Analysis of Salesforce's Kubernetes Infrastructure revealed that majority of containers utilize less than 20% of their requested CPU. A large percentage of the cloud spend is wasted on this under utilization. Resources for these services are defined in Helm charts that are scattered across thousands of repositories. This made optimizations risky and tedious. We built an autonomous AI agent, that optimizes these Helm charts safely at scale, based on real time utilization data. We iterate on context engineering and improve the accuracy of this agent to nearly 100%.