Observability for AI Agents

Strands Agents emits OpenTelemetry natively. No extra imports, no configuration. One method gives you cycles, token counts, tool timings, and duration. Then you turn on tracing and see the full decision path, step by step, with the tool inputs and outputs attached to each span. In this video we go over three layers of observability for AI agents: metrics (cycles, tokens, per-tool timing), traces (the full decision path with inputs and outputs on each span), and custom trace attributes (your business context stamped onto spans via hooks). All built with Strands Agents, which emits OpenTelemetry natively with zero config. Then I take the same agent to production on AgentCore, swap the console exporter for OTLP, and everything lands in CloudWatch GenAI Observability automatically. Demo is a travel agent that searches flights, checks weather, and books trips. Agent Toolkit for AWS: https://go.aws/46qZ0MX Read the Blog: https://go.aws/4yblpKf Github Repo: https://go.aws/46spZHV Follow AWS Developers! 🆇 X: https://go.aws/4zG2FnH 💼 LinkedIn: https://go.aws/45PEFAO 0:00 Your Agent is Running Blind 2:25 Layer 1: Metrics (Cycles, Tokens, Tool Timing) 8:01 Layer 2: Traces (The Path It Took) 12:42 Layer 3: Trace Attributes (Your Context on the Trace) 17:14 Production: AgentCore Observability + CloudWatch 24:00 Recap + What Comes Next (Evaluation) #AWS #Observability #StrandsAgents