Leadership1 publisher2 min readPublished
A school portal login and a timed-out payment page stopped Meta's Muse agent
Business Insider's reviewer gave Meta's new agent his inbox, his credit cards and his Apple Health data. Two of the six tasks finished as asked, and the failures landed on other companies' sign-in and checkout screens.
The Board Room · Leadership desk
What happened
- Meta's Muse agent launched earlier this month, and a Business Insider reviewer connected it to his email and calendar with read/write access, his credit and debit cards, and his Apple Health data.
- It handled two straightforward jobs well, producing a daily digest of important emails and combing his card statements to suggest he cancel Flickr.
- A band signup form sat inside a ParentSquare link, and the reviewer's ParentSquare login runs through his Google account, an option Muse's browser version of the site did not include.
- Muse got the Girl Scouts membership renewal as far as the final payment section but could not complete it before the site timed out, so the reviewer finished the job on his own computer.
- The login problem was only resolved after a Meta PR contact conferred with the engineering team and came back with troubleshooting steps for him.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- constraint A workflow that terminates on a third-party sign-in or checkout page stops there, and a better model will not get it past those screens.
- exposure The access is granted up front and in full, covering mail, payment cards and health records, against a return that depends on sites the grantor does not control.
- precedent When vendor engineers clear the blockers during an evaluation, the completion rate a buyer takes to a board includes their help, and getting that rate at scale would take the same staffing.
- contradiction The reviewer puts the birthday-party failure on human decision-making, so at least one of his failures cuts against the reading that institutional apps are the binding limit.
The obstacles in the household tests were access boundaries. Muse could not read the busy WhatsApp group where the Girl Scouts planning happens, so that whole workflow was invisible to it [13]. The other two blocked tasks stopped at authentication and at checkout [10][14].
Six tasks appear in the piece, and two finished as asked [17]. Both of those ran on data Muse reached directly, his mailbox and his card statements [4]. The trampoline park birthday he scored himself: "Failure, but mostly human failure", because the decisions belonged to the birthday boy and the venue needed a real phone call [7]. Meta has since updated Muse to make calls [8]. The resting heart rate task he summarised as "Sure, fine" [16].
He wrote that the ParentSquare fix worked, "although that kind of concierge service isn't available to most users" [12]. Enterprise pilots run the same way. Vendor engineers sit on the evaluation call, and their help is baked into the pass rate a buyer carries into a review.
A trampoline party is not a procurement workflow, and the reviewer says as much about his own test [7]. This is a first-person consumer review of one household [19]. What does transfer is the shape of the blocker: single sign-on delegated to a third-party identity provider, and a checkout page that fails on a slow client, are ordinary features of the software institutions actually run.
For anyone scoping agents this quarter, the practical line runs between workflows that end inside systems whose credentials you control and workflows that end on someone else's sign-in or payment page. Muse's two clean results came from the first kind [17].
What to watch
- Whether Meta's calling update changes outcomes on tasks that needed a phone call to a venue.
- Whether Muse gains read access to WhatsApp threads, which would close one of the three access gaps in this test.
- Whether any agent vendor publishes completion rates on third-party sites measured without engineer assistance.