Most people I know are using AI APIs as if they're deterministic - like calling a sorting algorithm. But in my experience, when you rely on AI in a critical step of a workflow, a wrong result can trigger a domino effect across the entire system.

I wrote about that butterfly effect here [1]. Curious how others are handling this in multi-step AI workflows. Whats worked for you? Any lessons learned from failures?

[1] https://manidoraisamy.com/developer-forever/post/the-butterfly-effect-in-agentic-workflows.anc-b855d56e-d986-44e1-ab65-b4e23142dc08.html