Agent observability

Agent observability is the practice of measuring what an AI agent does inside one request: how many model calls and tool calls it makes, how long each takes, what the run costs and whether it finished. An agent decides its own steps at runtime, so one endpoint can take one second or twenty.

Agent observability covers the agent loop only. For the model calls, see LLM observability. For the whole stack, see AI observability.

How it works

Record each run as a workflow span, with every model call (chat) and tool call (execute_tool) as a child span. Then track four signals: workflow and tool duration, turns per request, cost per completed run, and tool errors. In our demo agent, shallow questions took about 1 model turn per request. Deeper ones took about 2.2 turns and 3.4 tool calls, against a cap of 3 turns.

Why it matters in production

A runaway loop raises cost and latency without raising an error, and an HTTP latency graph cannot show it. Turns per request and the ratio of tool calls to final answers expose the loop on a dashboard, so you can alert on it. A run that hits its turn cap costs money and returns nothing to the user.