Anthropic has added a Cowork tab to the Claude mobile app on Android and iOS, letting users monitor AI tasks running on their desktop from their phone. The mobile integration is observational only: users can check status and approve pending actions, but cannot start new tasks from the phone itself—the agent still runs on the tethered desktop machine Engadget.

The more substantial engineering shift happened behind the scenes. Anthropic updated Cowork to execute tasks in the background without requiring a stable internet connection throughout the job. Previously, a dropped connection could stall or fail a long-running task; this update decouples task continuity from network connectivity. That architectural change is why the mobile monitoring feature is actually useful rather than merely decorative.

Permissioning remains the safeguard. When Cowork needs authorization to act—writing a file, sending an email, running a command with side effects—it sends a notification to the user's phone for approval. Anthropic's position is explicit: "Nothing ships until you've reviewed and approved it" Engadget. That approval gate keeps the background-capable agent from operating under full autonomy without human oversight.

Access is staged. Max subscribers get the mobile Cowork tab first, with rollout to other Claude plans—likely Pro and any team or enterprise tiers—planned over the coming weeks. No firm date was given for broader availability.

Anthropichas also outlined a longer-term direction: Cowork and the standard Claude chatbot will merge into a single interface, appearing first on web and desktop before moving to mobile. No timeline was specified.

Why connection independence matters more than the phone feature

The phone notification is the visible change, but the more important engineering shift is background execution that survives a lost connection. Agentic tools—Cowork, GitHub Copilot workspaces, autonomous-agent frameworks—typically assume the orchestrating session stays alive: a terminal, browser tab, or network socket. Removing that requirement means Cowork tasks can now survive a laptop sleeping, a Wi-Fi switch, or a closed lid mid-job. That's the actual foundation that makes phone-based monitoring worthwhile. Push notifications are the interface; connection-tolerant background execution is the plumbing that makes the interface non-trivial.

The approval-gate design warrants attention as a deliberate trust mechanism. Agentic AI products have struggled with the tension between usefulness—which grows with autonomy—and user comfort, which grows with oversight. A phone-push approval model is a pragmatic compromise: it keeps a human in the loop for consequential actions without requiring that human to remain at the original machine. Whether Anthropic's implementation actually flags every action that "ships," or whether some categories of low-risk action are auto-approved, was not disclosed in available reporting and remains a fair question for anyone deploying this in production.

The planned merger of Cowork and core Claude is structurally more interesting. Until now, Anthropic has kept its agentic task-execution tool separate from its conversational assistant—a distinction that echoes the broader industry uncertainty about whether "agent" and "chatbot" are different products or different modes of the same product. Merging them into a single interface, starting with web and desktop, suggests Anthropic has concluded they are different modes rather than different apps. That aligns with how Microsoft has integrated Copilot agent capabilities into its existing 365 products rather than shipping a standalone agent, and it may lower the friction cost of switching between "chatting with Claude" and "delegating a Cowork task."

Stagingthe mobile rollout behind the Max tier follows a familiar pattern: using premium subscriptions as early-access mechanisms for features still being refined at scale, rather than as permanent product segmentation. Whether the phone tab's utility justifies its subscription premium once it reaches Pro users will depend on how often people genuinely need to approve or monitor a Cowork task away from their desk—a use case likely more valuable to developers running long batch jobs than to casual users.