Debugging a classic application means reproducing the same fault with the same input. In agent systems that guarantee is gone, and the monitoring approach has to change with it.
01. Trace Records
A full trace must be kept for every run: which prompt went out, what the model returned, which tool was called with which parameters, what came back. Without that record, investigating a fault is impossible.
02. Step-Level Visibility
If the agent took several steps, each must be visible separately. A single "failed" record doesn't say at which step the fault occurred or why.
03. The Multi-Agent Difficulty
When dozens of specialist agents run in parallel and in sequence, monitoring designed for a single agent falls short. Which agent triggered which, and how total cost is distributed, need tracking too.
04. Metrics to Watch
Steps per run, token cost, duration, tool error rate, human intervention rate and task completion rate. Looking only at error count is misleading.
05. Silent Failure
The most dangerous case is an agent doing the wrong thing without raising an error. Catching it takes automated checks on output plus regular sampled human review.
06. Alert Thresholds
Set alerts when average step count or cost jumps. That's usually the first sign of a prompt change or a broken tool.