Monitor on-premises and multi-cloud AI agents with AgentCore Observability
Amazon Web Services Vipul Rajendra Gargav ● Covered by 2 sources
AWS showed how to watch AI agents outside AWS with AgentCore Observability. It uses ADOT, CloudWatch, and IAM so even GCP or on-prem agents show up in one dashboard.
Based on reporting by Amazon Web Services, Vipul Rajendra Gargav — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AWS is pushing AgentCore Observability beyond its natural home. The feature already traces and monitors agents on AgentCore runtime in AWS, but this walkthrough shows how to feed it telemetry from Strands Agents running on EKS, ECS, Lambda, on-premises servers, and even Google Cloud or Azure.
The trick is ADOT, AWS Distro for OpenTelemetry, running alongside the agent. It auto-instruments the Python app, patches boto3 and the Strands framework, and sends traces, logs, and metrics to CloudWatch’s OTLP endpoint with SigV4 signing. From there, AgentCore Observability can surface the agent’s reasoning chains, tool use, model calls, token usage, and other spans in one place.
The setup leans on a few moving parts: AWS credentials, OpenTelemetry environment variables, and CloudWatch Transaction Search turned on. AWS also calls out a production wrinkle here: on-premises deployments should consider IAM Roles Anywhere instead of long-lived access keys.
The example in the post uses Claude Haiku in us-east-1 and a simple travel assistant agent. With opentelemetry-instrument wrapping the Python process, the agent runs normally while telemetry appears in the console a couple of minutes later under a named agent such as my-external-agent or gcp-hosted-agent.
AWS tested the same pattern from Google Cloud Shell on GCP, and the result was the same dashboard view: sessions, traces, span metrics, latency, and token details. The company’s pitch is straightforward. If your agents are spread across clouds, you can still pull their behavior into one observability plane without rewriting the framework layer.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of boring infrastructure work: less keynote glitter, more “please show me what the agent actually did.” OpenTelemetry keeps winning because nobody wants five different dashboards for one bot that may or may not be hallucinating. The real story is not the model, it’s the audit trail.
Read more about this at: Amazon Web Services