Field Notes: 25 seconds to detect and hand off. Nothing executes until a human says yes.

A Strands agent on Amazon Bedrock AgentCore detected a real CloudWatch alarm, determined its significance, and communicated the finding to a LangGraph agent via A2A. The entire process took an average of 25.3 seconds across five runs. The LangGraph agent proposed a fix, but it remained unexecuted until a human provided approval.

This approval process is enforced by a Cedar policy on an AgentCore Gateway, not by a simple prompt in a prompt. The setup involved two agents built on different frameworks, detecting a genuine CloudWatch alarm and relaying it to another agent over A2A. The remediation agent merely suggested an action, while execution required human approval, resource tagging, and a Cedar policy on an AgentCore Gateway.

This system's robustness was demonstrated by attempting to break it: without approval, the proposal was denied, and using a resource without the correct tags also resulted in denial. The approval process was tested successfully, with both a successful approval and a successful rollback. The entire flow—from detection to hand-off to approval—was measured to ensure accuracy, and the results were consistent, demonstrating the system's reliability.