The computer-use landscape: APIs, agent products and AI browsers

0:00 The computer-use landscape

Here is a problem with learning about computer-use agents. By the time you finish a course, half the product names have changed. Between early 2025 and late 2026, the big vendors launched, renamed, merged and retired agent products several times. So in this lesson, you will learn the landscape as categories, and you will leave with a checklist that still works when the next product launches.

0:29 1. Model APIs

Category one is model APIs with a computer-use tool. You provide the environment, a virtual machine, a container or a browser, and the model proposes actions. Anthropic's Claude API has a generally available computer toolset for recent models, plus a reference Docker demo. OpenAI's computer-using agent model, which powered Operator, is available as a computer use tool in its Responses API, with safety checks you must acknowledge. Google offers computer use through the Gemini API, starting with a Gemini two point five Computer Use preview in October 2025. This is the category you use when a business process must be controlled and audited.

1:14 Why it matters

Why does this matter? Because the choice of tool decides where your data goes, what the agent can reach, and whether you can prove what it did. Pick a consumer product for a regulated process and you may not be able to log or restrict it properly. Pick a heavy custom build for an ad-hoc research task and you'll waste weeks. Matching the category to the job saves money, protects data and avoids painful migrations when the market shifts again.

1:49 Like early smartphones

Here's an analogy for this market. Think of smartphones around the late two thousands. New models, new names and new features appeared every few months, and many disappeared just as fast. The people who navigated that well didn't memorize model numbers. They knew what to look for: battery, camera, security updates, app support. Computer-use tools are at the same stage. Model names and product brands will keep changing. Your evaluation questions shouldn't.

2:20 Simple example: an AI browser trial

A simple example of using the checklist. Say a colleague wants to try an AI browser to research competitors' pricing pages. Run the first three questions. Where does it run? In their everyday browser. What can it reach? Everything they're logged into, including the company's ad accounts. Can we restrict domains in code? Not really. So the answer is: fine to try, but in a separate browser profile with no work logins, and nothing sensitive. Three questions, one minute, and a much safer experiment.

2:57 2. Agent products

Category two is finished agent products, where the vendor runs a cloud browser for you. The OpenAI story shows how fluid this is. Operator launched as a research preview in January 2025, was folded into ChatGPT as agent mode that July, and has been reorganized since. Google's Project Mariner prototype fed into Gemini's agent features and AI Mode in Search. These are great for individuals and one-off tasks. For repeatable business processes, you usually want your own harness.

3:31 3. AI browsers

Category three is AI browsers and browser assistants. Think Perplexity's Comet, Gemini features in Chrome including auto browse on paid plans in supported regions, AI built into Microsoft Edge, and extensions like Claude in Chrome. These work inside your real, logged-in browser. That is powerful, because the agent can use your accounts. It is also risky, for exactly the same reason. A malicious page can try to hijack an agent that has access to your email and admin panels.

4:05 4. Open source and infrastructure

Category four is open source and infrastructure. Playwright remains the standard for scripted browser automation, and the Playwright MCP server lets any MCP-capable model drive a browser using accessibility snapshots. Libraries such as Browser Use and Stagehand put model reasoning on top of browser automation. And hosted browser providers run fleets of isolated cloud browsers with session recording, so you do not have to operate your own.

4:34 Evaluation checklist

Now the checklist. Where does the environment run? What can the agent reach? Can you restrict actions and domains in code, not just in a prompt? Is there a built-in human confirmation step? What defenses exist against prompt injection? Can you log and replay every step? What are the data retention and training terms, which matters under GDPR, UK GDPR, Saudi Arabia's PDPL and the UAE's data law? What does it cost per task on your workload? And is it generally available, or still a preview?

5:11 Worked example: one agency, two choices

Here is how a Manchester agency used it. Account managers wanted ad-hoc research help across client sites. For that, an AI browser was fine, but in a separate profile with no client admin logins. They also needed a weekly audit of two hundred client landing pages. For that, they built an API-based harness with a headless cloud browser, a domain allowlist per client, read-only credentials and stored evidence, because clients want to see exactly what was checked and when.

5:45 Run a fair pilot

How do you run a fair pilot between two tools? Take twenty real tasks from your workflow. Run each on both tools, ideally more than once. For every run, record whether it succeeded when checked, how many times a person had to step in, how long it took, and what it cost. Keep the environment the same for both. At the end you'll have a simple table, and the decision usually becomes obvious. It almost never matches the impression you got from the demo videos.

6:22 Data terms matter

A word on data terms, because this is where legal and IT teams will stop you. When an agent sends screenshots to a model provider, those screenshots may contain customer names, emails or financial details. Check how long the provider keeps that data, whether it's used for training, where it's processed, and whether enterprise or zero-retention options exist. For teams in the UK, the EU, Saudi Arabia or the UAE, cross-border processing questions come up quickly. Settle them before the pilot, not after.

6:58 Three mistakes

Three mistakes in choosing tools. First, choosing by demo video, which is designed to succeed. Always pilot on your own tasks. Second, building a business process on a preview feature with no fallback. If the vendor changes or withdraws it, you need a plan. Third, ignoring data terms until legal or IT stops the project. Screenshots can contain personal data, so settle retention, training use and processing region at the start.

7:29 Try this now

Try this now. Take one real task from your work and fill in the evaluation checklist for two options: one API-based build and one finished product or AI browser. For each checklist question, write a one-line answer and where you found it, ideally in the vendor's documentation. Where you can't find an answer, write unknown. Then run each option on the same five examples of the task, and record success, human interventions, time and cost. You'll have a small, honest comparison instead of an opinion.

8:06 Recap

To recap. Learn categories, not names: model APIs, agent products, AI browsers, and open-source infrastructure. Use products for ad-hoc work and your own harness for repeatable processes. And evaluate every tool with the same checklist. Your next step: fill in the checklist for two tools you could use this week, against one real task. Next, we combine deterministic Playwright scripts with model reasoning to get the best of both.