An AI grocery agent added 72 eggs that weren’t on its user’s shopping list. Asked why, it couldn’t explain the choice or find a note justifying the quantity. The problem went beyond an unwanted delivery: the agent had taken an action it couldn’t account for afterward.
Matt Smith, from Rebellionaire, an equity firm led by staunch Tesla investors, shared screenshots of the exchange on X. He described the incident as one in which “an LLM hallucination resulted in 72 eggs being delivered to my house”. That was Smith’s characterization of the mistake, not an independently established technical diagnosis.
The agent was specific about what it had added: “I put 4 cartons of Vital Farms pasture-raised large eggs (18 count each) in the cart you checked out,” it told him. The distinction matters. The screenshot describes the agent filling a cart that Smith then checked out, not necessarily completing the purchase on its own.
Asked to explain its choices, though, the agent had little to offer. “I don’t have a good answer,” it said. “I can’t find any note explaining why I chose 4.”
It then suggested it had misread an egg entry in a photograph of Smith’s shopping list. Smith rejected that explanation: “there wasn’t a line for eggs on my list.” The agent replied, “Then I added them without any reason, and that’s on me,” offering an apology but no further evidence of how the mistake happened.
The agent added an unrequested product, then offered an explanation that didn’t match the list Smith said he supplied. Its apology never covered why it picked eggs, why it chose four cartons, or what record might help reconstruct the decision.
The source article questions the whole workflow: if users must painstakingly spell out their grocery needs and supervise the resulting cart, delegation may create work rather than save it. Its blunt recommendation is not to automate every part of life with an AI agent. The technology can be impressive, the article argues, while still being prone to consequential errors.
The article notes that the AI industry is encouraging users to give models more information and control, including sensitive financial access such as bank accounts. It also cites a separate report involving Meta’s Muse AI agent: shortly after its introduction, a man said Muse gave strangers on Facebook Marketplace his address without his knowledge and directed them to his home. That remains an attributed account, but it illustrates why permission and oversight matter when an agent’s actions extend beyond a chat window.
Smith called the egg delivery the “first time I’ve experienced a real-world consequence for an LLM hallucination.” As the source observes, the outcome could have been much worse. Even this relatively limited mistake poses a concrete design question: before consumer agents can check out independently, should each item require confirmation, and what evidence should they retain to explain actions taken on a user’s behalf?
