What is Observability?
The ability to understand a system's internal state from its external outputs, primarily through logs, metrics, and traces. A system is observable when you can answer "what is happening?" and "why?" without modifying it or guessing. In AI and agentic systems, observability extends further: tracking model inputs, outputs, latencies, token usage, retrieval quality, and agent decision traces.
My perspective
In practice
Observability is how you catch problems before your users do. In traditional software, it means structured logs, dashboards, and distributed tracing. In AI systems, it adds a new layer: you need to trace which prompt was sent, what context was retrieved, how many tokens were used, which tool the agent called, and why it made the decision it did. Without that trace, debugging a failing agent is guesswork. The principles are the same as in any distributed system; the artifacts are different.