How AI-Driven Kubernetes Optimization Reclaimed Millions

Salesforce's Hyperforce platform faced 47% idle Kubernetes capacity due to configuration sprawl across 8,000+ services. To address this, they built a Capacity Optimization Agent that uses LLMs for repository discovery and configuration parsing, but delegates all optimization decisions to a deterministic Integer Linear Programming solver. This hybrid approach overcomes the non-determinism and global reasoning limitations of pure LLM-based agents. The system generates pull requests with clear explanations and projected impacts, integrating into existing deployment pipelines to build service owner trust. The result is a closed-loop, automated workflow that reclaimed millions in infrastructure spend and is expanding to cover memory and HPA configurations.