Claude Cowork Goes Web and Mobile, and 90 Percent of Its Use Is Not Coding

July 9, 2026. Anthropic is rolling Claude Cowork out to web and mobile, announced July 7 on the Claude blog and shipping in beta over the next several weeks. Cowork is the delegated-work product: you hand Claude a task and it works across your files, calendar, email, messaging and connected tools until the job is done. Until this week it lived only in the desktop app, so the work stopped when you stepped away from your machine. Now sessions run remotely, follow you across devices, and keep going with the laptop closed. Anthropic also published first-party usage data the same day showing that more than 90 percent of Cowork activity is not software development, a finding VentureBeat led its coverage with. For agencies and small businesses, this is the clearest signal yet that AI agents are becoming an operations layer, not a developer tool.

What Anthropic shipped

1. Cowork on web and mobile. Sessions start from the claude.ai home screen on the web or from the sidebar of the Claude app on iOS and Android. Beta access rolls out over the next several weeks, starting with Max subscribers, with more plans to follow.
2. Background sessions. Sessions and files now save to your Claude account rather than a single machine. Start a task at your desk, check on it from your phone, close the laptop and the work keeps going.
3. Scheduled tasks that run with no device online. Recurring jobs now execute server-side. Anthropic's own example: set Monday client prep for 6 am, and Claude works through the email threads, transcripts and recent news, builds the briefing document, and leaves the follow-up email drafted but unsent for review.
4. Approvals that reach your phone. When Claude hits a decision only you can make, the question comes to your device. You can redirect a draft mid-meeting and the session continues on the corrected path. Nothing ships until you review and approve it.
5. One home for chat and Cowork. On web and desktop the two surfaces merge, with projects and artifacts shared across both. Desktop remains the full experience and keeps exclusive access to local files and the browser.
6. Doubled usage limits through August 5. Anthropic extended its Cowork usage promotion to mark the launch, so teams can trial bigger delegated tasks before normal limits return.

The usage data: more than 90 percent is not coding

AI agents made their name writing code, but Anthropic's companion usage report undercuts the stereotype. More than 90 percent of Cowork usage is not software development. The two largest categories, business operations and content creation, together account for roughly half of everything people delegate. The examples Anthropic highlights will be familiar to anyone running a service business: reconciling the quarter's spend and drafting the variance memo, turning a folder of contracts into a renewals tracker with the risks flagged, and building tomorrow's client deck from call transcripts and pipeline data.

Anthropic calls this the work around the work: tasks that are rarely in anyone's job description but consume a large share of everyone's week. That matches what we see in AI automation engagements. The highest-ROI agent deployments are rarely glamorous. They are the recurring reports, the reconciliations, the briefing documents and the follow-up drafts that quietly eat billable hours.

Why this matters for the 2026 agency stack

Every major vendor is now converging on the same pattern: agents that work in the background, on schedules, behind human approval gates. OpenAI moved its ChatGPT Workspace Agents onto per-run credit pricing on July 6. Anthropic already ships Claude Tag for Slack for team-level delegation, and Microsoft's separately named Copilot Cowork went usage-metered on July 1. The competitive question has shifted. It is no longer whether your stack includes an agent. It is whether your processes, files and approval rules are structured so an agent can actually run them unattended. Cross-device access removes the last logistical excuse: the agent no longer stops when the laptop lid closes, so the bottleneck is now the operator's process discipline, not the tooling.

What it means for operators

Here is the playbook we would run this week, whether you use Cowork or an equivalent.

1. Inventory the work around the work. List every recurring task that is not client-facing craft: weekly reports, meeting prep, invoice reconciliation, CRM hygiene, content repurposing. The usage data says these are exactly where agents stick, so they are your first delegation candidates.
2. Pilot one scheduled task. Anthropic's Monday 6 am client-prep pattern is a reusable template: defined inputs, one clear output, and a drafted-but-unsent final action. Copy that shape for your own Monday pipeline review or weekly client digest and let it run before you wake up.
3. Keep approval gates on anything that leaves the building. Cowork's default, where nothing ships without sign-off, is the right one. Mirror it across your stack: human approval on outbound email, payments, publishing and every client-facing deliverable, with the agent free to act on everything internal.
4. Watch the meter. The doubled limits end August 5, and delegated work everywhere is becoming metered work, as we mapped in the July 2 cost playbook. Budget per workflow rather than per seat, and set spend alerts before the first surprising bill arrives.
5. Map capability per surface before you commit a workflow. Local files and browser automation stay desktop-only. If a process depends on a local folder or a logged-in browser session, it still needs the desktop app, so split your workflows accordingly.
6. Fix the groundwork, because structure beats model choice. Clean folder structures, documented processes and a written definition of done are what make delegation work. That preparation, not the subscription fee, is the real adoption cost.

Risks and limits

Background agents raise the stakes on governance. An agent that runs while you sleep needs least-privilege access to connected tools, an audit trail, and a human on every irreversible step, the same rules we set out for Claude's new admin spend controls. Treat the mobile approval prompt as a control surface, not a rubber stamp: if you find yourself approving without reading, tighten the task definition instead. And remember the rollout is a staged beta. Max users get access first, so agencies on Team plans should plan their pilots for the coming weeks rather than today.

If you want this operating layer without the trial and error, our AI automation agency team builds delegation workflows, approval gates and scheduled agent operations for SMBs, and if you are bringing the capability in-house you can hire an AI engineer who has shipped this before.