An element of this that I'm still exploring is compounding errors. If interaction 1 is wrong, the error can become part of the context of interaction 2. There's *causality* involved, and the amount of unvalidated causality - feedback latency - becomes a dominant factor, I suspect.
@jasongorman.bsky.socialIf you're interested in the design of agentic workflows, remember this formula:
R = 1 - (1 - P)(1 - C)
