Updated
Updated · The New Yorker · Sep 16
Meta's Muse AI Agent Stumbles on Tasks, Selling 4 eBay Items in a Weekend Test
Updated
Updated · The New Yorker · Sep 16

Meta's Muse AI Agent Stumbles on Tasks, Selling 4 eBay Items in a Weekend Test

3 articles · Updated · The New Yorker · Sep 16

Summary

  • A weekend test of Meta’s new Muse agent found it could complete some chores—most notably moving listings that helped sell 4 eBay items—but it repeatedly failed at simpler requests like booking a restaurant reservation.
  • Muse’s usefulness depended on broad account access, asking to connect services such as OpenTable, calendars, cards and cloud storage; even with limited access, it hit login problems, bot checks and weak outputs like generic product titles.
  • The agent still produced a monthly budget within minutes and flagged a suspicious double charge, showing where automation worked better than consumer-facing tasks blocked by verification systems.
  • Meta markets Muse as a personal AI agent with little learning curve, but the test underscored privacy concerns because the product is designed to ingest emails, banking and other personal data despite the company’s long record of data controversies.
  • The broader takeaway was less about one app’s glitches than about AI agents pushing more daily judgment and labor into systems controlled by big tech, even when the time savings feel modest.

Insights

Are you willing to hand over your bank and email passwords to Meta just to save a few minutes on daily chores?
Why do highly advanced AI agents still fail at simple web logins, and will the internet fundamentally change to accommodate them?