Anthropic, after months of research into screen-based agents, has shipped the three pieces its platform was missing: computer use, the Skills API, and the Files API, released together for production use. The headline capability lets an agent operate software it can see, clicking and typing in any application the way a person would, while the two APIs hand the agent your documented expertise and your actual documents.
What Was Announced
The three releases form one story about production agents. Computer use is the ability for an agent to operate software it can see: given a screenshot, the agent clicks, types, and navigates in a real application, which matters for the vast number of business systems that have no API. The Skills API lets you hand the agent your expertise as a folder of instructions, so it follows your documented process instead of improvising. The Files API gives the agent read access to your documents, so it can work from the actual source material rather than pasted fragments.
The combined pitch is a claims agent that reads the intake document through the Files API, follows a skill that encodes the review process, and then, when a step needs a system that has no API, uses computer use to operate it directly. That is the full production-agent stack in one sentence, and it is aimed squarely at the enterprise workflows that have resisted automation for a decade because the software underneath has no integration surface.
Why Computer Use Is the Big One
Computer use has been the missing piece for agents in legacy environments. Healthcare systems, insurance platforms, government portals, and a long tail of internal tools were built without APIs, and every automation project hit the same wall: the agent could think but could not touch. Computer use removes that wall by treating the screen as the interface, the same way a human employee does.
The practical difference shows up in the testimonials Anthropic published. One customer describes agents working inside healthcare and insurance systems that have no API, using the new computer use tooling where their integration team previously had to build custom connectors for each system. Another describes the Skills API making specialized document creation straightforward inside a banking product. These are not toy examples; they are the kinds of workflows that enterprise teams have been waiting for an agent platform to handle.
The reliability question is the honest one. Screen-based operation is slower and more error-prone than API-based integration, and Anthropic does not claim otherwise. The value proposition is coverage: computer use handles the systems where an API will never exist, and teams should use APIs where they exist and computer use where they do not. Anyone building on it should design for the failure modes: screens change, elements move, and the agent needs verification steps built into the loop.
The Skills API and Files API
The Skills API is a folder of instructions the agent loads, encoding how your organization does a specific kind of work. Instead of re-explaining the process in every prompt, the skill carries the process, and the agent applies it consistently. That turns institutional knowledge into a versioned, reviewable artifact, which is a meaningful change for teams that have been keeping processes in heads and wikis.
The Files API covers the input side. Agents can read documents directly instead of working from pasted excerpts, which matters for anything involving contracts, reports, or correspondence where context and exact wording are the point. Combined with skills, it closes the loop: documents in, expertise applied, actions taken, all inside one agent session.
Anthropic positions the three as the foundation for production agents, and the release notes read like an ecosystem play: the company is clearly trying to make Claude the default agent runtime for enterprises, with the building blocks in place so that vendors build on top rather than around it.
The Competitive Context
This is Anthropic’s answer to the same question OpenAI, Google, and xAI are all answering: how do agents get from chat to real work? OpenAI has been pushing Codex and its computer-use research, Google has Antigravity and Gemini agents, and the Claude side has been steadily assembling the agent stack: Claude Code for terminal work, Cowork for sessions, and now computer use for the screen. The pattern across all of them is the same, agents need eyes, hands, and memory, and this release is Anthropic filling in the eyes and hands.
The pricing and quota details matter for anyone building on it, and the API economics of agent runs, where a single computer-use session can consume substantial tokens, are the thing to model before committing. Teams already running Claude Code will recognize the token patterns, and our Claude Code token pricing breakdown covers the cost model that applies to these API workloads too. For the coding-agent side of the same family, the OpenAI Codex guide is a useful comparison point.
The Honest Caveats
Computer use is not magic. It is slower than APIs, and screens are brittle: layout changes, pop-ups, and multi-step flows all need verification logic. Anthropic’s examples are curated, and the production reality will involve retries, human-in-the-loop checkpoints, and careful prompt design. The Skills API requires the expertise to be documented in the first place; if your process lives only in someone’s head, a skill cannot extract it. And the security surface is real: an agent that can click and type in any application is a powerful tool, and enterprises will need guardrails around what the agent is allowed to touch.
Who Should Care
Enterprise integration teams get the most from this release, because computer use targets exactly the legacy systems that have blocked automation. Developers building agent products should evaluate the three APIs as a stack rather than individually, since the value compounds when documents, skills, and screen control work together. Anyone with compliance requirements should check the data handling story before putting confidential documents through the Files API. And AI platform watchers should note that the agent stack race is now firmly about production workflows, not demos, with Anthropic making the strongest enterprise play yet.
The Bottom Line
Anthropic has shipped the pieces that turn Claude from a model into an agent platform for real work: computer use for the systems without APIs, the Skills API for the expertise, and the Files API for the documents. The release is aimed at enterprises, the testimonials are from real customers, and the direction is clear. The caveats are the standard ones for screen-based automation, reliability, security, and the cost of agent runs, but for the long tail of software that never got an API, this is the most credible automation path yet offered by a major lab.