On March 24, 2026, Anthropic shipped a notable update to Claude: Computer Use is now available in Claude Code and Cowork for Pro and Max plan subscribers, and Auto Mode has entered research preview.
What Computer Use Does
Computer Use lets Claude observe a screen and interact with it directly — opening files, running terminal commands, navigating browsers, clicking UI elements. This extends beyond text-based tool calling, which required wrapping every action into a structured API call.
With Computer Use, Claude can interact with GUI applications that don't expose APIs at all.
# Example: Delegating a multi-step debugging task
# Instruction: "Find why the integration test is failing and fix it"
# Claude executes:
# 1. Run the test suite in the terminal
# 2. Read the failure output
# 3. Open the relevant source file
# 4. Apply the fix
# 5. Re-run the test to confirmThe collaboration pattern shifts: instead of Claude proposing a fix and you applying it, you describe the task and Claude handles the execution.
Auto Mode: Controlled Autonomy
Auto Mode introduces judgment — Claude decides which actions are safe to take without explicit human confirmation. It's in research preview, meaning the scope is intentionally limited. Irreversible actions (deleting files, triggering external APIs, pushing to production) are blocked or require confirmation.
Anthropic's careful approach here reflects a real concern: an agent that makes a wrong call on a destructive action is worse than no agent at all.
Tasks that fit well within Auto Mode's current scope:
- Running linters and applying auto-fixes
- Executing test suites and summarizing failures
- Navigating codebases to identify relevant files before a refactor
- Updating comments or documentation based on code changes
Claude Sonnet 4.6 Improvements
This release also brings capability improvements across coding, long-context reasoning, agent planning, and design tasks. The 1M token context window (beta) makes whole-codebase analysis practical — passing an entire repository as context is no longer a workaround, it's a usable workflow.
What to Watch Before Adopting
Define action boundaries explicitly. Without clear scope, an agent may traverse directories or run commands you didn't intend. Specify allowed action types upfront.
Keep production access restricted. Computer Use combined with production credentials is a meaningful security consideration. Apply least-privilege principles to agent permissions.
Auto Mode is early. Research preview means behavior may change and edge cases exist. For production-critical automation, wait for GA.
The Broader Shift
The pattern emerging across Claude, GPT-5.4, and Gemini 3.1 Pro is consistent: AI assistants are moving from suggesting to executing. For developers, this reframes the work — you define task contracts and verify outcomes rather than reviewing suggestions line by line.
Start with read-only or easily reversible tasks. Build confidence in the agent's judgment within a bounded scope, then expand. The teams that will get the most value from this shift are those who invest now in writing clear task definitions for their agents, not just ad-hoc prompts.