Meta's Muse, reviewed: the consumer agent that asks before it acts

Muse is Meta's personal AI agent, and the first thing worth noticing is who it is for. The landing page is pitched at families — permission slips and strollers rather than pull requests. Claire Vo gave it her calendar, her email and her children's activities, and came back saying it is the best-designed personal agent she has tested 1.

Two things separate Muse from the coding-assistant lineage, and neither is capability.

The first is the permission model. Vo's comparison is the moment an agent gets access to email: her previous setup announced that it had the inbox and would go and read it, while Muse asked whether it could go and find out information about her. Permissions arrive when the work needs them, phrased as a question rather than a notification, which is most of the difference between an assistant and a breach. The same instinct shows up in the plumbing: a persistent isolated Linux virtual machine with its own browser for the agent to work in, stored credentials, a promised 1Password integration, and Stripe Link presented as the payment method.

The second is the activity feed, which records every task with the steps and tool calls behind it, alongside the approvals the user has granted. Watching a task unroll step by step, Vo said she wanted the same lineage in Codex, Claude Code and Grok Bot.

Around those sit three primitives. Goals are framed as things a person cares about — health, relationships, money, career — rather than as a task list, and each one produces a plan, a reminder and a tracking view. The library holds everything the agent has produced, including documents, websites, images, videos and podcasts; asked for one, Muse produced a six-minute audio programme that Vo could publish. An ideas panel offers starting points drawn from what the agent has inferred about the user.

The most useful part of the review is the worked example, because it shows what good output looks like when the agent has been given real access.

Vo asked for a one-page agenda she could print and put on the kitchen table: the week's schedule, highlights for each child, callouts for the parents, and something worth reading. Muse connected a Google Calendar, removed the soccer practice the middle child had quit, and then, given access to email, inferred the children's names, her husband's name, the activities they do — jujitsu, basketball, piano — that the family is civically engaged in San Francisco, that both parents work in tech, and that they order a lot of food. It listed what it had inferred and asked her to confirm before building.

The result was a single-page PDF with the schedule, a section of local news, and conversation prompts. Vo says she had not got anything that good in one shot out of Claude or Codex.

Shopping is the weak spot, and it fails in a way that matters for anyone building browser use. Asked to find a specific pair of New Balance 9060s in the right colourway, Muse could not locate the product, and separately produced a cinema listing that did not match the film she had asked about. Two of those errors are the kind that erodes trust in everything else the agent says.

Browser use is not uniformly broken. Buying a ticket for a three-hour film went through the cinema's site, reached guest checkout, and stopped where it should have, asking for a Stripe Link card. The agent could drive a checkout. It could not yet do the product research that makes the purchase worth delegating.

Muse is roughly as capable as the agents developers already use, and considerably more legible, and the legibility is doing the work. Permissions asked in context, a running record of what the agent did and why, goals framed as things a person cares about rather than as a queue, and output shaped so it can be handed to someone else at the table.

Vo's own summary of the gap between Muse and the agent she had been using is a product observation rather than a technical one: consumers want beautiful things, they want disclosures delivered progressively, and they want to be asked. Agents built by engineers for engineers have been shipping the same permissions as a settings screen for two years. Muse is the first consumer-scale attempt to show what that looks like when it is treated as interface design instead.