The Office Agent Is Leaving the Desk

The most important part of an office agent might be the moment it stops.

Claude Cowork is now on mobile and web, with cloud-running tasks, phone notifications, and work that can continue after your laptop closes. The Verge notes that the desktop app still has the fuller local-file experience, but the direction is obvious: agents are becoming portable background loops, not apps you babysit in a single window.

That sounds small until you picture the actual work. Pull the messy updates into a brief. Draft the follow-up. Reconcile the spreadsheet. Prepare the client prep packet. Pause when approval is needed.

This is where the coding-agent metaphor starts to break. Code agents can often be judged by tests, diffs, and build logs. Office agents operate in fuzzier territory: tone, policy, sensitive context, incomplete instructions, and a lot of “please do not send that yet.”

Microsoft is poking at the same underlying problem from a different angle with Frontier Tuning: teach agents inside the organization’s boundary, using its workflows, eval signals, access controls, and conventions. The phrase that matters is not “smarter model.” It is “the way your business does.”

For partners, the opportunity is not to make every workflow fully autonomous. That is the trap.

The better pattern is delegated work with crisp handoffs: start here, use these sources, stop at this decision, ask this person, leave this draft unsent, log what changed. The product design is the boundary design.

An agent that can run anywhere needs to know exactly where it is not allowed to go.