Meta's AI Agent Platform for Hyperscale Capacity Efficiency

The Challenge of Hyperscale Efficiency

Operating at Meta's scale, where code serves over 3 billion people, means even a tiny performance regression (e.g., 0.1%) translates into substantial power consumption and increased infrastructure costs. Traditional manual methods for identifying, root-causing, and resolving these issues become a significant bottleneck, consuming valuable engineering time that could otherwise be spent on innovation. This problem manifests in two primary areas: proactively finding optimization opportunities (offense) and reactively mitigating performance regressions (defense).

Unified AI Agent Architecture for Efficiency

Meta's solution is a unified AI agent platform designed to automate both offensive and defensive efficiency tasks. The key architectural insight was recognizing that both problems share a similar structure: gather context, apply domain expertise, and create a resolution. This allowed for a single platform built on two distinct layers:

• MCP Tools: Standardized interfaces that allow Large Language Models (LLMs) to interact with Meta's infrastructure. These tools perform specific actions like querying profiling data, fetching experiment results, retrieving configuration history, searching code, or extracting documentation.
• Skills: These encode the domain expertise of senior efficiency engineers. A skill guides the LLM on which tools to use and how to interpret their results, capturing complex reasoning patterns (e.g., "consult top GraphQL endpoints for latency regressions").

Offensive and Defensive Applications

The same underlying platform powers both proactive optimization and reactive regression handling:

• Defense: AI Regression Solver (FBDetect): When FBDetect, Meta's regression detection tool, identifies a performance drop, the AI Regression Solver automatically gathers context (symptoms, root cause PR), applies mitigation skills (e.g., increasing sampling for logging regressions), and generates a new pull request to fix the issue. This automates what was traditionally a manual root-cause analysis and fix deployment.
• Offense: Opportunity Resolution: For proactively identified efficiency opportunities, engineers can request an AI-generated pull request. The agent gathers opportunity metadata, documentation, examples, and relevant code files using tools, then applies specific optimization skills (e.g., memoizing functions) to produce a candidate fix. This drastically reduces the time from opportunity identification to code deployment.

Impact and Future Expansion

The Capacity Efficiency Program has recovered hundreds of megawatts of power. Beyond direct savings, it fundamentally shifts how engineers approach efficiency. Instead of time-consuming manual investigations, engineers review AI-generated analyses and code, enabling faster deployment of high-impact fixes. The unified architecture promotes compounding returns; new capabilities like conversational assistants, capacity planning agents, and personalized recommendations can be built by composing existing tools with new skills, minimizing data integration overheads and accelerating innovation.