Harnessing AI Agents for Hyperscale Efficiency: Inside Meta's Capacity Program

Meta's services reach over three billion users, so even minor performance issues can cause massive energy waste. To tackle this, the company developed a Capacity Efficiency Program that uses unified AI agents to automate detection and resolution of performance problems, saving significant power and engineering time. Below, we answer key questions about how this system works and its impact.

What Is Meta's Capacity Efficiency Program?

Meta's Capacity Efficiency Program is a strategic initiative focused on minimizing power consumption across its massive infrastructure. It operates through two complementary approaches: offense and defense. Offensive measures involve proactively identifying code optimizations that improve efficiency before deployment. Defensive measures monitor live production environments to instantly catch and mitigate performance regressions. The program leverages a unified AI agent platform that encodes the expertise of senior engineers into reusable skills, enabling automated investigation and resolution of efficiency issues. This system has already recovered hundreds of megawatts of power – enough to supply hundreds of thousands of homes for a year – and continues to scale without requiring proportional headcount growth.

How Do Unified AI Agents Boost Performance?

The unified AI agents act as a central platform that standardizes tool interfaces and embeds domain knowledge from top efficiency engineers. These agents are composed of reusable, composable skills that can be combined to automate both finding and fixing performance issues. On the defensive side, they accelerate the detection and root-cause analysis of regressions, reducing manual investigation from roughly ten hours to just thirty minutes. On the offensive side, they automatically discover optimization opportunities and even generate ready-to-review pull requests. By encoding expertise into software rather than requiring constant human intervention, these agents free engineers to focus on innovation while the program scales to handle growing infrastructure demands.

What Are the 'Offense' and 'Defense' Efficiency Strategies?

Meta's efficiency efforts split into two fronts: offense and defense. Offensive efficiency means proactively searching for opportunities to make existing systems more efficient – for example, rewriting code or adjusting configurations to use less power. These changes are deployed before they affect users. Defensive efficiency involves monitoring production in real time to detect regressions, or performance drops, that slip through. When a regression is found, the system quickly traces it back to a specific code change and deploys a fix. This two-sided approach ensures that improvements are continuously made while any accidental slowdowns are rapidly corrected. The AI agent platform automates both sides, enabling a self-sustaining cycle of efficiency gains.

How Does FBDetect Help Catch Regressions?

FBDetect is Meta's internal regression detection tool that acts as a security net for performance. Every week, it catches thousands of regressions – small performance dips that could compound into massive energy waste across the fleet. Without automation, each regression would require a human engineer to investigate, root-cause, and fix, consuming hours of time. The AI agent platform integrates with FBDetect to automate this process: it flags the regression, analyzes the code change responsible, and can even prepare a mitigation or fix. Faster, automated resolution prevents accumulated power waste and keeps the overall system efficient. By moving from manual triage to AI-driven response, Meta saves megawatts that would otherwise be lost over time.

What Power Savings Has the AI Platform Achieved?

The AI agent platform has successfully recovered hundreds of megawatts of power – an amount sufficient to power hundreds of thousands of American homes for an entire year. This achievement stems from automating both offensive efficiency improvements and defensive regression fixes. Previously, many potential optimizations were missed because engineers lacked time to investigate every opportunity. Now, AI agents continuously scan for gains and automatically deploy solutions. On the defense side, the rapid resolution of regressions prevents gradual power waste. These savings allow Meta to scale its services sustainably without escalating energy costs, demonstrating that AI-driven efficiency can make a real-world environmental and economic impact.

How Does the Platform Compress Investigation Time?

Manual investigation of a performance issue typically took engineers about ten hours to diagnose and resolve. The AI agent platform compresses this into roughly thirty minutes – a twentyfold improvement. It does this by combining encoded domain knowledge with automated data analysis. When a regression is detected, the agent instantly compares the problematic code against historical patterns, identifies common root causes, and proposes fixes. For opportunity discovery, the agent scans codebases and metrics to pinpoint efficiency gains, then writes and submits a pull request for review. This compression of time means that more issues can be addressed in less time, scaling the program's throughput without needing a larger team.

What Is the Future Goal of This Program?

The ultimate goal of Meta's Capacity Efficiency Program is to create a self-sustaining efficiency engine. In this vision, AI handles the long tail of efficiency opportunities and regressions, requiring minimal human oversight. Engineers shift from firefighting performance issues to innovating on new products and features. As the infrastructure grows, the program aims to keep increasing megawatt delivery without proportionally increasing headcount. This means that the AI agent platform will continue to expand to more product areas every half-year, handling a growing volume of wins that would otherwise be impossible to address manually. By fully automating the efficiency lifecycle, Meta hopes to maintain hyperscale growth while minimizing its energy footprint and freeing up its engineering talent for creative work.