A workflow crosses multiple agents and services. Which observability design supports diagnosing one user's failed operation?

1. Only a monthly total request count
2. Only each agent's description
3. Independent random log ids without relationships
4. Propagated trace context with correlated agent and tool spans