For the last few years, Meta struggled to put forward a convincing strategy, and I found its glasses and VR headsets unappealing. Judging by this year's Connect conference, it feels like Meta is finding its groove. Muse is a great idea. It feels like the missing piece that suddenly makes Meta's mission and product line-up, from social media apps to AI glasses, make sense.
AI power users who've been running personal agents like OpenClaw or Hermes will know how compelling it is to let an agent get on with work on a computer. Chatting with AI models was useful, even breathtaking, when ChatGPT became mainstream. But AI’s killer feature is its ability to actually do things for you.
Muse brings that experience to the masses. You don’t need to keep your own computer running all the time or give an agent access to everything on it. You simply sign up and start chatting with your new personal assistant, connecting the accounts and sharing the data it needs. Under the hood, Muse gives each user a Linux virtual machine (VM) in the cloud, complete with storage, a browser and a terminal for running commands and code. Access is initially limited to the US. Muse can read and send emails, make bookings, browse the web and complete purchases, with approval required for sensitive actions. It can take on much of the mundane work you'd otherwise have to do yourself.
Meta has also drawn on its experience building apps used by billions of people to produce an interface that looks fun and easy to use. Muse feels like a milestone to me: this is the type of app that could truly democratise AI. Not surprisingly, it has already topped the US App Store charts.
I think Muse could make personal agents running on their own virtual machines mainstream. What about in the workplace?
For me, the appeal of assistants like Muse is that they handle much of the complexity of using software. They launch sub-agents, call tools, operate a browser and read or write files in the background. The user doesn’t need to worry about whether a task calls for a terminal command, a script or an agent interpreting what's on screen and clicking through an app.
This allows the user to focus on tasks and projects, set goals, review the output, iterate and experiment, while the assistant carries on with the work. This is the experience we all want AI to deliver at work.
Contrast this with the traditional Microsoft 365 experience, which still revolves around using a wide catalogue of different web and desktop apps and automating workflows via clunky low-code or no-code tools like Power Automate. Trying to deliver on the promises of AI in this environment often proves messy and frustrating. Take Copilot agents, for example. Users need to identify a use case, create the agent, navigate roadblocks that require an IT admin's help or approval, share it with others, and remember to use it when the right circumstances arise.
So far, adding Copilot to every corner of Microsoft 365 hasn't done much to reduce that friction for me. In my testing, Copilot in Power Automate failed at most attempts to create fairly simple workflows. The irony is that tools which felt empowering five years ago now feel completely inadequate in the age of AI. Copilot can produce a Python script for the same task, yet struggle to build the Power Automate flow. The user is still left assembling the workflow and dealing with the boundaries between apps.
For organisations in the Microsoft ecosystem, Windows 365 for Agents looks like a much more promising direction. Windows 365 is Microsoft’s product that allows customers to run Windows virtual machines in the cloud. Windows 365 for Agents makes those VMs available to agents, where they can operate applications, run code and use tools, while users can intervene when needed.
I think this offers a great foundation for the same experience at work that Muse promises consumers. An assistant could operate semi-autonomously in a secure, managed environment, taking much of the complexity of using traditional software out of the user’s hands. Although the computing environment is there, Microsoft would have to adapt the product to turn it into a personal assistant for the workplace. Microsoft currently describes agents checking out a Cloud PC from a shared pool for a task and returning it afterwards, rather than each employee having their own permanently assigned virtual machine. Microsoft would also need to rework the Copilot interface to make it the main hub for each employee’s work.
Indeed, in this world, the assistant becomes the main interface for work, taking over much of the role Outlook and Teams play today. We’d access our work assistants on a laptop, tablet or phone, or perhaps even, in the future, through a small device like the Muse Charm announced at Connect. Spreadsheets, PowerPoint presentations and platforms like Workday and Concur would still exist, but agents would operate them in the background. Users would get stuff done through back-and-forth conversations with a digital assistant, much like advanced users of ChatGPT or Claude do today.
I’m confident that this way of using AI will become the norm at work as well. However, I see three major roadblocks: access permissions and security, shared standards for data and reporting, and running costs.
First, organisations need to define the right access rights for users and the agents acting on their behalf. Even in a managed environment, getting permissions right is complicated. They often reflect internal organisational structures as much as practical needs. Employees need enough freedom to build workflows that run from end to end, while agents need clear limits on what they can read, change and share, and when they must ask for approval. Lock everything down and the assistant becomes frustrating to use. Give it too much access and a mistake can expose data or make changes the user never intended. Finding that balance will take time, especially where it requires a change in company culture.
Second, organisations need common definitions and standards for how data is used and reported. If every employee builds workflows with their own assistant, those assistants need consistent instructions about what the data means, which sources to use and how metrics should be calculated. Otherwise, we risk making it much easier to produce conflicting reports. Giving agents access to data won't resolve disagreements about what counts as the right answer.
Finally, there is the cost. A computer-using agent working through a task onscreen can require more model calls and tokens than a specialised agent completing the same task through a direct connection to an application. The Cloud PC adds computing costs as well. The difference will depend on the task and how each agent is built, but it could quickly strain budgets as usage spreads across an organisation. Here, again, the viability of a fully AI-enabled workplace depends on powerful models being cheap and widely available enough to run everywhere, all the time.