Amazon’s New AI Agent Just Booked My Entire Flight

You just say: nova.act(”Click the login button”)

Imagine you spent hours on a browser automation, and then a developer renamed a single CSS class and broke the whole thing, that’s the 'Selector Tax' in action.

We spend roughly 80% of our time acting as “CSS janitors” babysitting scripts that break the moment a developer moves a button by 10 pixels. We’ve all been there: the tests are green locally, but the moment they hit the “real world” of dynamic IDs and shadow DOMs, they fall apart.

It’s frustrating, it’s expensive, and honestly? It’s boring. We should be building features, not arguing with XPaths.

The Problem: Why Traditional Automation Breaks

Traditional tools like Selenium or Playwright are “blind.” They don’t “see” the website; they just hunt for a specific string in the code. If a site uses React or Vue and generates dynamic class names, your automation script is basically guessing.

This leads to “flaky” tests, the kind that pass 8 out of 10 times for no reason, eroding trust in the whole system until your team eventually just ignores the failures.

Enter Nova Act: Automation with “Eyes”

Amazon Nova Act is a shift from telling a machine where to click to telling it what to do. It’s powered by a “computer use” model that actually looks at the screen.

Instead of writing: await page.click(’.submit-button-01-final-v2’)        

You just say: nova.act(”Click the login button”)        

Because the AI understands the intent and the visual layout, it doesn’t care if the button’s ID changed or if it moved to a different corner of the screen.

Why It’s Actually Reliable (The 90% Factor)

You’ll see a “90% reliability” stat thrown around. Here’s why it isn’t just a marketing number: Nova Act wasn’t just trained on the internet; it was trained in “web gyms.”

These are simulated enterprise environments, fake CRMs, travel sites, and HR portals where the AI plays “web chess” against itself millions of times. It learned how to handle pop-ups, cookie banners, and loading spinners by failing in a simulator until it got it right.

Real World Examples: No More Demos

This isn’t just for toy projects. Real teams are using it to solve actual headaches:

• Hertz: They have to test core rental flows (booking, canceling, modifying) across modern sites and ancient legacy apps. They went from spending weeks on test engineering to creating agents in under 30 minutes.
• Sola Systems: A startup reconciling hundreds of thousands of payment records and medical updates every month. They use Nova Act for the “messy” parts of the web where APIs don’t exist, maintaining a 90% success rate even when the target sites change.
• 1Password: They’re using it to simplify how users access logins across hundreds of different websites with a single prompt.

The Human in the Loop

We shouldn’t blindly trust an AI with a corporate credit card. Nova Act has a built-in “Human-in-the-Loop” (HITL) system.

If the agent hits a CAPTCHA or a weird MFA screen, it doesn’t just crash. It can pause, send you a screenshot, and let you take over the mouse and keyboard to solve the blocker. Once you’re past it, you hand control back to the agent to finish the job.

How to Get Started (In 3 Simple Steps)

You don’t need a massive infrastructure to test this out

• The Playground: Go to nova.amazon.com/act. Type a URL, give it a goal (like “Find the cheapest flight to NYC”), and watch it work in a live browser.
• The IDE: Once it works, download the Python script. Drop it into VS Code or Cursor using the Nova Act extension. You can debug step-by-step and add your own logic.

• The Cloud: When you’re ready to scale, hit “Deploy.” It packages your script into a container and runs it on AWS without you having to manage a fleet of “flaky” Selenium servers.What’s next?

Web automation has been a “maintenance tax” on developers for twenty years. We’ve accepted that tests will break and that 3:00 AM calls are just part of the job.

Nova Act is the first time we’ve been able to treat a browser like a teammate rather than a brittle series of code paths. It’s about spending less time fixing selectors and more time actually shipping code.

Shoutout to Amazon for sponsoring this post.