Meta's risky new AI agent

Meta launched Muse, an AI agent that requires access to health data, credit cards, and personal emails. The company has given people reasonable reasons to worry about such access over the last twenty years.

What Muse actually does

Unlike ChatGPT, Muse is designed for practical tasks. Users ask it to do things like organize inboxes, find products, or manage subscriptions instead of generating text or replacing Google searches.

Early wins with email and money

Muse successfully created a daily digest of important emails from a cluttered inbox and identified subscription services to cancel, including Flickr, which the user decided to keep anyway.

Birthday party planning fails

When asked to plan a child's birthday at a trampoline park, Muse failed because the task required too many human decisions and real phone calls to the venue to complete successfully.

School app chaos defeats Muse

Muse struggled to track school and activity notifications across multiple apps. It couldn't access Google login options on ParentSquare or view WhatsApp messages where Girl Scouts planning happened.

Girl Scouts website too much

Muse almost completed the Girl Scouts annual membership renewal but couldn't finish the final payment section because the website timed out, forcing the user to complete it manually.

Health data access, limited help

After connecting Apple Health data, Muse told the user their resting heart rate was fine but suggested better sleep. It failed to book an annual physical and couldn't sync menstrual cycle data with Walmart orders.

Muse's cautious approach

Meta spokesperson Jimmy Raimo said some friction is intentional. Muse stops when fees look high or logins seem off rather than pushing through, because the company would rather it ask than get things wrong.

“If fees look high or a login looks off, Muse is designed to stop and check rather than push through. We'd rather it ask than get it wrong.”

The real challenge with Muse

The biggest hurdle is knowing what tasks to give Muse. It requires knowledge of what AI agents can do and creativity to brainstorm suitable tasks, which many users may lack.

“80% of the battle here isn't how well Muse can execute a task; it's how well you can come up with a task for Muse to do.”

Promise without delivery yet

While Muse can accomplish some genuinely helpful tasks, like disputing health bills, it cannot yet handle the annoying everyday work that users most want to avoid doing themselves.