I think you guys are learning what CDP does and how to use it in practice. I've been deep it in for 8 years in BrowserBox and the model that works is a few set of primitives, roughly mapped to real human user actions and needs, implemented with CDP to match how a real browser works. I think for now you've given up on that and are letting the agent write its own CDP improvizations, through lack of understanding and deep experience with these primitives. There's a lot of quirks in CDP itself, undocumented behavior, unclear naming and so on. Playwright/PPTR code provides a great reference for some things (keyboard bindings, etc), but in other case (multi tab management, and more), it's not a great reference, and some of their abstractions lead in the wrong direction for what you want.

I think you've settled on the "easier / simpler" solution for now (let AI figure it out). But that solution, while logical and workable at your current juncture is not optimal long term if you really want to build in it.

That's okay - you guys are successful and making bank with browser use, so these considerations truly don't matter. Just keep doing what you're doing, because overall what you're doing is working for your needs!

My approach is always going to be slightly different - it's not necessarily better or worse than yours, but without knowing too much about your product - I believe we have different aims, so the different approaches make sense, perhaps! Or maybe I just know what I'm talking about, and you guys don't. Just kidding! I'm sure you know a lot.

At the end of the day - no problem in this domain is that hard - but it is interesting to work on and I'm glad to see you guys are reaching out and doing something as well.

Let me know if I didn't understrand anything about what you meant.