orchestration · the architecture hole
You're not closing one hole; you're designing the agent loop. A one-shot build passes the instances that dodge its blind spot and bleeds the rest; naming the rules lifts the floor; a verify-and-retry loop reaches back and recovers the tail. Generalization is the skill — and a capable model doesn't save the one-shot, because a frozen prompt can't adapt to inputs it never saw.
What good looks like
- Compose a pipeline of move-cards; freeze it.
- It runs unchanged across a hidden distribution sample.
- Score = generalization across instances, modulated by build efficiency.
The symptom
- One-shot decay: a frozen build only clears the instances that miss its blind spot.
- Decompose: name the recency rules and the floor jumps for the whole distribution.
- Verify + retry: read each failure and recover the residual tail, instance by instance.
- Efficiency: thrash-retrying without aiming costs builds it can't earn back.
Caddie
- Name the rules up front — decompose is the single biggest score mover.
- Add a verify gate whose probes cover the aspects the distribution stresses.
- Budget a retry that feeds the exact failure back — recover, don't thrash.
- Don't over-retry: efficiency gates a high score you didn't aim.