AI Agents Can Now Click Through Your Workflow. The Question Is: Should You Let Them?

If you've been treating "the agent can click through my software" as a demo you'll evaluate someday, August 19 moved the deadline up. Anthropic took computer use, browser use, the Files API, and the Skills API out of beta and into general availability — all four at once. GA means Anthropic will support these the way it supports anything else customers build production systems on. It stops being a preview you can safely ignore.

What changed, specifically

Browser use now reads a page's accessibility tree instead of guessing from pixels, which is why it handles unfamiliar sites more reliably than earlier versions. Computer use added batch actions, so an agent can queue a sequence of clicks and fills instead of pausing to reconsider after each one. One enterprise reference case Anthropic cited: an insurance claims workflow that took 32 minutes end to end dropped to 13, at 30% lower cost per claim. That's the pitch — and it's also exactly the kind of workflow where a wrong click has consequences a demo doesn't show you.

The distinction that matters isn't the new features. It's what GA means procedurally: a beta tool breaking is an accepted risk of testing something early. A GA tool breaking is a production incident, on a system now handling filled-out forms, submitted requests, and edited files with the same authority a person would have had doing it by hand.

What to do Monday morning

Before you connect a computer-use or browser-use agent to anything that changes state — a form submission, a file edit, an account setting — answer two questions in writing, not in your head.

What is this agent's task, exactly? "Handle the claims intake queue" is a goal, not a scope. "Read each new claim, fill the standard fields, flag anything with a dollar amount over $5,000 instead of submitting it" is a scope. The gap between those two sentences is where an agent decides on its own what counts as within bounds — and batch actions mean it can act on that decision several times before anyone notices.

Where does a human have to look before the action is final? Not after — a completed submission or a saved file is often not free to undo. Pick the one or two steps in the workflow where the cost of being wrong is highest, and require a person to see the agent's intended action there before it executes, even if everything else runs unattended.

An operations manager at a 40-person insurance brokerage

Dana runs claims intake and was the kind of person who read the 32-minute-to-13 case study and wanted it immediately. She connected a browser-use agent to the intake portal with instructions to "process incoming claims." The first week went fine — routine claims, small amounts, nothing unusual. In week two, a claim came in with a data-entry error upstream that made the payout field read $50,000 instead of $5,000. The agent filled the form exactly as instructed and submitted it. Nothing in "process incoming claims" told it a five-figure number needed a second look.

She didn't disconnect the agent. She rewrote the task: everything under $5,000 processes automatically, everything at or above it fills the form and stops for her review before submission. The workflow is still faster than doing it by hand. It's just not faster in the one place where fast was the wrong goal.

A different failure mode: a marketing coordinator at a 15-person agency

Contrast that with Marco, who gave a computer-use agent access to update ad campaign budgets across three client accounts based on a weekly performance report. He'd already scoped the task tightly — adjust budgets within a pre-approved range, nothing else. The problem wasn't scope. It was that the agent ran the full batch of budget changes overnight and emailed him a summary the next morning. By the time he read it, a client's budget had been reallocated for eighteen hours based on one bad data point in the report — reversible, but not free, and not something he'd have approved in real time if he'd been asked.

The takeaway

GA doesn't mean an agent knows where your workflow's expensive mistakes live. It means the tool is stable enough that finding out is now your responsibility, not the vendor's beta disclaimer's.