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

Amazon Web Services (AWS) announced that Amazon Bedrock AgentCore Observability now supports monitoring AI agents deployed on-premises or in multi-cloud environments, including Google Cloud Platform (GCP) and Microsoft Azure, in addition to AWS services. The solution uses the AWS Distro for OpenTelemetry (ADOT) auto-instrumentation to route telemetry to the AgentCore Observability dashboard, enabling detection of hallucinations, monitoring of harmful responses, and tracking token usage for cost governance.

Artificial Intelligence /blogs/machine-learning/ Monitor on-premises and multi-cloud AI agents with AgentCore Observability When you deploy AI agents built with frameworks like Strands Agents, LangGraph, and CrewAI, you need observability into their performance. This holds true whether they run on Amazon Elastic Kubernetes Service Amazon EKS , Amazon Elastic Container Service Amazon ECS , AWS Lambda, on-premises, or another cloud provider such as Google Cloud Platform GCP or Microsoft Azure. Amazon Bedrock AgentCore https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/what-is-bedrock-agentcore.html is a platform to build, connect, and optimize agents at scale, with any framework or model. Although Amazon Bedrock AgentCore Observability https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/observability.html , a capability of Amazon Bedrock AgentCore, provides native tracing, monitoring, and analytics that local cloud monitoring tools don’t offer out of the box, it natively supports only agents deployed on AgentCore runtime in the AWS Cloud. If your agents run anywhere else, you need additional configuration to send telemetry to the dashboard. In this post, we show you how to set up observability for agents running outside AWS. You learn how to configure the AWS Distro for OpenTelemetry ADOT auto-instrumentation in non-AWS environments, route telemetry to the AgentCore Observability dashboard, and validate the setup end to end. The following diagram shows the end-to-end observability pipeline and how telemetry flows from agents to the AgentCore Observability dashboard. Solution overview The solution uses the AWS Distro for OpenTelemetry ADOT https://aws-otel.github.io/ running in-process with the agent application. ADOT auto-instruments the agent framework and captures generative AI semantic convention spans, then exports the telemetry directly to the Amazon CloudWatch OpenTelemetry Protocol OTLP endpoint using SigV4 authentication with AWS Identity and Access Management IAM credentials. Sending telemetry from your AI agent to Amazon Bedrock AgentCore Observability requires three core components: ADOT auto-instrumentation : The AWS Distro for OpenTelemetry handles the complexities of exporting telemetry from non-AWS environments. IAM credentials : The ADOT uses these access keys to authenticate with CloudWatch and forward your agent’s telemetry traces, metrics, and logs to the AgentCore Observability dashboard. Environment variables : These contain specific OpenTelemetry settings related to routing and authentication. As seen in the following diagram, this cross-platform observability solution integrates several AWS services. Amazon CloudWatch serves as the foundation, handling telemetry ingestion and storage. Amazon Bedrock AgentCore Observability adds specialized monitoring dashboards for AI agents. AWS Distro for OpenTelemetry ADOT provides the cross-platform instrumentation capabilities. IAM secures the authentication between your external environments and AWS. Observability is a foundational pillar of responsible AI. By routing telemetry to AgentCore Observability, you gain visibility into agent reasoning chains, tool invocations, and model outputs. This allows you to detect hallucinations, monitor for harmful or off-topic responses, track token usage for cost governance, and audit agent behavior across environments. This is especially critical for agents running outside AWS, where problematic outputs might go unnoticed without centralized observability. Prerequisites Before you begin, verify that you have: - An AWS account: - with Amazon Bedrock model access configured this walkthrough uses Claude Haiku https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-haiku-4-5.html . For model availability by AWS Region, refer to supported models by AWS Region https://docs.aws.amazon.com/bedrock/latest/userguide/models-region-compatibility.html in Amazon Bedrock. - for designated AgentCore Observability and designated log group s . - with Amazon Bedrock model access configured this walkthrough uses CloudWatch Transaction Search https://docs.aws.amazon.com/AmazonCloudWatch/latest/monitoring/Enable-TransactionSearch.html turned on in your account one-time setup - Python 3.10 or later installed on your non-AWS environment. - IAM user credentials access key ID and secret access key with permissions for: bedrock:InvokeModel . logs:CreateLogGroup , logs:CreateLogStream , logs:PutLogEvents . xray:PutTraceSegments , xray:PutTelemetryRecords , xray:GetSamplingRules , and xray:GetSamplingTargets . cloudwatch:PutMetricData . - Outbound HTTPS access to AWS endpoints from your environment. Turn on CloudWatch Transaction Search If you haven’t turned on Transaction Search, run the following one-time per account : Verify it’s active: How it works The ADOT auto-instrumentation aws-opentelemetry-distro handles the complexity of exporting telemetry from non-AWS environments to CloudWatch: Auto-instrumentation : The opentelemetry-instrument command injects the ADOT into the Python runtime. It automatically patches boto3 for Amazon Bedrock calls and the Strands framework for agent reasoning spans to emit OpenTelemetry traces. SigV4 authentication : The aws configurator uses the boto3 credential chain to sign OTLP export requests with SigV4. From non-AWS environments, this uses the AWS ACCESS KEY ID and AWS SECRET ACCESS KEY environment variables. CloudWatch OTLP endpoint : The ADOT exports traces and logs to the CloudWatch native OTLP ingestion endpoint. The OTEL EXPORTER OTLP LOGS HEADERS header directs logs to the specific AgentCore log group, which is how CloudWatch indexes the data under the generative AI observability dashboard. For details on how the CloudWatch OTLP endpoint URL is determined and configured, see CloudWatch OTLP endpoint https://docs.aws.amazon.com/AmazonCloudWatch/latest/monitoring/CloudWatch-OTLPEndpoint.html . Generative AI semantic conventions : The Strands otel package emits spans following the OpenTelemetry generative AI semantic conventions https://opentelemetry.io/docs/specs/semconv/gen-ai/ , including agent reasoning steps, tool invocations, and model calls with token usage. The following diagram shows how telemetry export through ADOT auto-instrumentation works from non-AWS environments to CloudWatch. Walkthrough Follow these steps to configure and run a Strands agent in a non-AWS environment, with telemetry routed to AgentCore Observability. Step 1: Install dependencies On your non-AWS environment on-premises server, GCP VM, Azure VM, or a compute with internet access : The aws-opentelemetry-distro package includes the ADOT auto-instrumentation with OTLP exporters specific to AWS and the aws configurator that handles SigV4 authentication. The strands-agents otel package provides OpenTelemetry trace emission from the Strands framework. Step 2: Configure AWS credentials Set your IAM user credentials as environment variables. Security note : For production deployments, consider using IAM Roles Anywhere https://docs.aws.amazon.com/rolesanywhere/latest/userguide/introduction.html instead of long-lived access keys. With IAM Roles Anywhere, on-premises workloads can obtain temporary credentials using X.509 certificates. Step 3: Set OpenTelemetry environment variables These environment variables configure the ADOT to route telemetry to the AgentCore Observability https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/observability-get-started.html enabling-observability-non-runtime-hosted dashboard: Key configuration details: AGENT OBSERVABILITY ENABLED=true activates generative AI-specific telemetry processing in the ADOT. OTEL PYTHON DISTRO=aws distro and OTEL PYTHON CONFIGURATOR=aws configurator activate the OpenTelemetry configuration specific to AWS, including SigV4 signing for the CloudWatch OTLP endpoint. OTEL RESOURCE ATTRIBUTES with aws.log.group.names tells CloudWatch to index the telemetry under the AgentCore Observability dashboard. Without this, traces go to generic Amazon CloudWatch Logs. OTEL EXPORTER OTLP LOGS HEADERS with x-aws-metric-namespace=bedrock-agentcore routes metrics in embedded metric format https://docs.aws.amazon.com/AmazonCloudWatch/latest/monitoring/CloudWatch Embedded Metric Format.html to the correct CloudWatch namespace. Step 4: Create the agent application Create a file named agent test.py with a Strands agent: Step 5: Run with ADOT auto-instrumentation The opentelemetry-instrument command wraps your Python process with the ADOT, automatically instrumenting Amazon Bedrock calls and Strands framework operations: The agent’s response appears in the terminal. Behind the scenes, the ADOT captures traces, spans, and logs, and exports them to CloudWatch. Step 6: Verify in AgentCore Observability You see telemetry data within two to three minutes of execution. Open the Amazon CloudWatch console: - Choose GenAI Observability , then Bedrock AgentCore . - In the Agents tab, look for my-external-agent . - Choose the agent to view sessions, traces, and span metrics. The following screenshot shows the telemetry from the Strands agent my-external-agent running in a non-AWS environment, as seen in the AgentCore Observability dashboard in CloudWatch. The console shows: Agent name : my-external-agent . Sessions : at least one session. Traces : trace spans showing the agent’s reasoning and Amazon Bedrock model invocations. Span details : invoke agent , chat , execute event loop cycle , and chat.us.anthropic.claude-haiku spans with latency and token metrics. The following screenshot shows a successful trace from the Strands agent my-external-agent with four spans, model information, and latency and token details in the AgentCore Observability dashboard. Validating from Google Cloud Platform To confirm the solution works from a third-party cloud provider, we tested the same setup from Google Cloud Shell, a browser-based terminal running on GCP infrastructure. Set up the environment on Google Cloud Shell: Run the agent from GCP: The following screenshot shows the Strands agent gcp-hosted-agent running on Google Cloud Shell GCP and returning a successful response. Verify cross-cloud telemetry Within two to three minutes of execution, the gcp-hosted-agent appears in the AgentCore Observability dashboard alongside agents running on AgentCore runtime or other environments. The following screenshot shows a successful trace from the Strands agent gcp-hosted-agent running on GCP with four spans, model information, and latency and token details in the AgentCore Observability dashboard. The telemetry is identical to what an AgentCore runtime-hosted agent produces. Sessions, traces, span metrics, token usage, and latency are all visible in the same dashboard, regardless of where the agent runs. Although this walkthrough uses Strands Agents, the same ADOT-based pattern applies to other OpenTelemetry-compatible agent frameworks. When choosing how to deploy your AI agents, understanding the observability trade-offs between different runtime environments helps you make the right architectural decision. Agents deployed directly on Amazon Bedrock AgentCore runtime benefit from automatic observability configuration. Agents running in non-AWS environments require additional manual setup but offer greater deployment flexibility. The following comparison highlights the key differences in telemetry collection, credential management, and use cases to help you determine the best approach for your requirements. Aspect | Non-AWS Runtime | AgentCore runtime | | Telemetry supported | ADOT – manual OTEL variables required | ADOT – Automatic in-built OTEL variables | | Credential management | IAM access key/secret or IAM Roles Anywhere | Automatic IAM role | | Best for | Agents on-premises, GCP, Azure, or a non-AWS environment | Agents deployed on AWS with AgentCore | Validated environments We tested the ADOT auto-instrumentation approach across two non-AWS environments: Environment | Platform | Result | | On-premises simulated | Standalone server running in non-AWS environment | Strands agent reporting telemetry sessions, traces, spans in AgentCore Observability | | Google Cloud Shell GCP | Browser-based terminal running on Google Cloud Platform | Strands agent reporting telemetry sessions, traces, spans in AgentCore Observability | Best practices Based on our testing, we recommend the following when setting up cross-platform AgentCore Observability: Use consistent naming : The service.name in OTEL RESOURCE ATTRIBUTES becomes the agent name on the dashboard. Use descriptive names that identify the environment for example, prod-onprem-support-agent and staging-gcp-research-agent . Verify with : Before running the agent, confirm that your credentials work by running get-caller-identity first python -c "import boto3; print boto3.client 'sts' .get caller identity " . If this fails, the ADOT also fails silently. Use Python 3.10 or later : The ADOT requires Python 3.10 or later. We recommend Python 3.12 for the best compatibility with all dependencies. Set session IDs for multi-turn conversations : Use the OpenTelemetry baggage API to propagate session IDs: Rotate credentials regularly : For production deployments, avoid long-lived access keys. Consider IAM Roles Anywhere https://docs.aws.amazon.com/rolesanywhere/latest/userguide/introduction.html for on-premises workloads, or use your cloud provider’s identity federation to assume AWS IAM roles. Clean up To remove the resources created during this walkthrough: This walkthrough uses Amazon Bedrock, Amazon CloudWatch, and AWS X-Ray, which incur costs. See the respective pricing pages for details. Conclusion Amazon Bedrock AgentCore Observability isn’t limited to agents running on AgentCore runtime or within AWS. Using ADOT auto-instrumentation with IAM credentials and the correct OpenTelemetry environment variables, you can send telemetry from your choice of environment with internet access. Your agents can run on-premises, on GCP, on Azure, or anywhere else and still report to the same AgentCore Observability dashboard. The setup requires a pip install and a set of environment variables. The resulting telemetry is identical to what AgentCore runtime-hosted agents produce: sessions, traces, span metrics, and token usage, all in one unified view. To get started, clone the sample code from GitHub https://github.com/aws-samples/sample-external-agents-telemetry-on-agentcore-observability and follow the instructions in the README to configure and run the agent in your environment. For agents already running on AWS but outside AgentCore runtime EKS, ECS, Lambda , refer to the AgentCore Observability for EKS-hosted agents https://github.com/awslabs/agentcore-samples/tree/main/03-integrations/agents-hosted-outside-runtime tutorial. For agents on AgentCore runtime, observability is configured automatically. See Add observability to your AgentCore resources https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/observability-configure.html .