Anthropic Takes Its Agent Stack to General Availability: Computer Use, Browser Tool, Skills and Files APIs Go Production

On August 20, 2026, Anthropic shipped one of the most consequential platform updates of its year — and it didn’t arrive with a flashy model launch attached. Four tools that define what an AI agent can actually do on the Claude Platform moved from preview to general availability (GA) all at once: computer use, the new browser use tool, the Skills API, and the Files API. No waitlists, no experimental flags, no beta header on the new computer toolset. What was a research demo in October 2024 is now a production surface.

The significance is easiest to see in hindsight. Computer use debuted as an experimental capability when Claude 3.5 Sonnet and the “computer use” API first appeared in late 2024 — a striking demo of an AI clicking through a GUI, but one most enterprises wouldn’t touch. The GA milestone marks the end of that caution period. Anthropic is now telling developers: this is the stack for building agents that operate software, browse the web, apply your organization’s procedures, and manage documents — and it is supported, rate-limited, and compliance-ready.

What went GA, tool by tool

Computer use is production-ready and can read and write files, take screenshots, and control the mouse or keyboard. Critically, it now carries HIPAA eligibility for enterprises — a signal that Anthropic expects regulated industries to deploy it, not just developer hobbyists. The docs still carry a sober warning worth quoting: in some circumstances Claude “will follow commands found in content even when they conflict with your instructions,” meaning instructions on webpages or in images can hijack an agent mid-task. Prompt injection remains the tax on every screen-reading agent.

The browser use tool is the key addition to computer use in this release. Rather than controlling a pixel-level desktop, it gives agents native, in-browser control for tasks that require the web — navigating pages, filling forms, extracting data. For agentic workflows, this splits naturally: browser use for the web layer, computer use for everything else.

The Skills API got “a simpler API for uploading and versioning” custom skills. Skills are the mechanism for encoding procedural knowledge — a folder with a SKILL.md manifest (YAML frontmatter with name and description) plus any scripts, templates, and reference files a task needs, up to 30 MB uncompressed. Claude loads a skill’s instructions only when the task calls for it and executes bundled scripts in its sandboxed code environment. The GA release works on the standard API with just x-api-key and anthropic-version: 2023-06-01 — no beta flag — and it’s available through Microsoft Foundry as well. Anthropic-managed skills ship prebuilt with IDs like pptx, xlsx, docx, and pdf; custom skills are private to your workspace with generated IDs.

The Files API now features a rate limit five times higher than before and 1 TB of storage, letting developers upload and access files without hitting platform limits early. Files flow through uploads or the code execution tool, and they complete the production loop: a skill generates a document, and the response carries a file_id you retrieve through the Files API.

The production loop these four tools create

The GA bundle is designed to work as a system, and the developer experience documentation makes the loop explicit. A Messages request can carry up to 20 skills through the container parameter, riding on the code execution tool; Claude reads each skill’s description and loads only the instructions the task needs. When the skill produces a deck or a filled spreadsheet, the output comes back through the Files API. If the task needs the web, the browser tool takes over; if it needs local software, computer use drives it.

Two mechanics deserve developer attention. First, versioning is snapshots, not diffs: creating a new skill version re-uploads the entire file set, and files you omit are not carried over — treat skill folders like build artifacts kept in CI, with old versions addressable by their skver_* IDs as rollback anchors. Second, workspace scoping is the multi-tenant trap: custom skills are shared across an entire workspace and every API key in it, so any multi-tenant product needs one workspace per tenant (organizations get up to 100 before talking to an account team). Keys, files, and skills all inherit that boundary.

On the safety side, Anthropic says the Skills API operates under zero data retention (ZDR) — no data is retained on servers when the tool is called. Combined with HIPAA eligibility for computer use, the compliance posture is clearly aimed at enterprise buyers.

Why this matters: the agentic stack consolidates

Context matters here. The GA lands the same season Anthropic is reported to be preparing the largest IPO in history, with investors targeting a $2 trillion valuation as soon as October. In that light, this release is less about features than about de-risking the enterprise adoption funnel: stable APIs, compliance signals, and predictable limits are what procurement departments need before agents touch production systems.

It also lands one month after Anthropic updated the Model Context Protocol with the stateless 2026-07-28 specification — dropping protocol-level sessions and the initialize handshake so MCP servers can scale horizontally behind plain HTTP. MCP is how Claude reaches outside apps; the GA tools are how Claude acts. Together they form the two halves of an agentic architecture: a stateless, scalable protocol for connections, and production-grade tools for action.

The competitive read is straightforward. OpenAI and Google have leaned on consumer app growth; Anthropic keeps investing in the plumbing that businesses build on. Its connector directory has passed 950 MCP servers, used by millions daily. Every tool that goes GA converts an experiment into an invoice line — and every skills folder a company uploads deepens lock-in to the platform where those procedures live.

For developers, the practical takeaway: the training-wheels era of Claude agents is over. The stack is production-ready, and the sharp edges that remain — snapshot versioning, workspace scoping, prompt injection on untrusted screens — are documented, priced, and yours to engineer around. The companies that figure out governance for screen-controlling, web-browsing, procedure-following agents first will have a real head start.