Claude Cowork Moves to the Cloud: What Anthropic’s Architecture Shift Means for How You Work with AI Agents
Anthropic has shifted Claude Cowork's model inference and virtual machine from users' local devices to cloud-hosted sandboxes, eliminating battery drain and allowing tasks to continue when a laptop is closed. Non-technical professionals should note that this introduces internet dependency and raises data residency questions not yet fully addressed in the available source material.
Anthropic Moves Claude Cowork’s Brain to the Cloud
Anthropic has made a significant architectural change to Claude Cowork, its agentic AI product designed to let Claude take actions on your behalf — reading files, running tasks, and operating like a capable digital assistant. As Simon Willison’s analysis describes in a post published on 5 October 2026, quoting Anthropic engineer Felix Rieseberg, the change moves both model inference and the virtual machine (VM) powering Cowork from your local computer to Anthropic’s cloud infrastructure.
If those words already feel abstract, don’t worry. The next few sections break this down in plain language — and explain why it matters even if you have never touched a command line in your life.
What Was Happening Before (And Why It Was Frustrating)
To understand the change, you need a quick picture of how the old version worked. Claude Cowork needed a small virtual machine — essentially a self-contained mini-computer — running on your laptop or desktop. This VM was the secure environment where Claude could actually do things: access the files you gave it permission to touch, execute tasks, and run tool calls without touching anything outside its sandbox.
According to Felix Rieseberg’s statement quoted by Simon Willison, this design existed for good reasons: “capability, safety, and security.” The VM ensured that only data you explicitly added to a session was accessible to Claude. That is a thoughtful safeguard.
But it came with real costs that users felt every day:
- Battery drain: Running a VM locally is computationally heavy. Laptop batteries took a visible hit.
- Disk usage: The VM occupied local storage space.
- Performance slowdowns: Your machine was sharing resources between your actual work and Claude’s working environment.
- Work stopping when you closed your laptop: Because the VM lived on your device, shutting the lid meant the session ended. Any long-running task Claude was in the middle of would simply stop.
Rieseberg’s quote makes this tension explicit: people “loved what they were able to do with Claude but didn’t love the disk, battery, and performance cost.”
What the New Version Does Differently
The new architecture solves these problems by relocating everything to Anthropic’s infrastructure. Rieseberg describes it this way: the new version “runs model inference and the VM in the cloud. Each session gets its own sandbox, not sharing state with other sessions.”
Breaking that down:
- 1. Model inference in the cloud means Claude’s thinking — the actual processing of your requests — no longer draws on your machine’s resources.
- 2. The VM in the cloud means the secure working environment where Claude takes actions is also hosted remotely, not locally.
- 3. Each session gets its own sandbox means your work is isolated from every other Cowork user’s session. There is no cross-contamination of data between users.
The one remaining local responsibility is file access. When Claude needs something from your device — say, a document you want it to analyse — the desktop app on your machine handles that specific request. As Rieseberg explains, “when the VM needs something on the users’ device (like a file), the desktop app is responsible for that file access tool call.”
This is an elegant division of labour: heavy computation and the sandboxed environment live in the cloud, while your local app acts as a secure bridge for anything Claude needs from your physical machine.
A Concrete Scenario: A Chartered Accountant’s Office in Pune
Imagine a chartered accountant running a mid-sized practice in Pune. Her team handles GST filings, audit reports, and client correspondence for dozens of businesses. She has started using Claude Cowork to help draft summary reports from raw financial data that clients share.
Under the old system, whenever she set Claude to work on a large batch of client files, her laptop’s fan would spin up, the battery would drain noticeably within an hour, and if she needed to step away and close her laptop, she would return to find the task had simply stopped mid-way. Restarting it meant losing progress and reorienting Claude from scratch.
With the cloud-based architecture Rieseberg describes, that experience changes in three practical ways. First, she can start a Cowork session on her office laptop in the morning, then switch to reviewing it on her phone during lunch — something Rieseberg specifically mentions as a benefit: “using Cowork from a phone.” Second, the task continues running in Anthropic’s cloud even if her laptop goes to sleep. She returns to a completed draft. Third, her laptop stays cooler and her battery lasts through client meetings without the VM quietly draining it in the background.
The files themselves stay on her device until she explicitly shares them with a session, preserving the security model the original design was built around.
Why This Architecture Choice Matters Beyond Convenience
For non-technical professionals, it is tempting to see this as a simple quality-of-life improvement. And it is that. But the architectural shift also signals something larger about where agentic AI tools are headed.
When Claude can run tasks in a persistent cloud environment — working even when your device is off — the nature of what you can delegate changes. You stop thinking in terms of “I will sit here and watch Claude do this” and start thinking in terms of “I will set this going and check the result later.” That is a meaningful shift in how AI integrates into professional workflows.
The isolation of each session (“each session gets its own sandbox, not sharing state with other sessions”) is also important for professional use. If you are a lawyer in Delhi sharing sensitive client briefs with a Cowork session, knowing that your session’s data does not bleed into another user’s environment is not a minor technical footnote — it is a basic requirement for professional responsibility.
Limitations and Honest Tradeoffs to Keep in Mind
The source material does not provide exhaustive details about every tradeoff, so it is worth naming what remains uncertain or worth watching:
- Internet dependency: Moving the VM to the cloud means Cowork now requires a stable internet connection to function. Users in areas with unreliable connectivity — which remains a reality in parts of India outside Tier 1 cities — may find this a regression compared to local processing.
- Data residency questions: For Indian professionals working under data localisation norms or sector-specific regulations (such as SEBI guidelines for financial data or DPDP Act considerations), understanding where Anthropic’s cloud infrastructure is physically located and how data is handled in transit becomes an important question the source does not answer.
- Region availability: The source does not confirm which geographies have access to the new cloud-based Cowork or whether it is in a limited preview. Availability in India specifically is not confirmed in the material Simon Willison cites.
- File access through the desktop app: While this is described as a feature — Claude can still reach your local files — it also means you still need the desktop app installed for that bridge to work. A fully browser-based or mobile-only workflow is not clearly supported based on the available information.
- Privacy and enterprise controls: The source does not detail what enterprise data controls or audit logs, if any, govern the cloud VM sessions. For organisations with compliance teams, these details will matter before broad adoption.
What to Watch For Next
Anthropic’s help page linked in the Simon Willison post is worth bookmarking if you are an existing Cowork user or considering it. That page is likely to carry the most current details on session limits, supported platforms, and data handling.
More broadly, this architectural shift puts Claude Cowork in closer conversation with other cloud-hosted agentic environments that have been emerging across the industry. The direction is clearly toward persistent, always-on AI agents that work independently of whether your device is active — and Cowork’s update is an early, concrete example of that trajectory becoming a shipping product rather than a research demo.
If you are a professional curious about trying it, the practical starting point is checking whether Cowork is available in your region and reviewing Anthropic’s help documentation before handing it any sensitive files. The capability is genuinely interesting. The due diligence around data and availability is equally real.
