# Cowork’s Cloud Move Changes Agent Control - Date: 06 Oct 2026 (2026-10-06T16:03:24.000Z) - Summary: Anthropic’s October 6 move puts new Pro and Max Cowork tasks in the cloud, allowing work to continue after a laptop closes while local access still depends on the desktop app. The shift makes clear permissions, task status, and visible controls more important for persistent agents. - Tags: `digest`, `ai-discourse`, `claude-cowork`, `cloud-agents`, `agent-ux`, `permissions`, `human-control` ## Sources 1. [Anthropic - Use Claude Cowork on web, desktop, and mobile](https://support.claude.com/en/articles/15520349-use-claude-cowork-on-web-desktop-and-mobile#h_f951c27c48) (website) 2. [Simon Willison - A quote from Felix Rieseberg](https://simonwillison.net/2026/Oct/5/felix-rieseberg/) (website) 3. [Department of Product - The new UX of AI Agents](https://departmentofproduct.substack.com/p/the-new-ux-of-ai-agents) (website) ## Executive Summary Closing a laptop no longer necessarily ends an agent’s work—and that changes what users need to understand about delegation. Anthropic’s October 6 change moves new Cowork tasks on Pro and Max plans into the cloud, separating continued execution from an open desktop. The consequential shift is persistence, not a demonstrated jump in model capability: work can keep running while the person who initiated it is elsewhere. The accompanying product-design question is how to make that continuing activity understandable and controllable. Rich Holmes’s new agent-UX survey emphasizes visible memory, meaningful approvals, background work, and audit trails. Together, these developments sharpen a practical distinction: an agent that remains available is not necessarily an agent that retains access to everything it needs. ## What Actually Changed [Anthropic’s Cowork documentation](https://support.claude.com/en/articles/15520349-use-claude-cowork-on-web-desktop-and-mobile#h_f951c27c48) specifies that new Pro and Max tasks run in the cloud from October 6, with the “Only on your computer” setting removed. Existing local tasks stay local until completed. Cloud sessions can continue after the laptop closes, scheduled tasks can run without a device online, and users can resume sessions across desktop, web, and mobile. [Felix Rieseberg’s explanation, quoted by Simon Willison on October 5](https://simonwillison.net/2026/Oct/5/felix-rieseberg/), makes the architectural change explicit. Previously, inference ran in the cloud while tool calls executed inside an Anthropic-provided virtual machine on the user’s computer. Now both inference and that VM run in the cloud, with a separate sandbox for each session. Rieseberg attributes the move partly to complaints about disk, battery, performance, and work stopping when a laptop closed. Those are the builder’s account of the motivation, not independently measured improvements. The important qualification is **execution continuity does not imply access continuity**. Local files remain on the user’s computer. Cowork reaches connected folders through the desktop app while it is open; when a cloud task needs a file, it fetches a copy of that file. Anthropic says closing the app leaves the cloud session running but unable to reach local files. Local connectors, browser automation, and computer use also retain desktop dependencies. That makes “works while you’re away” a conditional promise. A cloud-only task and a task that needs another local file have different operational boundaries, even if they appear in the same interface. ## The UX Has to Explain the Boundary In [“The new UX of AI Agents”](https://departmentofproduct.substack.com/p/the-new-ux-of-ai-agents), Holmes organizes emerging interfaces around seven patterns. The publicly accessible introduction identifies editable memory, approvals before action, working directly on a canvas, manual output editing, background agents, visibility and control, and agents embedded where people already work. The detailed examples are subscriber-only; this is a design survey, not evidence that these patterns have solved trust or reliability. Its most useful contribution is treating controls as part of the product rather than an appendix to the chatbox. Persistent work makes that argument concrete. Users need to distinguish a task that is progressing from one waiting for permission or missing access. They also need to know which remembered context influenced the work and what actions have already happened. Those are design implications of the cloud move, not claims that Cowork currently provides every such control. A session sandbox describes an isolation boundary; it does not, by itself, establish that the task is correct or that its actions match the user’s intent. ## The Bigger Story Yesterday’s performance case reinforced bounded delegation: agents could search widely while humans retained authority over measurements, complexity, and rollout. Today complicates that view without overturning it. Delegation can outlast the user’s presence, so its boundaries must remain intelligible when nobody is watching the task unfold. The strongest signal is therefore a change in the execution model, not proof of greater autonomy or reliability. Cloud persistence removes one interruption—the sleeping laptop—but exposes the need to communicate permissions, dependencies, and unfinished work more clearly. The next useful standard for background agents is not merely whether they keep running. It is whether users can tell what continued, what stopped, and why. ## Further Reading - [Rich Holmes: “The new UX of AI Agents”](https://departmentofproduct.substack.com/p/the-new-ux-of-ai-agents) — a product-design framework for persistent agents; the introduction is public, with detailed examples behind a subscription.