Meta Muse virtual machine per U.S. user: 2 cores / 8GB RAM / 8GB storage (Source: Stratechery (September 28, 2026))
Two vendors, one design choice
In a September 28 analysis, Stratechery points to a shared design choice in two recent launches: the agent gets its own computer. Meta is provisioning every U.S. user of its Muse product with a virtual machine, which Stratechery calls the only way to make agents work for most people. Microsoft's Autopilot, an always-running enterprise agent inside Copilot, is described by Microsoft VP Jared Spataro as living "in your tenant with its own identity, memory, computer, and workspace."
The Microsoft version is the one an enterprise buyer will recognise. Per the material Stratechery quotes, Autopilot can watch Teams channels, run recurring tasks and handle follow-ups while its user is asleep. It can be @mentioned like a colleague, with "permissions, audit, and governance behind it." Those governance claims are Microsoft's own. The excerpt does not describe how audit or permissioning works in practice.
The supplier problem, as Stratechery frames it
Stratechery's argument is that doing things on the web or in apps is becoming abundant, and that what stays scarce is what it calls volition, meaning the user's decision about what to get done. In that framing, the apps and services an agent uses become "implementation details." Their owners become suppliers to whoever operates the agent, much as publications became suppliers to search and social aggregators.
The author's evidence is personal. He says agents could already operate any app with a command line. Since GPT-5.6 Sol they can operate any app with a user interface, "albeit slowly," and Astra made that fast. He asked Muse to organise his saved Instagram recipes and received a working app of his own. This is one writer's account, not a measured adoption trend. It is still a clear statement of the risk for any organisation whose revenue depends on customers opening its interface.
What the source does not establish
The analysis offers no adoption data, no security findings and no enterprise deployment figures. Microsoft's Code tab, which lets anyone create an app, tracker, dashboard or automation and share it as a cloud-hosted internal app, is a product description. Nothing in the excerpt measures how organisations use it. Both launches are single announcements from two companies. They show where these vendors are heading, not a market-wide pattern.
For a budget-holder, the useful reading is narrower. An agent with its own identity, memory and machine is a new kind of actor in the environment. It is not simply another application, and it needs an owner, a permission model and an audit trail before it has one by default.
Questions to put to your team
First: if Autopilot or a comparable agent is enabled in our tenant, which identity does it act under, and who approves what it can read and change? Second: when staff use the Code tab to create internal apps and automations, who reviews them, and where are they inventoried? Third: which of our customer journeys assume a person opening our app or site, and what do we know about traffic that arrives through an agent instead? Fourth: what do we contractually require from a vendor whose agent runs on a virtual machine holding our employees' data?
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Stratechery.
