Capacity Efficiency at Meta (engineering.fb.com)

Meta has unveiled its Capacity Efficiency Program, introducing a unified AI agent platform designed to automate the identification and resolution of performance issues within its infrastructure. By leveraging a combination of standardized tool interfaces and encoded domain expertise from senior engineers, these AI agents have drastically reduced the time required for manual regression investigations—from 10 hours to approximately 30 minutes—while recovering hundreds of megawatts (MW) of power. This innovative dual approach to efficiency integrates offensive strategies, which proactively seek optimization opportunities, with defensive measures that catch regressions, ensuring a virtually self-sustaining efficiency engine. The significance of this initiative lies in its potential to fundamentally transform engineering workflows at Meta, allowing engineers to focus on innovation rather than troubleshooting. The technology employs tools such as FBDetect for regression detection and AI Regression Solvers for automating fixes, streamlining the process and enhancing overall engineering velocity. By systematically utilizing AI in both the offensive and defensive arenas of performance management, Meta not only improves its power efficiency but also reinforces a culture of continuous improvement, where AI assists in executing complex tasks that would traditionally take much longer, thus supporting a scalable and sustainable approach to software efficiency at hyperscale.