Call My A.I. Agent

Summary

The episode examines Muse, Meta’s AI agent, through New York Times technology reporter Eli Tan’s several-week trial. Unlike a chatbot that mainly responds to prompts, Muse can use a computer interface—clicking, typing, and navigating websites—to carry out tasks in connected accounts, sometimes working autonomously in the background.

Features and development

• Muse can connect to services and personal information such as email, calendars, payment accounts, and delivery apps. With those connections, it can perform tasks like ordering food, organizing subscriptions, booking travel, and handling phone calls.
• Meta positioned Muse as a polished, consumer-friendly assistant for everyday users. The company released it for free, aiming to attract a large user base. It was briefly the top app in the App Store.
• The discussion traces Meta’s push to build Muse partly to the influence of OpenClaw, an earlier agent that impressed developers with its ability to work independently for extended periods. OpenClaw was powerful but technically demanding, unreliable, and concerning from a security standpoint.

Tan’s hands-on review

What worked: Muse generally handled straightforward errands, including grocery orders and subscription checks. In one notable test, Tan asked it to contact his dental insurer. Muse found relevant details in his email, navigated the call process, and connected him to the right representative. During the interview, it also successfully ordered the host a lunch delivery.

What fell short: Performance was inconsistent. Muse took about 15 minutes to buy a movie ticket that Tan could purchase himself in 35 seconds. Its AI-generated podcast was dull, and its fantasy-football advice was poor, sometimes relying on outdated or irrelevant reasoning.

Privacy and broader concerns

Using Muse effectively required Tan to share highly sensitive information, including financial accounts, personal email, contact details, and information about his girlfriend. He withheld access to his work email because he feared the agent might contact sources or expose work-related information. Tan said his primary concern was not simply Meta’s access to his data, but the possibility of an agent making mistakes, acting unpredictably, or being compromised.

Tan also found himself turning to Muse for tasks he could easily do himself—including activities he valued for their human interaction or enjoyment. He described this growing reliance as unsettling. Meta’s optimistic case is that agents will handle chores and return time to users; Tan was unsure whether that time would necessarily lead to more meaningful engagement with the world.

Review note

This is an experiential product review, not a step-by-step tutorial. It assesses Muse through practical tasks and highlights both successful uses and limitations.

Main speakers and sources

• Natalie Kitroeff, host of The Daily
• Eli Tan, New York Times technology reporter and Muse tester