Claude Cowork goes mobile, and most of its work isn't code

On July 7, the official Claude account announced that Cowork, Anthropic's task-delegation agent, is coming to mobile and web. @claudeai framed it in one image: hand Claude a task at your desk, then pick up the finished work from your phone. The line that stuck with everyone: "Close the laptop and Claude keeps going." A companion post from @claudeai added that Chat and Cowork are merging into "one home on web and desktop," so delegating a task starts the same way a conversation does.

If you have not followed Cowork: it launched on desktop in January as Anthropic's agent for non-coders. You connect your files, calendar, email, and web tools, describe an outcome, and it works the task until it is done. Until this week, it lived on your computer. Now the beta is rolling out to web, iPhone, iPad, and Android over the next several weeks, starting with the Max plan, and scheduled tasks can run even when no device of yours is online.

The number buried in the announcement

The rollout is not the real story. The usage data is. Anthropic's own announcement says that over 90 percent of what people delegate to Cowork is not software development. The two biggest categories, business operations and content creation, account for roughly half of all usage. The examples Anthropic highlights are almost aggressively unglamorous: reconciling quarterly spending, turning a folder of contracts into a renewals tracker with flagged risks, building client presentations out of call transcripts, and a scheduled agent that reads overnight email and news to have a client briefing ready by 6 am.

AI commentator @rohanpaul_ai called the move a strategic inflection for Anthropic, arguing it connects the company's coding-agent technology to "the much larger knowledge-worker market that never opens a terminal." That framing matches the numbers. Anthropic built its reputation on Claude Code and developers, and its own telemetry is now telling it that the bigger market never wanted code at all.

My honest read

Three things stand out to me.

First, the agent race has officially left the terminal. OpenAI pointed its Codex-powered agent at office work with ChatGPT Work two days later, and Anthropic is doing the same motion from the other direction. The technology that was sold as a programmer's tool in 2025 is being repackaged as a delegation tool for everyone else, because that is where the demand showed up.

Second, the phone is becoming the review surface, not the workplace. The design here is delegation with checkpoints: you assign the task at your desk, Claude pings your phone when it hits a decision point, you approve or redirect, and you collect the output later. That approval loop is the actual product. Nobody serious is shipping agents that run unsupervised end to end, and that should tell you something about how to deploy them in your own company.

Third, the caveats are real. This is a beta, it starts on the expensive Max tier, and the desktop app is still required for local files, browser use, and computer control. And a scheduled agent that works your email and documents while no device is online is a cloud service with standing access to your business data. That is fine, but it deserves the same permission review you would give a new employee, not a casual toggle.

What this means if you run a business

The 90 percent figure is the most useful market research a small business will get this year, and Anthropic paid for it. When hundreds of thousands of organizations get a frontier agent, they do not ask it for software. They hand it the boring middle of the week: pulling scattered updates into one report, reconciling spreadsheets, drafting the follow-ups nobody gets to. That is exactly the work most Fox Valley businesses are still doing by hand between customer calls.

You do not need a Max subscription to act on that. You need a list of the repeatable processes in your business that look like Anthropic's examples, ranked by hours lost. That list is what we build in New Face Design's free process audit, and it is usually short, specific, and worth more than any tool announcement. The agents are finally leaving the terminal. The businesses that win will be the ones that know which tasks to hand them first.