Introduction: The Challenge of Efficiency at Hyperscale

When your code serves over three billion people, even a tiny performance slip—say, 0.1%—can translate into massive extra power consumption. For Meta, keeping this in check is the job of the Capacity Efficiency Program. Traditionally, finding and fixing inefficiencies required armies of engineers manually sifting through regressions and hunting for optimization opportunities. That approach no longer scales.

To break through this bottleneck, Meta built a unified AI agent platform that encodes years of domain expertise from senior efficiency engineers into reusable, composable skills. These AI agents now automate both the detection and remediation of performance issues, recovering hundreds of megawatts (MW) of power—enough to power hundreds of thousands of American homes for a year—and slashing manual investigation time from hours to minutes.

A Two-Sided Efficiency Strategy: Offense and Defense

Meta views efficiency as a dual effort: offense (proactively finding optimizations) and defense (catching regressions before they compound). Both sides benefit from AI automation.

Offense: Proactive Optimization at Scale

On the offensive side, AI-assisted opportunity resolution expands to more product areas each half. The system identifies code changes that can improve performance, drafts ready-to-review pull requests, and accelerates deployment. This frees engineers from low-level optimization grunt work, letting them focus on innovation.

Defense: Rapid Regression Detection and Fixing

For defense, Meta relies on FBDetect, an in-house regression detection tool that catches thousands of regressions weekly. In the past, each regression required an engineer to manually investigate, root-cause, and fix it—a process that could take ten hours. Now, AI agents compress that into roughly 30 minutes, fully automating the path from detection to a ready-to-review pull request. Faster automated resolution means fewer megawatts wasted while regressions compound across the fleet.

How the Unified AI Agent Platform Works

The platform standardizes tool interfaces and encodes domain expertise into composable skills. Agents can share knowledge, reuse capabilities, and operate consistently across the entire infrastructure. This unified approach enables the Capacity Efficiency Program to scale megawatt delivery without proportionally increasing headcount.

Key Components

• Encoded Domain Expertise: Senior engineers’ knowledge is captured as reusable, composable skills that any agent can invoke.
• Standardized Interface: All tools, from FBDetect to optimization scanners, speak a common language, reducing integration overhead.
• Automated Investigation: Agents traverse performance data, identify root causes, generate fixes, and create pull requests—all without human intervention.

Results: Real-World Impact

The program has already recovered hundreds of megawatts of power. By automating the long tail of efficiency work, Meta is building a self-sustaining efficiency engine where AI handles routine fixes and engineers focus on breakthrough improvements. The goal is to keep growing MW delivery without proportionally growing the team.

Looking Ahead

Meta plans to extend AI-assisted opportunity resolution to even more product areas, further reducing manual effort. The platform’s modular design allows new skills to be added easily, enabling the system to tackle increasingly complex efficiency challenges. Ultimately, the vision is a fully autonomous efficiency loop: detect, diagnose, fix, deploy—all driven by AI agents.

For more details, explore Meta’s internal efficiency blog or learn about offensive optimization and defensive regression management.