A multi-agent customer-onboarding system intermittently takes over a minute to respond. Each service logs its own events, but logs can't be joined across the orchestrator, three subagents, two MCP servers and the CRM. The team wants to find the cause. What is the best first investment?

Answer: B — Propagate one trace ID from the entry point through all agents, tool servers and the CRM, with a span per step

The problem is that events can't be connected, so a propagated trace with spans per step is what reveals which step in which request is slow. (a) produces more disconnected data plus cost and privacy exposure. (c) shows averages per service and hides intermittent outliers and cross-service chains. (d) is a change made without knowing where the time goes.