Really like the framing of "Time to Accountability" as the bottleneck. Curious how you think about this for the orchestration layer specifically. A lot of agent frameworks give you observability into the graph. You can see which node fired, what tools were called. But the reasoning that led to those decisions is still opaque. Is the answer better tracing/logging of model decisions? Or more constrained orchestration logic?