Anthropic's Claude Cowork became available on web and mobile for Max subscribers on July 7, 2026, expanding an agent that launched as a desktop-only app in January TechCrunch.

The expansion lets a user kick off a task from a desktop machine, check status from a phone, and retrieve the finished output later without the originating laptop staying open or connected TechCrunch. Running as a multi-platform app rather than a single-device client means Cowork's background tasks no longer depend on one machine staying online to keep executing.

The move follows Anthropic's June 2026 launch of Claude Tag, an always-on Claude instance that sits inside Slack functioning as a persistent AI teammate rather than a tool invoked on demand TechCrunch. Taken together, the two products point toward Anthropic's stated ambition for Cowork: an agentic administrative coworker operating continuously across devices, rather than a tool positioned primarily for developers.

What the usage data shows

Alongside the platform expansion, Anthropic published early usage data drawn from 1.2 million anonymized and aggregated Cowork sessions across more than 600,000 organizations, sampled over the final two weeks of May 2026 TechCrunch.

Business process operation was the single largest category, accounting for 33.4% of sessions. Anthropic's examples include consolidating scattered status updates into reports, building onboarding checklists, and reconciling spreadsheets — the kind of recurring, cross-tool clerical work that rarely justifies a dedicated headcount but consumes disproportionate time for the people who inherit it TechCrunch.

Content creation and copywriting came second at 16.4%, covering drafts, slide decks, social posts, and proposals. Software development, the category around which the coding-agent market has largely been built and marketed over the past two years, accounted for only 8.7% of sampled sessions TechCrunch.

That ordering is notable given where the underlying agent technology originated. Cowork descends from the same coding-agent lineage — autonomous, multi-step task execution with tool use and file access — that Anthropic, OpenAI, and others have spent the better part of two years refining for engineering workflows. The May sample suggests that lineage of capability is finding its largest audience somewhere else entirely: among operations, marketing, and administrative staff who never touch a terminal.

The broader context here is that the "coding agent" framing, which has dominated agentic AI coverage since 2024, may have understated how portable the underlying architecture is. An agent built to plan a multi-file refactor, run tests, and iterate on failures turns out to be structurally well-suited to pulling five people's status updates into one document, or reconciling a expense spreadsheet against receipts — tasks that share the same shape of "read scattered inputs, apply judgment, produce a structured output" even though the domain is entirely different.

It is worth flagging that Anthropic's data reflects self-reported category tagging on its own platform during a single two-week window, shortly after the January desktop launch and well before this week's cross-device expansion. Usage patterns during a product's early adopter phase, concentrated among a subset of Max subscribers willing to install a new desktop app, may not hold once the tool is reachable from any browser or phone and used more casually by staff outside early-adopter functions.

Anthropic's positioning of Cowork against Claude Tag also signals a bifurcation in how the company sees agentic work getting done: Tag as a persistent, ambient presence inside an existing collaboration surface like Slack, and Cowork as a more directed, task-initiated agent that a user explicitly hands work to and later collects results from. Competitors including OpenAI, Google, and a widening field of coding-agent startups have been racing to extend similar capabilities beyond software engineering, and the usage split Anthropic has now disclosed gives those rivals a concrete data point to react to, whether or not their own telemetry looks the same.

For technology leaders evaluating where to deploy agentic tools inside their own organizations, the practical takeaway is less about Cowork specifically and more about where the return on agent deployment is actually showing up. If the largest share of session volume on a coding-lineage product is going toward business process and content work rather than software development, that argues for treating agent procurement as an operations and knowledge-work decision as much as an engineering-tooling one — a reallocation of budget conversation that many IT organizations have not yet had.