Tracking AI Agent Performance: Metrics, Tracing, and Root Causes of Failures

Standard server and application monitoring is insufficient for agents: an agent may return HTTP 200, consume a large number of tokens, and still provide the client with an incorrect response or enter an infinite loop.