A script fails loudly and stops. An agentic pipeline fails quietly and keeps going — down a queue nobody is watching, making decisions on stale assumptions. The difference between the two isn't intelligence; it's whether anyone can see what happened.
Trace the decision, not just the call
Traditional logging records that a function ran. That's not enough here. For each step you want the inputs it saw, the option it chose, why it chose it, and the shape of what it returned — linked by a run ID so a single case reads as one story instead of scattered lines. When a reviewer asks "why did this application get rejected," the answer should be a trace, not an archaeology dig.
Three signals that actually page someone
Most agent dashboards measure everything and alert on nothing useful. Watch three things: the rate of steps that fall back to human review (rising means the world changed), end-to-end run latency (long-running work hides stalls), and output-validation failures (the workflow catching itself). Everything else is context you pull after one of those three wakes you up.