Claude's agent stack goes GA: computer use, browser, Skills, Files
Anthropic pulls its agent toolkit out of beta: multi-action computer use, ref-based browser control, versioned Skills, and a Files API with 5x limits, 1TB.
Copy markdown
Computer use is GA — and takes several actions per turn
The computer use tool leaves beta across Fable 5, Mythos 5, Opus 5, Sonnet 5, and Opus 4.8, now taking several actions per model call instead of one — fewer round-trips and lower cost per task. It's also BAA-eligible for HIPAA workloads.
A browser tool that drives pages by accessibility tree
The new browser_toolset_20260801 skips pixel-hunting: read_page hands Claude an accessibility tree with element refs like ref_3 to act on, exposes 27 operations, and adds roughly 6,600 input tokens per request. You host the browser yourself, and it isn't in Managed Agents yet.
Skills API drops the beta header
The /v1/skills endpoint is GA — upload and version your own skills that run inside Claude's code-execution sandbox, with no separate hosting to stand up. It's also live via Microsoft Foundry.
Files API: 5x rate limits and 1TB per org
Files graduates from beta with automatic expiration, pagination, 5x higher rate limits, and 1 TB of storage per organization — enough to stop hand-managing document lifecycles in agent pipelines.
Guardrails: Enterprise Admin API GA, plus agent domain controls
Enterprise user-management endpoints — members, invites, groups, and custom roles — lose their beta headers, and Managed Agents can now allow- or block-list domains for web_search and web_fetch, so you can fence in what an agent is allowed to reach.