I always thought that a clever AI agent system would create deterministic systems that had been explored randomly through user request. Like, Lovable should have a giant treasure of data that can be used to essentially create building blocks that can be put together for a fraction of the cost of having the coding agent live generate everything all at once. It only needs some sort of knowledge structure like a graph to be able to traverse things and then a predictive structure to show where a building block might need some custom thing that an AI agent could do. Or from an ML perspective, why ask an LLM to calculate the sentiment of some text, have an LLM write a code that calculates the sentiment using VADER (or whatever). I could be wrong about all of this or speaking in too much abstraction.