Anthropic Ships Claude's Agent Stack With 40% Fewer Round Trips

Anthropic just pushed its agent stack out of beta. Four building blocks that developers have been experimenting with for a year are now generally available on the Claude Platform: computer use, a new browser tool, the Skills API, and the Files API. The headline change is architectural. Claude no longer stops after every click.

One turn, many actions

The old computer use loop was chatty. Claude would take a screenshot, send it back, wait for the next model call, click a button, wait again, then type. Each action was one round trip. The new computer_toolset_20260801 lets Claude plan and return a short sequence of actions in a single response, what Anthropic calls a batch action.

Claude can plan a short sequence of actions, such as click, type, and then take a screenshot, and return them together in one response. This is called a batch action; it uses the same response shape as parallel tool use with one difference: you run the blocks in order rather than concurrently. Your application executes them sequentially, stops at the first failure, and returns one tool_result block per action. The economics matter: the updated computer use tool lets Claude take several actions per turn instead of one per model call, so tasks finish in fewer calls and less time. Early-access users reported 20-40% fewer round trips per task.

The concrete numbers from Asteroid, a healthcare automation shop, tell the story better than benchmarks. Their longest claims workflow went from 32 minutes to 13, cost per task fell about 30% across every workflow tested, and completion hit 100%, with no changes to prompts.

Browser tool stops fighting the DOM

Pixel-based automation breaks the moment a website ships a redesign. The new browser use tool takes a different approach. Alongside the screenshot, the agent reads the structure of the page and acts on a specific field or button rather than a position on screen. Claude gets element references plus coordinates, so a shifted layout doesn't invalidate the target. The two toolsets are declared independently and can run in the same request, each in its own coordinate frame.

Skills: procedures as versioned artifacts

A skill is a folder that Anthropic describes plainly: a folder of instructions, scripts, and templates that Claude loads only when a task calls for it. The Skills API lets you upload that folder, version it, and attach it to any request by pinning a specific version_id or accepting latest. They run inside Claude's code execution sandbox, so there's no hosting to manage.

The pattern is familiar to anyone who has shipped software. A team encodes its procedure once. Changes are versioned. Requests declare which version they expect. This is what turns a demo into a workflow you can debug when it drifts.

Files API grows up

The Files API is the storage layer for agents. Upload once, reference by file_id across future requests, download whatever the agent produces. This release adds three things developers were waiting on:

• expires_in_seconds: automatic file expiration, so you stop paying for artifacts nobody reads
• 500 RPM: a 5x rate limit bump
• 1 TB per organization: enough headroom for real production loads

Combined with Skills, this is what makes Claude Managed Agents possible. Your procedure lives in Skills, your documents live in Files, and the agent stitches them together per request.

What the loop actually looks like

The developer-visible change is that your agent loop has to iterate over every tool_use block in a response, not just the first one, and dispatch on the block's name together with toolset_name. A response mid-task now looks like this:

{
  "content": [
    {"type": "tool_use", "name": "left_click",
     "toolset_name": "computer", "input": {"coordinate": [640, 60]}},
    {"type": "tool_use", "name": "type",
     "toolset_name": "computer", "input": {"text": "pictures of cats"}},
    {"type": "tool_use", "name": "screenshot",
     "toolset_name": "computer", "input": {}}
  ],
  "stop_reason": "tool_use"
}

If any action in a batch fails, you return is_error: true for that block and stamp every later block in the batch with the halt text Not executed: an earlier computer action in this turn failed. Claude then sees the partial trace and replans on its next turn. A request that leaves any tool_use block unanswered gets rejected.

New defaults worth knowing

A few migration details will bite people upgrading from computer_20251124:

• The beta header is gone. Requests go through the standard API.
• Zoom is now a first-class member tool, enabled by default. If your environment doesn't support it, disable it explicitly with configs rather than returning errors.
• Display dimensions are no longer declared. Coordinates always live in the pixel space of whatever screenshot you return, so the toolset rejects a screenshot or zoom image that exceeds the model's image limits instead of downscaling it. You resize before sending and scale coordinates back yourself.
• The key action supports a repeat count from 1 to 100.

Screenshots stay expensive. Each runs roughly 1,000 to 1,800 input tokens, and requests carrying more than 20 images get held to stricter per-side limits. Anthropic recommends keeping the last three screenshots and pruning in batches every 25 turns so the cached prefix stays byte-identical between prunes.

Where this lands

Computer use has always been the flashiest demo in the agent space and the least trustworthy in production. Latency, click precision, and prompt injection in screenshots kept it in the lab. The batched action model doesn't fix precision, but it changes the cost curve enough that longer-horizon tasks become viable. A 32-minute claims workflow dropping to 13 minutes is the difference between an interesting research artifact and something you put in front of a paying customer.

The pairing with Skills and Files is the strategic move. Computer use alone is a puppet. Add versioned procedures and persistent storage and you get something that looks more like a junior employee with a runbook. Anthropic is also flagging that computer use is now eligible for HIPAA-regulated workloads under our BAA, which opens the door for the healthcare and insurance use cases that dominated the early-access list.

Prompt injection risk hasn't gone away. Anthropic runs classifiers over screenshots to flag suspicious instructions and steer Claude toward asking the user before acting, but the docs are explicit that these precautions are necessary, not sufficient. For anything touching real money or medical records, the human-in-the-loop pattern is still the right default.