Anthropic is pushing the boundaries of agent autonomy with a significant update allowing Claude Code and Cowork to operate directly on your machine. Starting this week, Pro and Max subscribers can grant the AI permission to open files, navigate browsers, and execute development tools without manual intervention. This shift marks a move from chat-based assistance to active system participation.

Currently restricted to macOS, the feature functions as a research preview. It prioritizes API connectors for services like Slack or Google Workspace but falls back to direct interface control—manipulating the mouse, keyboard, and display—when integrations aren't available. This builds on foundational work from the 3.5 Sonnet release in 2024, now tailored specifically for engineering workflows where context switching costs time.

Setup requires the Claude desktop app running on a Mac, paired with the mobile companion. This enables "Dispatch," allowing engineers to assign tasks from their phone that the desktop agent executes later, even when the user is away from their desk. Anthropic emphasizes safety, noting the system requests explicit approval before clicking or scrolling through sensitive areas.

Performance varies depending on task complexity. While efficient for standard workflows, intricate operations may require retries. The company acknowledges that screen-based control introduces latency compared to native integrations. By releasing early, Anthropic aims to gather data on failure modes rather than presenting a polished product. For ML engineers building similar agent systems, this transparency offers a real-world look at the trade-offs between flexible UI automation and rigid API connections. It signals a future where agents handle routine maintenance, leaving humans to focus on architecture.

Source: The Verge