Meta's Muse AI Hit a Captcha. It Asked a Human for Help. The Agentic Web Has a Scaffolding Problem.

A journalist testing Meta's Muse AI app asked it to book a flight using stored airline login credentials. Muse navigated to a captcha page and paused, asking the user whether it should attempt to prove it was human or hand control back. The user had already found the same flight manually on Google Flights. Muse stores credentials in a vault but cannot independently pass bot-detection challenges.

This illustrates what I call the agentic handoff boundary. An AI agent can plan, navigate, and use credentials, but it cannot cross authentication barriers designed to exclude non-humans. The practical lesson is that autonomous AI tools work until they meet infrastructure built to refuse them. At that boundary, the system must either escalate to the human or fail. Muse chose escalation, which is the correct design. But it means these tools remain collaborative, not autonomous.

Meta, under Mark Zuckerberg, built the Muse AI app. A Slate journalist tested it by booking a flight and found Muse could not independently pass a captcha.