A practical walkthrough of instrumenting a LangGraph agent with Datadog's AI Agent Monitoring and LLM Observability SDK. The sample agent uses Claude Sonnet via Amazon Bedrock, Tavily for web search, and Amazon SNS for output routing. The guide covers: configuring LLM Observability to capture traces, flame graphs, and token usage; analyzing latency bottlenecks and cost across agent runs; running automated LLM-as-a-judge evaluations for output quality and prompt injection detection; and correlating agent traces with APM, logs, and infrastructure metrics to debug failures across the full application stack.
Table of contents
Our sample LangGraph agentConfigure AI Agent Monitoring for a LangGraph agentTrace LangGraph tool calls and LLM latencyAnalyze latency, cost, and errors across agent runsEvaluate agent output quality with LLM-as-a-judgeCorrelate LangGraph agent traces with APM, logs, and infrastructure dataGet started with Datadog AI Agent Monitoring135 Impressions