> When I think about planning I think about mapping out a possibility space, identifying trees of dependencies, assigning priorities, and then solving for some

> When I think about planning I think about mapping out a possibility space, identifying trees of dependencies, assigning priorities, and then solving for some kind of weighted shortest path. but that's two things: planning and execution of a plan. The plan is the dependency graph (!), assignment of priorities and shortest path is execution. This is very important in the context of agents, autonomous or not. If you want the agent to self-correct it has to understand that there can be multiple start points or multiple end points (or both) to backtrack and pivot. And as long as it is glorified "next token predictor" it cannot really do that. Of course some tasks indeed are ifttt style linear-ish flows where next token prediction may prove to be adequate. However, if your agent is incapable of understanding non-linear flows, can it reasonably back off from one true way hen faced with such flow?