Monitor on-premises and multi-cloud AI agents with AgentCore Observability

The post explains how to set up observability for AI agents running outside AWS using ADOT auto-instrumentation. The solution routes telemetry to the Amazon CloudWatch OTLP endpoint with SigV4 authentication, using IAM credentials. The setup requires three components: ADOT auto-instrumentation, IAM credentials, and environment variables. The walkthrough uses a Strands agent, but the pattern applies to other OpenTelemetry-compatible frameworks like LangGraph and CrewAI. The solution was validated in an on-premises simulated environment and on Google Cloud Shell. The console displays sessions, traces, span metrics, token usage, and latency in the AgentCore Observability dashboard. The post also compares telemetry collection and credential management between non-AWS runtimes and AgentCore runtime, and provides best practices such as consistent naming and verifying credentials with get-caller-identity.