Capacity Efficiency at Meta: How Unified AI Agents Optimize Performance at Hyperscale

Meta's Capacity Efficiency Program uses a unified AI agent platform to automate both finding and fixing performance issues at hyperscale. The platform is built on two layers: MCP Tools (standardized LLM interfaces for querying profiling data, code, configs, etc.) and Skills (encoded domain expertise from senior engineers). On defense, the AI Regression Solver integrates with FBDetect to automatically generate pull requests that fix performance regressions, compressing ~10 hours of manual investigation into ~30 minutes. On offense, AI agents turn efficiency opportunities into ready-to-review pull requests. Both sides share the same tool interfaces, with only the skills differing. The program has recovered hundreds of megawatts of power and continues to scale without proportionally growing headcount.