Claude Completes the Agent Puzzle: Can Now Directly Manipulate Computers and Browsers

Dynamic Beating AI News Flash: Anthropic has officially unveiled Claude's computer operation capabilities and browser tool. Previously, each mouse click, input, or other action by Claude required the model to be called again; now, it can sequentially complete clicks, inputs, keystrokes, and screenshots in a single task.

The new browser tool no longer relies solely on screenshots to locate buttons; instead, it simultaneously reads the webpage structure and visual information, enabling direct manipulation of the page based on its elements. This approach makes it less vulnerable to button position changes caused by webpage layout modifications.

Anthropic has also opened up the Skills API and Files API. Developers can encapsulate enterprise processes into reusable Skills, allowing Claude to execute tasks following a predefined workflow; uploaded files can now be reused across requests.