Boosting Hyperscale Efficiency with AI Agents at Meta

Meta's Capacity Efficiency Program leverages a unified AI agent platform to automate performance optimization at hyperscale. By encoding senior engineers' expertise into reusable skills, these agents both detect and resolve issues, saving hundreds of megawatts of power and freeing engineering time for innovation. Here we explore how this system works and its impact.

What is the Capacity Efficiency Program at Meta?

The Capacity Efficiency Program is Meta's strategic initiative to optimize performance across its vast infrastructure serving over 3 billion users. It addresses the reality that even a 0.1% performance regression can lead to massive power consumption. The program divides efforts into two complementary approaches: offense—proactively finding and deploying code optimizations to improve efficiency—and defense—monitoring production systems to detect regressions, identify root causes, and apply fixes. Together, these efforts ensure that Meta's growing user base doesn't proportionally increase energy usage. The ultimate goal is to create a self-sustaining efficiency engine where AI handles the majority of optimization tasks, allowing human engineers to focus on innovation.

Boosting Hyperscale Efficiency with AI Agents at Meta — Source: engineering.fb.com

How does the unified AI agent platform work?

Meta built a platform that encapsulates domain expertise from senior efficiency engineers into composable, reusable skills. These AI agents interact through a standardized tool interface, enabling them to automate both the discovery and resolution of performance issues. For example, an agent can autonomously investigate a regression, pinpoint the responsible code change, and even generate a ready-to-review pull request. This approach compresses tasks that once took engineers approximately 10 hours down to just 30 minutes. The platform is designed to scale across multiple product areas without requiring proportional headcount increases, making it a cornerstone of Meta's efficiency strategy.

What are the 'offense' and 'defense' sides of efficiency?

Meta views efficiency as a two-sided effort. Offense involves proactively searching for opportunities to make existing systems more efficient—for instance, by rewriting algorithms or reducing resource consumption. Engineers deploy these optimizations to improve baseline performance. Defense, on the other hand, handles regressions that slip into production. Using tools like FBDetect, the team monitors resource usage, detects anomalies, traces them back to specific pull requests, and rolls out mitigations. While both sides have been effective for years, the bottleneck was always human engineering time. The AI agent platform now accelerates resolution on both fronts, automating lengthy investigations and freeing engineers for higher-value work.

What role does FBDetect play in this system?

FBDetect is Meta's in-house regression detection tool. It scans production systems weekly and catches thousands of regressions that could waste power. Previously, each detection required manual investigation to find the root cause—a time-consuming process. Now, the AI agent platform integrates with FBDetect to automate this step. When FBDetect flags a regression, an agent immediately begins analysis, identifies the offending code change, and proposes a fix. Faster automated resolution means fewer wasted megawatts compound across the fleet. This tight integration between detection and automated response is key to scaling Meta's efficiency without linearly scaling the engineering team.

What energy savings has this program achieved?

The Capacity Efficiency Program has recovered hundreds of megawatts of power—enough to supply electricity to hundreds of thousands of American homes for a year. These savings come from both proactive optimizations (offense) and rapid regression fixes (defense). By automating diagnosis and resolution, the program compresses effort that used to take hours into minutes. This efficiency gain allows Meta to deliver more megawatt savings each half without proportionally increasing headcount. The long-term vision is a fully autonomous efficiency engine that handles routine tasks, enabling engineers to focus on breakthrough innovations while the AI manages the continuous optimization cycle.

What is the future direction of AI-assisted efficiency at Meta?

Meta aims to evolve the Capacity Efficiency Program into a self-sustaining engine. Currently, AI agents handle a growing number of opportunity resolutions that engineers would never reach manually. Over time, the platform will expand to more product areas, with agents taking on increasing responsibility for both detecting and fixing performance issues. The end state is an automated loop where AI discovers, analyzes, and resolves efficiency opportunities—closing the gap between regressions and proactive optimizations. This will minimize human intervention, allowing engineers to concentrate on innovative products while the infrastructure continuously fine-tunes itself for optimal power and performance.