You talk to Claude. Claude talks back. That's been the deal — a pleasant text exchange where you do all the actual clicking, form-filling, and spreadsheet-wrestling yourself. The AI thinks. You do.
Anthropically speaking, that arrangement just expired.
On March 23, 2026, Anthropic shipped two features that redefine what "AI assistant" actually means. Computer Use gives Claude literal control over your Mac — opening applications, navigating browsers, clicking buttons, typing into forms. Think of it as lending your AI a mouse and a pair of hands. Auto Mode is the safety net underneath: a separate AI layer that reviews every single action Claude plans to take before it executes. An AI watching an AI. The philosophical comedy writes itself.
How the hands work
Computer Use is available to Claude Pro and Max subscribers on macOS. You pair your desktop app with Anthropic's mobile app, then use Dispatch — a new mobile tool for assigning tasks remotely. Tell Claude to clean up your spreadsheet from your phone, go make coffee, come back to find it done. The feature runs on Sonnet 4.6 and Opus 4.6 in research preview.
Auto Mode is where things get genuinely clever. A separate AI system screens every action Claude wants to perform before execution. It checks for risky behavior, prompt injection — when someone tricks an AI into doing something it wasn't asked to do — and actions that drift beyond your original request. The concept is straightforward: build an internal compliance department inside the AI itself, so it second-guesses every move before making it.
What actually happens
In Anthropic's demos, Claude handles multi-step workflows — research across browser tabs, data entry, form navigation — tasks that normally eat fifteen minutes of your life in mindless clicking. The reviewer layer caught attempts to access unauthorized files and blocked actions outside the user's original scope. On paper, solid. In practice — that's what research previews exist to find out.
The fine print
macOS only. Research preview. And here's the telling detail: Anthropic recommends running this in sandboxed environments — isolated spaces where Claude physically cannot reach your real files. When the company building the feature suggests you put it in a cage first, you know the confidence level. This isn't production-ready software. It's a promising prototype wrapped in caution tape.
There's also the recursive problem. An AI reviewing another AI's actions is automated bureaucracy. Who reviews the reviewer? What happens when the compliance layer hallucinates a risk that isn't there, or misses one that is? It's turtles all the way down — except these turtles have access to your filesystem.
What this means for you
If you're on Mac with a Pro or Max subscription, you can experiment right now. Send tasks from your phone, let Claude handle the grunt work, see what breaks. For everyone else — Windows, Linux, the free tier — this is a trailer for what's coming. The gap between "AI that suggests" and "AI that does" just shrank noticeably.
Anthropic isn't building a better chatbot anymore. They're building a coworker — one that runs every mouse click past its own internal review board before executing. Slower than a human, arguably more careful. The assistant just got promoted. Whether you trust the new hire is entirely your problem.
