MENLO PARK, CA — Meta has deployed a new AI agent platform that automatically identifies and resolves performance issues across its massive infrastructure, recovering hundreds of megawatts of power and slashing investigation times from hours to minutes, the company announced today.
The system, built as part of Meta’s Capacity Efficiency Program, encodes the domain expertise of senior efficiency engineers into reusable, composable skills. These AI agents now handle both proactive optimization and regression detection at hyperscale, enabling the program to scale without adding proportional headcount.
“This is a paradigm shift,” said a Meta efficiency engineer speaking on condition of anonymity. “What used to take ten hours of manual digging now takes about thirty minutes, and the agents are getting faster every quarter.”
How It Works: Offense and Defense
Meta frames efficiency as a two-sided effort. On the offensive side, AI-assisted opportunity resolution scans codebases for proactive changes that can make systems more efficient, then automatically generates ready-to-review pull requests. On defense, the in-house tool FBDetect catches thousands of regressions weekly.
“Those regressions, if left unchecked, compound across the fleet and waste precious megawatts,” the engineer added. “Our AI now automates the path from detection to mitigation, compressing days into minutes.”
Background
The Capacity Efficiency Program has been central to Meta’s operations for years, but manually resolving issues created a bottleneck as infrastructure grew. The unified AI agent platform solves this by standardizing tool interfaces and encoding domain expertise into a common system.
Meta says the program has already recovered enough power to run hundreds of thousands of American homes for a year. The ultimate goal, the company notes, is a self-sustaining efficiency engine where AI handles the long tail of performance bugs.
What This Means
Industry analysts see Meta’s approach as a template for other hyperscalers facing similar scaling challenges. “Automating efficiency at this level isn’t just about cost savings—it’s a competitive necessity,” said Dr. Lena Park, a data center efficiency researcher at Stanford University. “Meta is showing that AI can close the loop between detection and action in ways that were previously impossible.”
The platform also frees engineers from grunt work, allowing them to focus on innovation. “When you serve over 3 billion people, even a 0.1% regression matters,” the Meta engineer said. “Now our engineers can spend their time inventing the next generation of products instead of chasing down bugs.”
Meta plans to expand the AI agent platform to more product areas every half-year, further scaling the program’s impact without adding proportional headcount. The company expects the system to become increasingly autonomous over time, learning from each resolved issue.