AI Agent Observability: Instrument for Correctness

AI agents can return incorrect or unsafe answers even when all infrastructure signals appear healthy. This means a 200 OK message, good performance, and being within budget do not guarantee a correct output. The problem is that traditional monitoring tools don't catch these "confidently wrong" answers. An agent's mistake doesn't typically cause an error or a crash. To fix this, the focus needs to shift from log lines to span trees. A span tree captures each step of an AI agent's process, including model calls, token usage, and latency. This provides visibility into the agent's reasoning. OpenTelemetry offers conventions for these spans, though they are still in development. The bottom line is that correctness is a critical fourth pillar of observability that must be intentionally instrumented. Without it, you are debugging a black box.