How I evaluate agent handoffs, the part that actually breaks

Here is a failure that taught me where multi-agent systems really break. A four-agent setup, planner to researcher to critic to writer, passes CI at 0.91 on the final answer. A week later it is at 0.8