Why Agent Observability Is Hard
Metrics, logs, and traces were built for human operators. Autonomous agents demand a new observability primitive: the reasoning trace.
Classical observability answers three questions: what happened, where did it happen, and how long did it take? These are sufficient when a human interprets the output. They are insufficient when the consumer is another machine that must decide what to do next.
An agent observability system must answer a fourth question: why did the agent believe this was the right action? This requires reasoning traces, structured records of observations, hypotheses, evaluations, and decisions.
Reasoning traces are not just logs. They must be queryable, comparable, and evaluable. We need to ask whether, given the same situation, the agent would make the same decision, and if not, what changed.
Vigilo experiments with this primitive. It watches host-level signals directly and organizes them around episodes and decisions rather than services and spans, using a fast per-event alert path plus a slower correlating pass, not a traces/metrics/logs pipeline. The goal is to make agent-adjacent behavior inspectable, debuggable, and ultimately improvable.