AWS AgentCore Observability now monitors AI agents across on-prem and multi-cloud

Amazon Bedrock AgentCore Observability, a capability of Amazon Bedrock AgentCore, delivers native tracing, monitoring, and analytics for AI agents—but until now, it only supported agents running on the AgentCore runtime in AWS. This blog post demonstrates how to extend that observability to agents deployed anywhere, including on-premises, Google Cloud Platform, Microsoft Azure, Amazon EKS, Amazon ECS, or AWS Lambda. The solution leverages AWS Distro for OpenTelemetry (ADOT) running in-process with the agent application. ADOT auto-instruments agent frameworks like Strands Agents, LangGraph, and CrewAI, capturing generative AI semantic convention spans and exporting them directly to the Amazon CloudWatch OTLP endpoint using SigV4 authentication with IAM credentials.

Setting up the pipeline requires three core components: ADOT auto-instrumentation to handle telemetry export from non-AWS environments, IAM credentials to authenticate and forward traces, metrics, and logs to the AgentCore Observability dashboard, and environment variables for OpenTelemetry routing and authentication configuration. Prerequisites include an AWS account with Bedrock model access (the walkthrough uses Claude Haiku), enabling CloudWatch Transaction Search, Python 3.10+, and IAM permissions for Bedrock invocation, CloudWatch Logs, X-Ray, and CloudWatch metrics. Once configured, teams gain centralized visibility into agent reasoning chains, tool invocations, and model outputs. This enables detection of hallucinations, monitoring for harmful or off-topic responses, tracking token usage for cost governance, and auditing agent behavior across all deployment locations—critical for responsible AI when agents run outside AWS where problematic outputs might otherwise go unnoticed.