Maybe It Is Not Yet Time To Bring Every AI Demo To Production (opens in new tab)

There is a sentence I keep hearing in AI engineering that sounds innocent, practical, and mature: “Just add a fallback provider.” Clean. Elegant. Wonderful. The kind of sentence that usually survives only until production starts touching it. Because in a demo, fallback means: Provider A fails, call Provider B. In production, fallback means something very different. Will Provider B interpret the prompt in the same way? Will it serialize the tool schema in the same way? Will it respect cache di...