The VM Problem Nobody Talks About

Claude CoWork runs inside a virtual machine. That sounds fine in theory - sandboxed, isolated, safe. In practice, it means your AI coworker is one VM hiccup away from losing everything it was doing.

You're 45 minutes into a complex workflow. CoWork is editing a spreadsheet, has three browser tabs open, and is halfway through drafting an email. Then the VM crashes. Not a graceful shutdown - a hard crash. When it restarts, it has no memory of what it was doing. You start over from scratch.

Babysitting an AI Defeats the Purpose

The whole point of an AI agent is to save you time. If you're spending that time monitoring whether the VM is still alive, checking if the session dropped, and re-explaining context after a restart, you haven't saved anything. You've just shifted your work from doing tasks to supervising a fragile system.

VM restarts aren't instant either. You're waiting 30 seconds to a minute for the environment to spin back up, then another few minutes for CoWork to re-establish its state. That adds up fast across a workday.

Local Agents Skip the VM Entirely

A native macOS agent runs directly on your machine. No virtualization layer, no remote environment to maintain, no session state that evaporates when something goes wrong. If the agent process crashes, it restarts in seconds with its persistent memory intact.

The reliability difference is significant. When your agent runs as a native process on your Mac, it has the same uptime as any other app on your machine. It doesn't depend on a network connection to a remote VM. It doesn't lose context because a container was recycled.

The best coworker is one you don't have to babysit. A local agent that just runs - reliably, quietly, on your hardware - is closer to that ideal than a VM that needs constant supervision.

Fazm is an open source macOS AI agent. Open source on GitHub.