Agent observability
Also known as: llm observability, ai observability, agent tracing, agent monitoring
Traditional software observability tracks errors and latency. Agent observability adds a layer specific to AI systems: capturing the sequence of model calls, tool invocations, token usage, reasoning traces, and intermediate outputs that make up an agent run. Without this, debugging a failed agent run is extremely difficult because failures are often reasoning failures rather than code errors.
Tools like LangSmith, Langfuse, Helicone, and Google's Agent Observability feature visualize agent execution as structured traces. You can see exactly which tool call returned unexpected output, where the model's reasoning went off track, and which step consumed most of the tokens. That visibility makes the difference between shipping improvements in hours versus days of log archaeology.
In 2026, agent observability became a standard expectation for production agent deployments. Builders running background agents especially need it: when an agent runs for an hour in an isolated VM and returns an incorrect result, the trace is the only way to understand what went wrong. Common metrics tracked include task completion rate, token cost per run, latency by step, tool call success rate, and human intervention rate.