Meta's Capacity Efficiency Program has been using AI agents to automate the detection and resolution of performance issues across its massive infrastructure. This Q&A explores how the program works, the role of AI agents, and the impact on power savings and engineering productivity.
What Is Meta's Capacity Efficiency Program?
The Capacity Efficiency Program at Meta focuses on optimizing performance across the entire infrastructure to reduce power consumption and free up engineering resources. It operates on two fronts: offense and defense. Offense involves proactively identifying opportunities to make existing systems more efficient through code changes. Defense uses tools like FBDetect to monitor production for regressions—small performance drops that, at Meta's scale, can waste significant power. By automating both sides with AI agents, the program has already saved hundreds of megawatts of power, enough to supply hundreds of thousands of homes for a year.
How Do Unified AI Agents Enhance Efficiency at Hyperscale?
Meta built a unified AI agent platform that encodes the domain expertise of senior efficiency engineers into reusable, composable skills. These agents can automatically investigate and resolve performance issues. For example, a single agent might analyze a regression detected by FBDetect, root-cause the problem to a specific pull request, and even prepare a fix. The platform standardizes tool interfaces, making it easy to chain agents together. This approach compresses what used to be up to 10 hours of manual investigation into roughly 30 minutes, and many fixes are now fully automated from opportunity to ready-to-review code.
What Role Do Offense and Defense Play in the Program?
Efficiency at Meta is a two-sided effort. Defense catches regressions that slip into production using FBDetect, which flags thousands of issues weekly. Faster automated resolution means fewer megawatts wasted over time. Offense uses AI agents to proactively find and implement efficiency improvements across more product areas each half-year. While human engineers previously had to manually handle each win, AI now takes on a growing volume of optimizations that would otherwise be unattainable. This dual approach ensures continuous power savings even as the infrastructure expands.
How Much Time and Power Does the AI Automation Save?
The results are dramatic. AI agents can compress roughly 10 hours of manual regression investigation into about 30 minutes—a 20x time savings. By quickly addressing regressions, the program recovers hundreds of megawatts of power that would have been lost to inefficiencies. Additionally, the automation allows the Capacity Efficiency team to scale MW delivery across more product areas without proportionally increasing headcount. Engineers are freed from repetitive debugging to focus on innovation. In essence, the AI agent platform turns hours of human work into minutes of automated processing, delivering both time and energy savings.
What Tools Underpin the AI Agent Platform?
Key tools include FBDetect, Meta’s in-house regression detection system that monitors production resource usage. It catches thousands of regressions every week. The AI agents then use FBDetect's output as input for automated root cause analysis and mitigation. Another critical component is the unified agent platform itself, which combines standardized tool interfaces with encoded domain expertise. Agents can call various internal APIs and services to trace issues to specific code changes. The platform is designed to be extensible, so new skills can be added as new types of efficiency opportunities emerge.
What Is the Ultimate Goal of the Capacity Efficiency Program?
The long-term vision is a self-sustaining efficiency engine: an AI-driven system that handles the long tail of performance issues automatically, without requiring proportional growth in the team. Currently, the program already recovers hundreds of megawatts of power and compresses investigation times drastically. The goal is to expand AI-assisted opportunity resolution to more product areas each half-year, making the program more autonomous. By doing so, Meta aims to keep delivering increasing MW savings while maintaining a small, highly skilled team that focuses on high-level strategy and new innovations, rather than routine debugging.