You can't debug what you can't see
When a user reports 'the agent gave me a wrong answer' or 'it got stuck,' that report alone is nearly useless for debugging. Was it the plan, a bad tool call, a misread result, or a hallucinated fact? Without a record of the agent's actual steps, reasoning, tool calls, and their results, you're left guessing and trying to reproduce a problem you can't actually see. Observability is what turns a silent failure into a diagnosable one. It's the difference between staring at a bug report and staring at an execution trace.
Production agents should log every step of every run: which tool was called and with what arguments, what the tool returned, what the agent's intermediate reasoning or plan update was, and how long each step took. This isn't optional polish, it's the foundation everything else in this lesson depends on. Evaluation, cost control, and human review all need a trace to work from.