Claude Cowork Goes Mobile

Anthropic moved Claude Cowork off the desktop this week. It now runs on the web and on phones, in beta for paid users, and the part that actually matters is what happens when you leave: a Cowork session keeps running in the cloud after you close the laptop and pocket the phone. You hand it an outcome, walk into a meeting, and come back to a draft waiting for review.

That’s the launch. Most of the coverage filed it under the coding-agent wars, another front in the fight between Claude Code and every rival trying to automate the developer. Anthropic’s own numbers say that framing misses what the tool is for.

Alongside the rollout, Anthropic published what it saw across 1.2 million anonymized Cowork sessions from more than 600,000 organizations, sampled over three weeks in May. More than 90% of that work had nothing to do with software development. Coding came in at 8.7%. The largest category, at 33.4%, was business process and operations: pulling scattered updates into a single report, building onboarding checklists, reconciling spreadsheets. Content creation and copywriting was second at 16.4%. Anthropic has a name for it, “the work around the work,” the connective tasks that show up in nearly every job and sit in no one’s job description.

Source: Anthropic

If you’ve run an operation on Claude for any stretch of time, none of that lands as a surprise. Coding demos get the airtime because they’re legible; you can screenshot a diff and it looks like magic. The actual daily load is the glue work: the status update nobody wants to write, the five spreadsheets that have to be reconciled before a real decision can happen, the recurring report that eats a morning. That’s what quietly consumes an operator’s week, and it’s exactly where Cowork’s users pointed the thing.

So the mobile launch isn’t a coding app getting a phone. Read the usage split and what Anthropic built is obvious: an agent for the operations layer of a business, now detached from any single machine. This is the same company that’s been shipping a run of releases aimed at people who don’t write code, like its tool for visual prototyping and presentations, and reframing the “AI is coming for your job” anxiety as AI taking the work nobody wanted anyway. The usage data is the evidence for that pitch. It’s also, conveniently, the product that captures it.

Untethering an agent from the desktop is a bigger shift than an app-store listing suggests, and the phone is the least interesting piece. The real change is that a Cowork task now runs unattended in the cloud, on whatever cadence you set, whether or not any of your devices are even powered on. You can kick something off from your phone at midnight and let it grind while you sleep.

An agent you can’t see is a different trust problem than one running in a window in front of you, and this is where Anthropic’s design choice carries the load. Cowork asks before it acts. With permissions set, it stops at the consequential decision points and waits for a human; you can redirect it or shut it down at any point. That gate is the load-bearing part of the entire product.

Anyone who has run unsupervised agents knows the failure mode already: the model reaches the hard part, decides it isn’t important right now, quietly works around it, and keeps building, and you find the hole three steps later, load-bearing and buried. Move that loop into the cloud, out of your sight, running overnight, and the permission gate stops being a nice-to-have. It becomes the line between delegation and abdication.

For the enterprise buyer Anthropic is plainly courting here, or the kind of institution that recently handed its entire state government over to Claude, that gate is the compliance story: admin controls, spend limits, activity streaming to a SIEM. For a solo operator it’s simpler and more personal. It’s whether you can trust the thing to run your operations unattended, or whether you’ll end up re-checking everything it produced, which erases the entire reason you handed it the work.

The easy mistake is to over-index on the phone. Most people will download it, run two tasks, and forget it exists. The operators who get compounding value are the ones who take the “work around the work” finding at face value and hand Cowork the recurring operational glue: the Monday report, the inbox triage, the deck that gets rebuilt every week from the same five inputs. This is the connective coordination work that enterprise AI keeps stalling on, and it’s exactly what an unattended agent is built to absorb. Set the permission gate tight. Then actually let it run, and resist the urge to hover.

That’s the move the data is pointing at. Not a coding agent that happens to have a phone app. An operations agent that finally lives where the operations already do, as long as you keep a hand on the switch.