This actually matches my own experience very close to 100%: static pipelines work much much better than orchestrators for specific automation goals.

I think this is because the orchestrator is extremely vulnerable as every little issue gets amplified in the overall flow. Hallucination, poisoned context (or context overflow!), and even token triggers from feedback at the central point will get exaggerated throughout the flow. The risk of strange results multiplies with each call. So calls should be kept to a minimum.