Does this mean the CLI generates a reusable script or abstraction for a specific site so it doesn't have to re-evaluate the entire page structure on every run?
Yes, there is a difference in the cost of early labeling work, and the execution costs later on will show an even greater gap. Our execution costs are near zero.
I haven’t tried it yet cuz setting up the Chrome extension feels kinda annoying ngl, but the idea is actually pretty exciting.
I feel like this could solve 3 big problems: anyone can use it to make skills that are actually personalized to their own use case; token cost gets way lower for repeat stuff, which is basically what this repo is saying; a lot of businesses are scared to use agents for core workflows cuz they don’t want the agent to randomly go off and mess things up. But this creates a more predictable boundary. Like in theory, all business digital workflows should be automated in a predictable way, not just “trust the AI bro” style.