I'd counter that n8n (or most other workflow tools) can handle as much ambiguity as OpenClaw - it has a LLM call node. Stuff the ambiguous parts in there, but don't burn a rainforests worth of compute figuring out how to call the weather API each and every time.

Also, in the olden days of pre-AI, if our weather workflow did not notify us because conditions juuust failed to be met, we adjusted the thresholds. Uphill, both ways.

Don't get me wrong, I use a bunch of LLMs for automations. By prompting the model "here is what I want to achieve, here are the tools I have, figure out how to stitch them together". Actual workflows run (mostly) deterministically, with a sprinkling of "classify this image" or "summarize this text" nodes thrown in for a good measure.