Claude Cowork Leaves the Desktop — and the Data Shows Most Users Aren't Coding

What actually shipped

Starting this week, Claude Cowork — Anthropic's agent for handing off open-ended work tasks — is available on web and mobile, not just the desktop app it launched on. Access is rolling out first to Max subscribers, with other paid plans following over the coming weeks via the Claude app on iPhone, iPad, and Android, and at claude.ai on the web.

The practical change is bigger than "another platform." Cowork tasks can now run fully in Anthropic's cloud instead of depending on a laptop staying open and connected. That flips mobile from a launcher into a review surface: you kick off a task on desktop, get status pings on your phone, and approve or reject the output before anything ships — Anthropic is explicit that nothing goes out without user review.

The usage data is the real story

Alongside the rollout, Anthropic published a breakdown of 1.2 million anonymized Cowork sessions from more than 600,000 organizations, sampled between May 11 and May 31. The split undercuts the assumption that agentic AI adoption is primarily a developer phenomenon:

• Business process operations — 33.4% (reports, checklists, spreadsheet reconciliation)
• Content creation and copywriting — 16.4% (drafts, decks, social posts, proposals)
• Software development — 8.7%

For a company built on Claude Code's developer traction, that's a notable signal: the largest agentic workload inside organizations using Cowork is back-office ops, not code. Anthropic is also extending doubled Cowork usage limits through August 5 — a straightforward push to lock in habits during the mobile rollout.

Cowork is its own vibecoding case study

Cowork's origin is arguably the more interesting data point for builders. It launched as a research preview on macOS on January 12, reportedly built in about ten days by an Anthropic engineer using Claude Code itself, then reached general availability across paid plans in April. Anthropic frames it plainly: "Cowork is where you hand Claude a task, and it works across your files, calendar, email, messaging app, the web, and the other tools you connect until the job is done." That's the same permission-scoped, tool-connected agent loop dev teams are now wiring into internal products — just aimed at spreadsheets and inboxes instead of pull requests.

Why builders should pay attention

Two patterns here are worth stealing for anyone shipping agentic features: first, decouple execution from the device — cloud-native background tasks with a mobile approval step are becoming the default UX for agent products, not a nice-to-have. Second, the access model matters as much as the model itself — Cowork's value comes from scoped, revocable connections to files, calendar, and email, with an explicit human-approval gate before anything leaves the sandbox. That's the harder engineering problem, and it's the one competitors chasing the "agent for knowledge work" category will actually be judged on.