Implementing synthetic monitoring using Amazon Nova Act

AWS published a technical post describing an agent-driven approach to synthetic monitoring using Amazon Nova Act and Amazon Bedrock AgentCore, with a sample repository containing a complete implementation.

The post frames the problem around traditional browser automation frameworks such as Selenium and Playwright, which rely on explicit DOM locators and selectors. According to the post, such scripts are brittle: small UI changes can break tests and require constant maintenance. Amazon Nova Act instead uses a multimodal large language model that processes UI screenshots rather than DOM selectors. The post states that, in early enterprise customer use cases, Nova Act has demonstrated over 90% accuracy on browser workflows, and advises teams to test on their own sites and design monitors with retry logic for cases where adaptation does not succeed.

The described architecture combines several AWS services. Amazon Nova Act defines and executes UI workflows through natural-language actions. Amazon Bedrock AgentCore Runtime provides serverless execution with session isolation and stable invocation endpoints, and the AgentCore Browser tool provides an isolated remote browser environment per test run. Amazon EventBridge Scheduler triggers runs on a defined schedule, stated as every 5 minutes to hourly depending on workflow criticality, calling InvokeAgentRuntime directly. Failures are published through Amazon SNS to subscribed endpoints such as email, chat integrations and incident response tooling.

The post lists prerequisites including an AWS account with access to the relevant services, Python 3.11 or later, Docker, AWS CLI v2, and Node.js 18 or later with the AWS CDK for the production infrastructure path. Deployment uses Nova Act CLI act workflow commands, or a CDK stack that provisions the schedule, SNS topic, an SQS dead-letter queue and CloudWatch alarms. A separate deploy.py script is also provided.

Operational guidance includes starting with 3-5 critical workflows and validating meaningful outcomes rather than every DOM element. The sample agent uses single-attempt execution per step to minimize Browser session cost; the post notes this may occasionally produce false alerts because Nova Act adaptation succeeds in approximately 90% of cases, and that step-level retry can lower false positives at the cost of longer session durations.

Costs are attributed to five components: AgentCore Runtime invocations, Browser tool sessions, Nova Act inference, EventBridge Scheduler invocations and SNS notifications. The post gives a cost approximation for a six-step ecommerce journey running every 5 minutes, approximately 8,640 invocations per month, and directs readers to pricing pages.

Security details include AgentCore Runtime session isolation using a one-session-one-microVM model built on Firecracker microVMs, with full memory sanitization by default. The post recommends ephemeral sessions for synthetic monitors. It also describes IAM integration for least-privilege permissions, AWS Secrets Manager for test credentials with audit through AWS CloudTrail, multi-Region deployment using infrastructure as code, and cleanup commands for both the deploy.py and CDK paths.