--- schema: "swft.publication/v1" id: "company-meta-capacity-agents" title: "Meta's capacity agents: production signals become review-ready fixes" description: "How Meta connects production regression detection and optimization opportunities to shared tools, expert skills, agent-prepared fixes, and human review." summary: "Meta starts agent work from measured production regressions or reviewed optimization opportunities. Shared tools gather evidence, reusable skills apply expert methods, and the agent prepares a fix for the responsible engineer. The company reports much faster investigations, while the public evidence does not isolate quality or total operating impact caused by agents." canonical: "https://swft.io/companies/meta-capacity-agents" author: "SWFT Editorial" author_type: "Organization" published: "2026-09-01" modified: "2026-09-02" kind: "case-study" section: "Companies" tags: ["Meta capacity agents", "AI regression solver", "performance engineering agents", "MCP tools", "agent skills", "production-to-PR"] evidence_labels: ["INFERENCE", "SELF-REPORT"] source_ids: ["meta-capacity-efficiency"] authorship_disclosure: "AI-drafted from the cited public sources and independently checked by a second AI editorial-review agent (Codex) for source fit, claim boundaries, overlap, and reader utility. SWFT Editorial is responsible for corrections." --- # Meta's capacity agents: production signals become review-ready fixes Meta combines regression detection, profiling data, code history, reusable expert skills, agent implementation, and human review in one performance-efficiency line. > **Authorship:** AI-drafted from the cited public sources and independently checked by a second AI editorial-review agent (Codex) for source fit, claim boundaries, overlap, and reader utility. SWFT Editorial is responsible for corrections. ## Quick answer Meta starts agent work from measured production regressions or reviewed optimization opportunities. Shared tools gather evidence, reusable skills apply expert methods, and the agent prepares a fix for the responsible engineer. The company reports much faster investigations, while the public evidence does not isolate quality or total operating impact caused by agents. Meta connected production performance signals to agent-prepared code changes. Its capacity-efficiency platform detects regressions or queues optimization opportunities, gathers code and operating context, applies reusable expert playbooks, and returns a pull request or editor-ready fix for a person to review. The factory begins with measured production work, not a blank prompt. ## What are Meta's capacity-efficiency agents? Meta's [April 2026 operating account](https://engineering.fb.com/2026/04/16/developer-tools/capacity-efficiency-at-meta-how-unified-ai-agents-optimize-performance-at-hyperscale/) describes one internal agent platform serving two related jobs. The defensive job responds to a performance regression that has already reached production. The offensive job starts from a proposed optimization, such as reducing CPU use in a known hot function. Both jobs need profiling data, code, documentation, configuration history, examples, and a way to prepare a safe change. Meta built shared Model Context Protocol tools for those information sources. A tool performs a bounded operation, such as searching code or retrieving an experiment result. **Skills** encode the reasoning patterns of experienced efficiency engineers, including which tools to call and how to interpret the result. This combination turns expert knowledge into repeatable task specifications and supplied context. New use cases can reuse the same tools while changing the skill that guides the work. ## How does Meta turn a production regression into a pull request? FBDetect is Meta's internal regression detector. The company says it finds thousands of performance regressions each week and can detect changes as small as 0.005 percent in noisy production data. Once ordinary detection techniques connect a regression to a code or configuration change, the AI Regression Solver performs three stages: 1. Gather the symptoms, affected functions, causative pull request, and exact lines that changed. 2. Load mitigation guidance for the codebase, language, and type of regression. 3. Prepare a corrective pull request and send it to the author of the original change for review. The agent does not decide that a production metric moved. A separate measurement system supplies the signal and likely cause. The agent handles investigation and implementation after that evidence exists. ## How does Meta turn an optimization idea into code? On the offensive side, an engineer can choose an efficiency opportunity and request a proposed implementation. The agent gathers the opportunity description, optimization documentation, examples of similar work, relevant functions, and validation criteria. A skill supplies the domain method, such as when memoization is appropriate. The system then proposes a guarded change, verifies syntax and style, and checks that it addresses the intended opportunity. Meta says the result appears in the engineer's editor and can be applied with one click. That is a different handoff from the defensive path's ready-to-review pull request, so the two should not be described as equally autonomous. Meta reports that automated diagnosis can compress roughly ten hours of manual investigation into about 30 minutes. This is a company estimate for the described workflow, not an independent time study. ## What impact has Meta reported? Meta says its broader Capacity Efficiency Program has recovered hundreds of megawatts of power. It also says the agent systems for offense and defense contribute to that program and help the team expand into more product areas without matching headcount growth. The public article does not isolate how many megawatts came from agent-generated changes. It does not publish proposal counts, acceptance or merge rates, false-positive rates, reverted fixes, defect rates, review time, or the distribution behind the ten-hours-to-30-minutes example. The platform is internal. The article describes its architecture and operating use but does not link a public implementation of the tools, skills, or solver. ## Why is this a software factory rather than an AI coding assistant? The work enters through a measured queue. Production detection or a reviewed opportunity defines the target. Shared tools assemble evidence. A reusable skill applies specialist knowledge. The agent prepares a bounded result. Existing review and deployment controls decide what ships. That repeated line is more important than the chat interface. It makes task selection, context, proof, and ownership visible to the team. ## What can a product team copy? Start with an operating signal that already has an owner and a trusted detector. Keep detection deterministic. Create narrow tools for logs, traces, code history, documentation, and experiments. Encode one expert resolution pattern as a skill. Require the agent to name its evidence and validation criteria. Return the result to the person who owns the affected system. Track proposals, accepted fixes, investigation time, review time, false alarms, regressions after release, and the operating value actually recovered. A large program-level result should not be attributed to the agent without that causal bridge. ## How we know - **First-party report (SELF-REPORT)** Meta describes a shared platform of MCP tools and expert skills that supports both regression response and proactive performance optimization. Sources: [Capacity Efficiency at Meta: How Unified AI Agents Optimize Performance at Hyperscale](https://engineering.fb.com/2026/04/16/developer-tools/capacity-efficiency-at-meta-how-unified-ai-agents-optimize-performance-at-hyperscale/). - **First-party report (SELF-REPORT)** Meta reports that FBDetect finds thousands of regressions weekly and that the AI Regression Solver can gather context, apply mitigation knowledge, and prepare a corrective pull request. Sources: [Capacity Efficiency at Meta: How Unified AI Agents Optimize Performance at Hyperscale](https://engineering.fb.com/2026/04/16/developer-tools/capacity-efficiency-at-meta-how-unified-ai-agents-optimize-performance-at-hyperscale/). - **First-party report (SELF-REPORT)** Meta reports compressing about ten hours of diagnosis to about 30 minutes and says AI systems contribute to a broader capacity program that recovered hundreds of megawatts. Sources: [Capacity Efficiency at Meta: How Unified AI Agents Optimize Performance at Hyperscale](https://engineering.fb.com/2026/04/16/developer-tools/capacity-efficiency-at-meta-how-unified-ai-agents-optimize-performance-at-hyperscale/). - **Analysis (INFERENCE)** SWFT treats trusted production signals, reusable expert skills, and an existing owner review path as the most transferable parts of Meta's factory. Sources: [Capacity Efficiency at Meta: How Unified AI Agents Optimize Performance at Hyperscale](https://engineering.fb.com/2026/04/16/developer-tools/capacity-efficiency-at-meta-how-unified-ai-agents-optimize-performance-at-hyperscale/). ## Sources - **First-party report (SELF-REPORT)** [Capacity Efficiency at Meta: How Unified AI Agents Optimize Performance at Hyperscale](https://engineering.fb.com/2026/04/16/developer-tools/capacity-efficiency-at-meta-how-unified-ai-agents-optimize-performance-at-hyperscale/) — Engineering at Meta; published 2026-04-16; accessed 2026-09-01. Meta's account of production-regression detection, standardized tools, reusable efficiency skills, context gathering, candidate fixes, validation criteria, and ready-to-review pull requests. ## Read next - [Cloudflare's AI engineering stack: a control plane for company-wide agents](/companies/cloudflare-ai-engineering-stack) - [Spotify Honk: adding an agent to a software factory that already worked](/companies/spotify-honk) - [Software factory metrics that matter](/software-factory-metrics)