代理平台
Claude Makes Computer Use, Skills, and the Files API Generally Available as Browser Agents Shift to Element-Level Control
Anthropic’s updated tools can perform multiple interface actions in a single model turn while combining page structure, versioned skills, and persistent files. Existing beta integrations will continue to work, but upgrading the computer use tool changes the request format, and engineering teams must still manage browser isolation and prompt-injection risks themselves.
Anthropic announced on August 20 that computer use, the Skills API, and the Files API on the Claude Platform are now generally available, alongside a new browser use tool designed specifically for web applications. This update goes beyond simply removing the beta label: `computer_toolset_20260801` can perform consecutive clicks, text entry, or screenshots within a single model turn, reducing the round-trip latency caused by calling the model again for every action. In addition to pixel-based screenshots, browser use reads the accessibility tree, forms, tabs, and element references, allowing agents to target fields or buttons instead of relying on fixed coordinates that can easily break when layouts change.
The Skills API packages instructions, scripts, and templates into uploadable, versioned folders. Requests can specify either a fixed version or the latest version, and execution takes place in Claude’s code sandbox. The Files API allows a document to be uploaded once and then referenced by ID across requests. The GA format adds `expires_in_seconds`, `expires_at`, pagination, and ID filtering, while Anthropic has increased the organization storage limit to 1 TB. Python SDK 0.124.0 includes the corresponding types and toolsets, confirming that these interfaces have moved beyond product announcements.
Migration still has pitfalls. Existing beta headers will temporarily preserve the old response format, but the new computer use release changes the request format and tool-handling flow, so upgrading is not as simple as deleting a header. Applications that depend on the previous file-list structure should also update their parsers first. More importantly, browser use relies on an application-hosted browser rather than a complete desktop isolation layer managed by Anthropic. Engineering teams must still restrict accessible domains and download and upload permissions, isolate login credentials, and treat web content as untrusted input that may contain prompt injections. The next metric worth watching is the success rate of multi-action turns across real-world website redesigns, error recovery, and human-approval workflows—not merely the latency reductions reported in vendor customer case studies.