Anthropic wants to turn Claude into the coworker who never leaves the office.
The company's Claude Cowork tool can now control your desktop autonomously through a feature called Dispatch. The idea: assign Claude a task, walk away from your computer, come back to finished work. It can analyze spreadsheets, draft presentations, organize files and run research, clicking through apps and websites the way a person would.
Setup takes about two minutes. Install the app, scan a QR code, start delegating. Claude runs in a sandboxed environment and asks for permission before doing anything sensitive.
Dispatch is the centerpiece of a broader product sprint from Anthropic — 74 features shipped in 52 days across February and March, with 28 in Claude Code alone. Scott White, Anthropic's head of product for Claude Enterprise, said the feature is a direct answer to what companies keep asking for. "We've heard loud and clear from enterprises: you want Claude to work the way that your company works."
The vision is compelling. The results are a coin flip. In hands-on testing, Dispatch has roughly a 50% success rate
David Eastman at The New Stack tried asking Dispatch to find a passport scan via Google Drive. After a series of permission prompts, he reported, "it just gave up on the query and died." His takeaway: the cautious approach might leave Anthropic behind more aggressive tools like OpenClaw, which don't stop to ask before acting.
That trade-off is by design. Anthropic is betting that controlled agents will win the enterprise customers who actually write big checks, even if power users find the guardrails frustrating. But even willing enterprises aren't ready. A Cisco survey of major enterprise customers found that 85% are experimenting with AI agents, yet just 5% have moved them into production.
Anthropic's own users feel that tension. In a survey of roughly 81,000 Claude users last December, unreliability was the top concern at 26.7%, followed by jobs and economic fears at 22%, and autonomy and agency worries at 17%. Users want Claude to do more. They're also nervous about what happens when it does.
Dispatch is available across Anthropic's paid tiers, starting at $20 per month for the Pro Plan, with Max plans at $100 and $200 per month for heavier usage.
Valley View
Anthropic is betting that enterprises will choose the cautious agent over the capable one. That's probably right — a 50% success rate in a sandbox is manageable; a 50% success rate with full desktop access is a nightmare. But reliability is the bottleneck now, not permissions. The companies that push autonomous task completion well past the coin-flip threshold will own the next phase of AI adoption. Anthropic has the trust. It needs the consistency.
