Anthropic Makes Four Agent Tools GA: Computer Use, Browser Tool, Skills API, and Files API

On August 20, Anthropic announced that Computer Use, a browser tool, the Skills API, and the Files API are now generally available (GA) on the Claude Platform, aiming to let agents operate software, incorporate team knowledge, and read/write files, thereby extending automation to systems without native APIs.

Browser automation: from pixels to elements

The new browser tool, identified by the code browser_toolset_20260801, moves away from pixel‑based automation, which breaks when layouts shift, to DOM‑based interaction using element references and coordinates.

Batch actions reduce calls and latency

Previously, each model invocation could perform only a single action; the updated Computer Use now returns multiple actions (click, type, screenshot) in one batch, cutting the number of model calls and lowering latency. It also meets HIPAA compliance under a BAA, making it suitable for medical use.

Skills API: versioned team workflows

The Skills API addresses workflow solidification by storing instructions, scripts, and templates in a folder hierarchy that agents load on demand. Names and short descriptions stay in context (~100 tokens), while full SKILL.md files load only when triggered, and scripts run via bash with only their output consuming tokens. Versions are managed via version_id or the keyword latest, allowing teams to pin requests to specific workflow versions.

Files API: upload once, reference everywhere

Files API works by uploading a document once to obtain a file_id, which can be reused in subsequent requests without re‑uploading. The GA release adds three features: an expires_in_seconds field for automatic expiration, a rate limit of 500 RPM (five times the previous limit), and a per‑organization storage cap of 1 TB.

Performance numbers

Anthropic’s blog cites two case studies: Asteroid reduced a medical claims workflow from 32 minutes to 13 minutes, cut model calls by 32‑52 %, lowered per‑task cost by 25‑32 %, and raised completion from 77 % to 100 %; Box created a loan‑approval skill that extracts methodology and memo formats, reads financial statements from Box, and generates audit‑ready loan memos for analysts.

Observations on cost and limitations

Each Computer Use action captures a screenshot, which consumes tokens; users note that “most agent demos die from cost, not capability—every click burns a screenshot.” Batch actions reduce call count but screenshots still incur cost. Anthropic suggests scaling screenshots, limiting history, and caching to mitigate expense. Currently, Computer Use cannot be used with Claude Managed Agents; migration guides advise removing the beta header, switching to the new toolset, and adapting to batch semantics.

With all four tools now GA, Claude moves closer to autonomous operation, though the economics of running such agents remain a challenge.