Claude’s Computer Use Tool Goes GA with Multi-Action Turns, Requiring Legacy Agent Loops to Handle Ordering and Abort Semantics

Anthropic has made computer use, the Skills API, and the Files API generally available, and introduced browser use for web-based tasks. The most important technical migration point is `computer_toolset_20260801`: Claude can now output multiple computer actions in a single turn, eliminating the need to call the model again after every click, text entry, or screenshot. In addition to screen pixels, browser use can read page structure and target specific fields or buttons, making it suitable for work contained within browser tabs. Full computer use, meanwhile, still requires developers to provide the desktop environment and execution loop.

Multi-action turns are not a format extension that can simply be ignored. Applications must execute every `tool_use` block in a response sequentially because later actions may depend on state produced by earlier ones. If the first action fails, subsequent items must be marked as not executed, and each `tool_result` must include the corresponding `toolset_name`. Processing only the first block, running actions in parallel, or failing to return a result for any action can cause errors or divergence in UI state. Migrating from the legacy `computer_20251124` version also involves changes to tool schemas, default scaling behavior, repeated key presses, and oversized screenshot handling. Anthropic says older beta integrations will continue to work temporarily, allowing engineering teams to migrate in stages.

The Skills API now supports uploading and versioning folders containing instructions, scripts, and templates, then attaching them on a per-request basis. Their contents are loaded into Anthropic’s code execution sandbox. The Files API allows PDFs or spreadsheets to be uploaded once and reused by ID. Its GA release adds automatic expiration, fivefold higher rate limits, and 1 TB of storage per organization. Together, the two APIs can connect procedural knowledge, input documents, on-screen actions, and output files into a single agent workflow.

General availability does not eliminate the need to isolate computer agents. Web pages and images may still contain prompt injections, while the integration provider remains responsible for controlling the desktop environment, login credentials, network egress, and confirmation of high-impact actions. The next step is to test how batch size affects latency and screenshot token usage, record complete action traces, and verify in a low-privilege sandbox that failure-triggered aborts truly prevent subsequent actions from executing.