I Gave My Agent a Deployment Tool. It Could Finally Answer.

In a recent experiment, a reporter discovered a method to determine if a production environment is healthy. Initially, their CloudOps agent was unable to provide an answer to this question. However, by granting the agent access to deployment data, they were able to obtain detailed information about the service, deployed version, instance count, active alarms, and observation time. This was made possible by providing a controlled way for the agent to query another system for evidence.

The setup was deliberately kept simple, with the user posing the question "Is production healthy?" to the AgentCore. The AgentCore then utilized a harness to select a method called get_deployment_status. This method, in turn, invoked a registered target using AWS Lambda, which returned a deterministic snapshot of the deployment status.

To ensure the accuracy of the agent's responses, the reporter defined what constituted a healthy environment in code. The Lambda function compared three conditions: whether the desired and running versions matched, if the number of healthy instances equaled the desired instances, and if there were no active alarms. The function returned "healthy" if all conditions were met, and "unhealthy" otherwise.

The reporter also provided specific instructions for the agent, such as only accepting the production environment and rejecting any unsupported environments. To facilitate testing, the reporter utilized AWS Serverless Application Model (SAM) and Docker to create a local Lambda test environment. Initially, the local attempt reported an error regarding Docker compatibility, but after updating SAM, the function was successfully invoked and returned a detailed snapshot of the production environment's health status.