Anthropic Gives Claude Code an Autopilot Mode — With Guardrails
Claude Code's new auto mode lets the AI execute multi-step tasks with fewer interruptions, using built-in safeguards to block risky actions before they run. Anthropic also shipped computer use in Cowork, letting Claude operate your Mac directly.
Anthropic shipped two significant upgrades to its agentic products on March 24: auto mode for Claude Code and computer use inside Claude Cowork. Both expand what Claude can do without stopping to ask permission — and both add new layers of automated safety checks to match.
Auto mode is the headline feature. Until now, Claude Code required explicit user approval for most actions. Auto mode removes that friction, letting Claude execute long chains of tasks — reading files, running commands, editing code, searching the web — without interrupting the session. The catch is that every action is reviewed by an AI safeguard layer before execution. The system checks for behavior the user didn’t request and scans for prompt injection attempts, where a malicious instruction embedded in an external file tries to redirect Claude’s actions.
This addresses a real attack surface. As AI agents gain write access to file systems, terminals, and browser sessions, they become targets for prompt injection via the content they process. Anthropic’s solution is to interpose an automated reviewer between the agent’s plan and its execution. It’s not a silver bullet, but it raises the cost of a successful attack.
Dispatch for Cowork is the companion product. Cowork — Anthropic’s visual, non-terminal interface built on Claude Code — now lets users delegate tasks to a persistent agent thread and walk away. When paired with computer use, Claude can point and click through macOS applications, open files, navigate browser windows, and return a finished result.
The underlying capability isn’t new. Computer use launched in Claude’s API in late 2024. What’s new is the surface: Cowork wraps it in an interface aimed at non-engineers. Marketing teams and analysts at Anthropic reportedly started using Claude Code internally because of its multi-step capability, bypassing the chat interface entirely. Cowork formalizes that migration path.
Availability is tiered. Auto mode is in research preview. Cowork with computer use is rolling out to Max subscribers first, with Pro plan access following within days (macOS only).
The Economic Index findings Anthropic released alongside this launch are worth noting. Users with more Claude experience don’t hand over more autonomy — they hand over less. They iterate more carefully, define task boundaries more precisely, and retain oversight on higher-stakes decisions. That pattern argues against the assumption that more powerful agents will naturally lead to users switching off.
The Dispatch + auto mode combination puts Anthropic squarely in competition with Devin, Cursor’s background agents, and OpenAI’s Operator. The differentiator Anthropic is betting on: safety controls that aren’t bolted on after the fact, but integrated into the execution loop.