AI agent observability applies the ideas of AI observability to agents that act over many steps. For each run, a trace records the prompts and model calls, each tool the agent invoked and its arguments, the documents or rows it retrieved, the decisions in between, and the final action. Teams add latency, cost, and error data to the same trace. Open standards such as the OpenTelemetry semantic conventions for generative AI let different tools record these traces in a common format.
Traces help a team answer why an agent produced a specific result. When a support agent issues the wrong refund, the trace shows whether it read an outdated policy, called the wrong tool, or misread a correct record. Compared with LLM observability, which centers on single model calls, agent observability follows the whole chain. Recording which data version the agent read makes a past run easier to reproduce.