REA, the Autonomous AI Agent Accelerating Meta's Ad Ranking Innovation

Meta's Ranking Engineer Agent (REA) autonomously carries out the entire ML experimentation process for ad ranking models. It goes from hypothesis generation to running training jobs, debugging failures, and iterating based on results, with humans handling only key strategic decisions.

Traditional ML experimentation was a bottleneck where each step—writing hypotheses, designing experiments, running training, analyzing failures, and retrying—was manually chained together, taking days to weeks. To solve this, Meta built an agent that can carry out long-horizon workflows on its own, and REA is specifically focused on running multi-day experiments without interruption.

REA has three core elements.

• Long-Horizon Autonomy: A hibernate-and-wake mechanism that waits for training jobs to finish and then wakes back up, continuously managing experiments that span days or weeks.
• High-Quality, Diverse Hypothesis Generation: Combines a Historical Insights Database that aggregates past experiments with an ML Research Agent that conducts in-depth ML research, producing hypotheses that would be hard to generate through a single approach alone.
• Resilient Execution: Handles failures such as OOM, loss explosion, and infrastructure errors using runbooks and priority logic, adjusting plans within an allowed range without engineer intervention.

The execution structure is split into REA Planner and REA Executor, supported by a shared Skill, Knowledge and Tool System that provides experiment data, internal infrastructure integration, and code exploration capabilities. The Planner proposes experiment plans and GPU costs, while the Executor runs asynchronous jobs and accumulates results in logs, which are then fed back into generating the next hypotheses.

The operating approach is organized into a three-stage plan: Validation → Combination → Exploitation. First, individual hypotheses are validated in parallel; next, promising ideas are combined; and finally, the most promising candidates are intensively optimized within the approved compute budget.

In its first production rollout, REA doubled average accuracy across 6 models and increased engineering output by 5x. Improvement work that used to require 2 engineers per model can now be handled by 3 engineers who can even produce proposals for 8 models.

Going forward, Meta plans to further advance REA by refining specialized models for hypothesis generation, expanding analysis tools, and extending the approach to other domains. The role of ML engineering is being reshaped so that humans move away from repetitive tasks and focus more on strategic judgment and final approval.