- Published on
Quoting Felix Rieseberg
- Authors
- Name
- DeepHorus
Anthropic has unveiled a rearchitected version of its Cowork agentic product, moving both model inference and the virtual machine runtime into the cloud. In the previous iteration, model inference ran in Anthropic's cloud while tool calls executed inside a VM the company shipped directly to each user's computer. The company said it originally deployed the local VM for capability, safety and security reasons, mapping only the data a user explicitly added to a session. With the redesign, announced October 5, 2026, the VM now runs remotely alongside the model itself.
The shift follows persistent user complaints about the practical costs of keeping the VM on local hardware. Anthropic acknowledged that people enjoyed what Claude could accomplish but disliked the disk usage, battery drain and general performance hit caused by running a VM locally. A second frustration compounded the first: closing a laptop lid stopped all work in its tracks. Those two constraints effectively tethered Cowork sessions to a powered, open machine, limiting its usefulness as a background agent.
Under the new architecture, every session is provisioned with its own isolated sandbox, with no state shared between sessions. When the cloud-hosted VM requires something residing on a user's device, such as a document or folder, the desktop application mediates that file access tool call. This arrangement keeps device data under local control while computation happens remotely. Anthropic framed the change as solving several problems it had heard repeatedly from users.
Among the benefits Anthropic cited are the ability to run Cowork from a phone, the ability to keep work running without a laptop staying open, and access to full capability without sacrificing battery life. The company also suggested users would get the same power as before, but without the local overhead. Felix Rieseberg of Anthropic detailed the reasoning behind the new design, with the company pointing users to a help page for additional implementation details.
The announcement arrives amid heightened scrutiny of agent deployment models across the industry. Recent commentary from the same period has raised questions about default hard budget caps on AI systems generally, while OpenAI's DevDay 2026 and broader 2026 large language model developments have kept competitive pressure high. Anthropic's move toward fully cloud-hosted sandboxes positions Cowork against rivals that are similarly pushing always-on, remotely executed agents.
Security observers will note the tension between offloading compute and preserving local data boundaries. Anthropic's answer is per-session isolation plus user-mediated file access, a design that keeps sensitive files on-device while the reasoning and tool execution live elsewhere. Whether that satisfies enterprise security teams remains to be seen, and the success of the new Cowork will likely hinge on that balance.
Looking ahead, the redesign sets a clearer expectation for Cowork as a persistent, device-agnostic assistant rather than a locally supervised process. If Anthropic can demonstrate that remote sandboxes preserve both safety and responsiveness, the transition could become a template for how agent products are delivered. Users, meanwhile, gain sessions that survive closed laptops and reach across phones and desktops.
