Claude Cowork goes mobile: what background AI agents mean for your SMB

On July 7, 2026, Anthropic announced the extension of Claude Cowork — its agent capable of executing tasks autonomously — to the web and to iOS and Android apps. The beta opens first to Max-plan subscribers, with a gradual rollout to other plans expected over the coming weeks, on no announced schedule. Launched on January 12, 2026 as a research preview on the macOS desktop app, then made available to every paid plan in April 2026, Cowork takes a further step: a task started on a computer can now keep running in the background, even with no device online, and be tracked, resumed or approved from a phone.

What actually changes

• →Background execution with no device connected: the agent keeps working on a task even after the app is closed or the computer is off.
• →Scheduled tasks: you can ask Claude to run a job at a set time, with no manual trigger.
• →Mobile tracking and approval: the user checks progress, resumes a session, or approves a step from their phone.
• →The desktop app remains necessary for local files, local connectors and computer control — web and mobile cover the rest of the knowledge work.

A data point more telling than the announcement itself

To back up the launch, Anthropic published an analysis of 1.2 million anonymized Cowork sessions, run between May 11 and May 31, 2026 across more than 600,000 organizations. Software development accounted for only 8.7% of recorded usage. Business operations came first at 33.4%, followed by content creation at 16.4%. In other words: the autonomous AI agent is no longer a developer tool — it is already used mostly for administrative and operational work, exactly the kind of task that fills an SMB's day. This shift confirms a trend already under way in May 2026, when Anthropic launched Claude for Small Business, an offering aimed at small companies with connectors to QuickBooks, HubSpot, DocuSign, Canva and PayPal.

What it changes for your SMB

Until now, an AI agent working for you stayed, in practice, in plain sight: an open window on a computer, a session watched live. With background execution and remote approval, the agent becomes a collaborator that acts while you are elsewhere — and reaches you by notification exactly when a decision is needed. That shifts the central question: it is no longer « should we adopt an AI agent » but « how do we supervise one that works while we are not watching it ». Three things then become non-negotiable: the agent's access scope (which accounts, which folders, which connectors it can touch), a traceable log of every action it takes, and the exact point where human approval is mandatory before an action has a real effect — sending an email, approving a payment, editing a contractual document.

This is precisely the logic we apply to the agents we deploy for clients — a regulatory-watch agent monitoring external sources, an operations watchdog tracking indicators continuously, or an email-sorting agent that pre-classifies an inbox before human intervention: access scoped to the task, a log of every action, and a clear line between what the agent decides alone and what it submits for approval.

Before adopting this kind of agent, four checks

• →List every account and connector the agent would access, and strip out anything beyond the task at hand — a reporting agent does not need write access to your CRM.
• →Write down, explicitly, which actions the agent can execute alone and which require human approval before any real-world effect (payment, external send, contractual edit).
• →Require a reviewable log of every action taken in the background — without a history, mobile supervision is just one more notification.
• →Start with a single, narrowly scoped recurring task before widening the agent's scope — the beta, still limited to Max subscribers, is a reasonable window to test without rushing.

Agentic AI that works while you sleep is no longer a conference promise — it is a feature you can turn on this week. The question that decides whether it serves you or exposes you is the same one you'd ask about a new hire given a set of keys: what scope, what traceability, what approval. Settle it before you flip it on, not after.