Maybe there’s a middle ground where a small local model can roll with the variations in a site that would break a script, while saving the per token costs?
We are thinking through on self healing mechanisms like falling back to a live web agent and rewriting script.