Today, Anthropic announced Cowork’s tool execution is moving from local VMs to cloud sandboxes. The shift reveals what really happens when theoretical security benefits collide with battery drain, disk usage, and performance lag. For SMBs evaluating or building agent platforms, this is the moment to stop assuming architecture choices are ideological and start doing the math.
When Anthropic first shipped Cowork, the architecture was philosophically pure: model inference ran in the cloud, but tool execution—the agent’s ability to actually touch your files and make system calls—lived in a local Anthropic-managed VM on your machine. The reasoning made sense: keep the agent’s hands in a sandbox you control, on your hardware, where your data doesn’t have to leave your desk.
It’s a security architecture that sounds right. Constrained execution, explicit data access, no surprise API calls. Felix Rieseberg, who leads the Cowork effort at Anthropic, spent months defending exactly this design. For corporate environments and regulated workloads, local execution was the feature.
Then the real world intervened.
What users actually complained about
The VM worked. But the cost was material. Disk usage accumulated in ways users found frustrating. Startup overhead was noticeable. Battery drain became tangible. Performance degradation, especially on mid-range laptops, was persistent. Nothing catastrophic, nothing that made the product unusable, but enough that the friction added up. Close your laptop, and work stopped. Run it all day, and your battery took a hit.
The security benefit—your files stayed on your machine—existed in theory. In practice, most users never thought about it. They closed their laptop, expected Cowork to continue, and got frustrated when it didn’t.
As of today, that’s all changed. Anthropic is moving tool execution to the cloud. Both the model and the VM now run on Anthropic’s servers, bound to your account. Your desktop app becomes a client: when you need a file, the app handles the access. Sessions persist. Work continues whether your machine is on or off. No battery drain. No local disk tax. No startup penalty.
Why this matters more than a feature update
This is a structural decision, and it tells you something important about how agentic AI platforms will actually deploy in the wild: theoretical security architecture loses to real-world user experience, even at a company that can afford to be principled about it.
Anthropic isn’t a scrappy startup optimizing for engagement at any cost. They’re shipping enterprise-grade AI with serious security considerations. They had the conviction to build local execution. And they abandoned it not because it was broken, but because the user experience cost was too high.
What does that mean for you? If you’re building an AI agent platform, it means that whatever architecture you choose—local, cloud, hybrid—needs to be defended not by argument but by math. Quantify the real infrastructure cost. What’s the disk footprint? The battery drain per hour of use? The startup latency? The CPU/memory when idle? If you can’t defend those numbers, don’t ship the architecture.
If you’re evaluating agent platforms for your team, it means you should stop asking “where does the code run?” and start asking “what’s the real-world cost of that choice?” A vendor tells you their agents run locally for security? Ask them: what’s the battery impact? How much disk? How fast does it start? If they handwave, they haven’t done the math themselves.
What this means for your decision
The Cowork move doesn’t mean local execution is wrong. For regulated industries—healthcare, finance, law—where data residency is a hard constraint, local execution might be the only option, and the battery cost might be acceptable. But even then, you need to know it.
What it does mean is this: performance, battery, and disk overhead are not minor details. They’re material enough to reshape deployment strategy at one of the best-resourced AI companies in the world. Imagine a 40-person logistics firm evaluating two agent platforms. One promises local execution and security. The other runs in the cloud but cuts startup overhead substantially and saves battery on daily use. The math becomes obvious pretty fast.
For most SMBs, the takeaway is simpler: when a vendor pushes a particular architecture (local, cloud, edge, hybrid), ask them to show you the tradeoffs, not the theory. What’s your battery cost? Your disk footprint? Your latency profile? Make them defend it with numbers. Cowork just gave you permission to do that.
0 Comments