uygnoey The production agent stack on the Claude Platform
Computer use, the Skills API, and the Files API reached general availability on the Claude Platform on 2026-08-20, and computer use gained a new browser use tool for agents working in web applications. Code execution and web search were already generally available. Together these are the pieces of an agent that can operate software it can see, apply your expertise, and hand back finished files.
Per-capability detail is in references/capabilities.md; availability and migration notes are in references/availability.md.
Instructions
1. Reach for computer use when there is no API
Computer use lets you build agents that operate software they can see. Given a screenshot, the agent clicks, types, and scrolls the way someone at the keyboard would — which is what makes it work in applications that were never built for automation.
That framing is also the selection rule. If a real API exists, use it: it is cheaper, faster, and deterministic. Computer use earns its place specifically where the surface has no programmatic entry point — legacy internal tools, vendor portals, systems of record that a partner controls.
The GA update changes the economics: 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.
2. On the web, prefer the browser use tool
The browser use tool is new in computer use as of this release, and it is the better default for anything running in a browser. 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.
Two consequences worth designing around:
- Reliability. Targeting a named element survives layout shifts, resolution differences, and re-renders that move a button a few pixels. Pixel coordinates do not.
- It keeps the multi-action turns. The browser use tool uses the same several-actions-per-turn model as the updated computer use tool, so you get the latency and cost benefit as well as the structural targeting.
Use pixel-level computer use for the desktop surfaces the browser tool cannot reach, and the browser use tool for web applications.
3. Package your procedure as a skill, not as a prompt
A skill is a folder of instructions, scripts, and templates that Claude loads only when a task calls for it. With the Skills API you upload and version your own skills, then attach them to any request. They run in Claude's code execution sandbox, so there is nothing for you to host.
Design implications:
- Loaded on demand. Because a skill is pulled in only when the task calls for it, the expertise you encode does not have to compete for room in every request's prompt.
- Versioned as an artifact. Upload and version the skill rather than editing a prompt string in application code. That gives the procedure its own change history, separate from the code that calls it.
- Scripts and templates travel with it. A skill is a folder, so deterministic helpers and output templates ship alongside the instructions instead of being reconstructed by the model each run.
4. Give documents identity with the Files API
The Files API is 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 the files the agent creates.
The GA update adds automatic file expiration, 5x higher rate limits, and 1 TB of storage per organization.
Reference-by-ID is the part that changes agent design: a multi-turn workflow stops re-transmitting the same source document on every request, and the agent's output becomes an artifact you can retrieve rather than text you have to parse out of a response.
5. Compose the loop
The four pieces are meant to be used together, and code execution and web search fit into the same loop. The canonical shape from the announcement:
- Read the input document from the Files API.
- Follow a skill that encodes the team's procedure.
- Act in the target application — the browser use tool for a web portal, computer use for a desktop surface.
- Save the result back as a file.
When you scope a new agent, walk those four questions in order: what does it read, whose procedure does it follow, what software does it have to operate, and what does it hand back?
Examples
A claims agent. The example given in the announcement: it reads the intake document from the Files API, follows a skill that encodes the team's filing procedure, completes the submission in an insurer's web portal with the browser use tool, and saves the confirmation back as a file.
Agents inside systems with no API. A research engineer quoted in the announcement, on agents working inside healthcare and insurance systems that have no API: on the new computer use tool 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 their prompts.
A skill as the unit of domain expertise. Box built specialized document creation into Box Agent with the Skills API: for a bank, a skill captures the firm's credit methodology and approved memo format, and Box Agent applies it to the financial statements and deal documents already in Box to produce a source-grounded credit memo for analyst review. The point made is that customers get agents for complex workflows without building each one from scratch.
More worked compositions are in examples/workflow-shapes.md.
Source
Distilled from Build production agents with computer use, the Skills API, and the Files API (published 2026-08-20).