Developer Uses OpenTelemetry and SigNoz to Catch AI Agent Calling Same Tool Repeatedly

A developer building a ReAct-style AI agent noticed that some queries took unusually long or required multiple attempts to resolve, despite returning correct answers. To investigate, they instrumented the agent's full decision loop using OpenTelemetry and the SigNoz observability platform, going beyond standard request-level tracing to capture every LLM call and tool invocation. Within 30 minutes of setup, the traces revealed that the model was calling a calculator tool three times in a row with identical inputs before delivering a final answer. A separate query showed the model sending malformed action responses twice before correcting itself on the third attempt. The exercise demonstrated that correct outputs alone can mask inefficient or redundant internal behavior that only distributed tracing can expose.