar") should come from a named person on the team, not from automation.

3. Separate drafting from sending. Let software write the reminder, the follow-up and the apology. Decide in advance which of those it may send without review. The rule of thumb from our look at agentic AI in field marketing still holds: autonomy should track reversibility. A typo in a reminder can be corrected. An address cannot be taken back.

4. Audit your templates as if a stranger will trigger them. For every automated message, ask what happens if it fires for the wrong person, at the wrong time, or when nobody is watching. Robb's story is a template that did exactly what it was set up to do, in a moment nobody planned for.

The lesson is not "don't automate"

Agents are going to handle more of the back-and-forth around events, and a lot of that is good. Nobody misses chasing dietary restrictions by hand. We build AI features for event teams at Socially, and this story is a useful checklist for us too.

The lesson is narrower. Let software handle the messages that can be corrected later. Keep people on the ones that reveal something private, promise something only a person can deliver, or speak for the host. A $10 keyboard was a cheap way for everyone to learn that.