Where the playwright MCP uses the browser as a developer (access to HTML, errors, network requests), nova-act uses the browser as a user. It clicks on X/Y coordinates on the screen, and reads the page as displayed.

...and it has a much different loop. It takes an action, then takes a screenshot, highlights areas of the page based on content and interactivity, decides what to do next using it's tooling.

It solved it in the first go. I even saw it pass a captcha test.

Nova-act was much faster, and it has

Playwright MCP was a bit slow. It also didn't seem aware it could scroll down. it did the job, but seemed prone to all the same selenium-style issues - waits, wrong selectors, giant html pages.

ultimately couldn't get it to add to cart. It got tripped up a lot on the way

That's helpful - I was looking into this last night. I'm surprised there's not better monitoring there for bedrock.

But there's a lot of detail in those logs. Good to know

I started hearing about observability as a category of tools since pre 2020. Prior to that, all the tooling and dashboarding was an ops job - things like sending statsd metrics from backends to a grafana dashboard running off a computer under someone's desk.

Good ol days