Meta's Capacity Efficiency: How Unified AI Agents Optimize Hyperscale Performance

Meta's Capacity Efficiency Program built an AI agent platform that automates the work of finding and fixing performance issues across its infrastructure. The core idea is combining a standardized tool interface with skills that embed domain knowledge, turning the judgment of experienced engineers into a reusable form.

The platform has already recovered hundreds of MW of power, and cut analysis that used to take about 10 hours manually down to about 30 minutes. In an environment serving more than 3 billion people, even a 0.1% performance degradation leads to massive power waste at scale, so speed and automation translate directly into efficiency.

Efficiency work runs on two tracks.

• Defense: FBDetect catches thousands of regressions per week, and AI traces the root cause down to the code/config change before generating a fix-forward correction as a pull request.
• Offense: When an engineer selects an efficiency improvement opportunity, AI gathers relevant documentation, similar cases, target files and functions, and validation criteria to produce an immediately applicable fix.

Both flows share the same structure, with the difference lying in skills. Common tools handle profiling data, experiment results, configuration history, code search, and documentation extraction, while skills tailored to each problem type determine which tools to use, in what order, and how to interpret the results.

For example, in regression response, the platform locates the function where the symptom appeared and the file/line of the causal PR, and for log-related regressions it applies mitigation strategies such as increased sampling. In efficiency opportunity response, it turns specific optimization patterns—like memoization—into executable code, checks syntax, style, and fit for the problem, and produces a reviewable PR.

This unified structure has since been extended further. Conversational efficiency Q&A, a capacity planning agent, personalized opportunity recommendations, guided investigation workflow, and AI-assisted validation could all be built on the same foundation, without needing to significantly scale up separate data integration for each new feature.

Ultimately, what Meta is aiming for is a self-sustaining efficiency engine where AI handles long-tail investigation and fixing work, while engineers focus on high-value product development.