Claude's Agent Stack Is Production-Ready: Computer Use, Browser Use, Skills API, and Files API Hit GA

On August 20, 2026, Anthropic announced general availability for four pillars of agent building on the Claude Platform: the computer use tool, the Skills API, and the Files API — plus a brand-new browser use tool. The subtext is clear: these capabilities are out of beta and into production, and computer_toolset_20260801 is now the identifier to build against.

What Happened

Four pieces, one loop:

Computer use: an agent that sees a screenshot and acts the way a person at the keyboard would — clicking, typing, scrolling. The GA upgrade: the model now takes several actions per turn instead of one per model call, so tasks finish in fewer calls. Computer use is also now eligible for HIPAA-regulated workloads under Anthropic’s BAA.
Browser use (new): the same idea extended to the web. Alongside the screenshot, the agent reads the structure of the page, so it targets a specific field or button rather than a pixel position — dramatically more reliable in web apps.
Skills API: a way to package your team’s expertise (instructions, scripts, templates) as folders you upload, version, and attach to any request. Skills run inside Claude’s code execution sandbox — nothing for you to host.
Files API: storage for the documents an agent reads and writes. Upload a PDF or spreadsheet once, reference it by ID in later requests instead of re-sending it, and download what the agent produces. GA brings automatic file expiration, 5x higher rate limits, and 1 TB of storage per organization.

Why It Matters

Plenty of enterprise software — think insurance, healthcare, banking back-offices — has no API at all. Computer use is the only realistic bridge to automating it. Anthropic’s launch post quotes a healthcare services provider (Asteroid): their longest claims workflow dropped from 32 minutes to 13, cost per task fell about 30% across every workflow tested, and completion hit 100% — with no prompt changes. Vendor numbers, certainly, but they’re consistent with the mechanics: multi-action turns cut model calls, and calls are where time and money go.

The deeper play: Anthropic is assembling the full production agent stack in one platform — screen vision (computer use), web navigation (browser use), institutional knowledge (Skills), and document state (Files) — on top of the already-GA code execution and web search. The entire agent loop now lives under one roof.

Building Your First Agent Today

The canonical starter: an agent that reads documents and then works in a web app that has no API.

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-sonnet-5",
    max_tokens=2048,
    tools=[
        # Computer use toolset: GA shape, multi-action turns
        {"type": "computer_20260801", "name": "computer",
         "display_width_px": 1440, "display_height_px": 900},
    ],
    messages=[{"role": "user", "content": [
        {"type": "text",
         "text": "Open claims.pdf from the Files API, extract the claim "
                 "number and status, then enter both into the insurance "
                 "portal open in the browser."},
    ]}],
)

A practical adoption path:

Pick your bottleneck. A repetitive manual process inside a desktop or web app with no API is the ideal first target.
Encode expertise as Skills. Approval rules, memo formats, filing procedures — these belong in versioned skill folders, reviewable like code.
Upload files once. Reference by file_id instead of re-sending documents every request.
Design the guardrails. An agent that clicks and types freely needs an isolated environment, least-privilege credentials, and human approval on sensitive actions (payments, deletions, sends).

What Changed vs Beta

Beyond multi-action turns and throughput, the enterprise-relevant shift is commitment: GA means interface stability and support agreements — no more beta headers. Availability is distributed across three clouds: the Skills API and Files API are also live through Microsoft Foundry, and the updated computer/browser use tools are coming to Google Cloud Vertex AI. Existing beta integrations keep working during migration — nothing breaks on day one.

Limitations

Governance precedes deployment. An agent operating real systems with real permissions needs audit logs and rollback paths — that’s your responsibility, not the platform’s.
Dense desktop UIs remain hard. Non-standard desktop applications are still trickier than web apps, where browser use reads actual page structure.
Costs shift, they don’t vanish. Multi-action turns reduce request count but consume more tokens per turn; measure cost per completed task, not per request.

What Developers Should Do

If you have a beta integration, plan migration to the GA toolset identifiers (computer_toolset_20260801) this week. If you’re starting fresh, begin with low-risk read-then-fill agents: extract data from documents and enter it into internal web forms. The official documentation covers all four tools with complete examples.