Ranking Engineer Agent (REA)
Meta built the Ranking Engineer Agent (REA) to run machine learning experimentation for its ads ranking models. REA generates hypotheses from a historical experiment database and a deep ML research agent, proposes an exploration strategy and GPU compute budget that an engineer confirms, then launches training jobs, debugs failures, and iterates across workflows that span days to weeks, hibernating while a job runs and waking when it completes. Engineers confirm the plan and approve the resulting improvement proposals.
- Company
- Meta
- Approach type
- Agent
- Human involvement
- Drafts reviewed
- Invocation
- Interactive, Background, Event-driven
- Deployment stage
- Deployed
- Evidence strength
- Detailed primary
- Entry reviewed
How it works
Hypothesis generation through an engineer-confirmed plan, multiround training experiments, and a model-improvement proposal returned for approval.
REA consults a curated historical insights database of past experiments and a deep ML research agent that investigates baseline model configurations, then synthesizes both to surface configurations unlikely to emerge from either source alone[Source 1]
Before executing any plan REA proposes a detailed exploration strategy, estimates total GPU compute cost, and confirms the approach with an engineer[Source 1]
Validation tests individual hypotheses from different sources in parallel to establish quality baselines, combination merges promising hypotheses to search for synergistic improvements, and exploitation explores the most promising candidates aggressively within the approved compute budget[Source 1]
When REA launches a training job it delegates the wait to a background system, shuts down to conserve resources, and automatically resumes where it left off when the job completes, so workflows continue across extended time frames without constant human monitoring[Source 1]
A dedicated experiment logger records outcomes, key metrics, and configurations into the centralized hypothesis experiment insight database, and the hypothesis generator draws on those insights to propose hypotheses for each subsequent round[Source 1]
Where people stay involved
People review, approve, or guide the agent’s work here.
Reported: An engineer confirms the exploration strategy and GPU compute budget before any plan runs, and humans make the final approvals on the resulting proposals; routine failures are handled inside the guardrails instead of being escalated.
Supervision level
Outcome review
A person evaluates tests, behavior, or outcomes.
Level 4 of 5
- engineer-approved experiment plan → multiround training experiments → model-improvement proposal reviewed for launch[Source 1]
Implementation details
- Model
- Not reportedNo model, model family, or provider is named for REA; the post says only that Meta is fine-tuning specialized models for hypothesis generation as continuing work.
- Harness
- Built on Confucius, an internal Meta AI agent framework for complex, multistep reasoning that supplies code generation and an SDK for integrating with Meta internal tooling; REA itself splits into a REA Planner and a REA Executor over a shared skill, knowledge and tool system, with the executor running an agent loop and a wait state[Source 1]
- Sandbox
- Not reportedThe post names Meta internal job schedulers, a codebase restriction, and access controls granted by preflight checklist review, but no execution isolation boundary.
- Tool access
- Integrates through the Confucius SDK with Meta internal job schedulers, experiment tracking infrastructure, and codebase navigation tools; launches and manages training jobs and works exclusively on the Meta ads ranking model codebase under engineer-granted access controls and a confirmed compute budget[Source 1]
- Knowledge
- A curated historical insights database of past experiments supports in-context learning across prior successes and failures; a dedicated experiment logger writes outcomes, key metrics, and configurations into a centralized hypothesis experiment insight database that both the deep ML research agent and the hypothesis generator read[Source 1]
- Context management
- A hibernate-and-wake mechanism delegates the wait for a training job to a background system, shuts the agent down to conserve resources, and resumes it where it left off when the job completes; REA maintains persistent state and memory across multiround workflows spanning days or weeks[Source 1]
- Credentials
- Not reportedEngineers grant explicit access controls through preflight checklist reviews, but how REA authenticates to job schedulers, experiment tracking, or the codebase is not described.
- Interfaces
- Not reportedThe post describes an engineer collaborating with the hypothesis generator through the REA Planner and a Confucius SDK for internal tooling, but names no user-facing interface surface.
Mechanisms
The planner builds a detailed experiment plan with an engineer, the plan is exported to the executor, which manages asynchronous job execution through an agent loop and wait state, and a shared skill, knowledge and tool system supplies ML capabilities, historical experiment data, and integrations with Meta internal infrastructure[Source 1]
Validation and failure handling
REA consults a runbook of common failure patterns, applies prioritization logic such as excluding jobs with clear out-of-memory errors or training instability signals such as loss explosions, debugs preliminary infrastructure failures from first principles, and halts or pauses runs when a compute threshold is reached, rather than surfacing routine interruptions to engineers[Source 1]
Reported observations
These results come from the sources. Notes below each result explain what it covers and any reported limits.
REA-driven iterations doubled average model accuracy over baseline approaches across six ads ranking models in the first production validation (self-reported, March 2026).[Source 1]
- Reported by
- Meta
- Scope
- Average model accuracy of REA-driven iterations in the first production validation
- Denominator
- A set of six ads ranking models
- Method
- Comparison of REA-driven iterations against baseline approaches; the accuracy measure is not defined
Observed in March 2026
About this result
- Category
- Effectiveness
- Evidence type
- Reported measurement
- What it covers
- Average model accuracy of REA-driven iterations against baseline approaches across six ads ranking models in the first production validation
Three engineers delivered proposals to launch improvements for eight models, work Meta says historically required two engineers per model; Meta labels this a 5x engineering output gain.[Source 1]
- Reported by
- Meta
- Scope
- Engineer staffing for model-improvement proposals produced with REA
- Denominator
- Eight ads ranking models staffed by three engineers
- Method
- Ratio derived from a stated historical staffing norm of two engineers per model; no measured baseline throughput is reported
Observed in March 2026
About this result
- Category
- Adoption output
- Evidence type
- Estimate
- What it covers
- Engineer staffing for model-improvement proposals: three engineers across eight models against a stated historical norm of two engineers per model
Meta reports that early adopters increased their model-improvement proposals from one to five in the same time frame.[Source 1]
- Reported by
- Meta
- Scope
- Model-improvement proposals per early adopter using REA
- Method
- Self-reported before-and-after proposal count over an interval described only as the same time frame
The source does not report the denominator of this figure.
Observed in March 2026
About this result
- Category
- Adoption output
- Evidence type
- Reported measurement
- What it covers
- Model-improvement proposals per early adopter, reported as a rise from one to five in the same time frame
Lessons
Meta keeps engineers out of routine failure handling on REA: the agent consults a runbook of common failure patterns and adapts within predefined guardrails, and engineers provide periodic oversight rather than continuous monitoring.[Source 1]
Meta bounds the agent by scope and budget: REA works exclusively on the ads ranking model codebase, engineers grant explicit access controls through preflight checklist reviews, and REA confirms compute budgets up front and halts or pauses runs when a threshold is reached.[Source 1]
For Meta ads ranking, pairing a historical experiment database with a deep research agent surfaced configurations that Meta says neither source produced alone, and the most impactful improvements combined architectural optimizations with training-efficiency techniques.[Source 1]
Reported opinion: The Meta team holds that agents should handle the iterative mechanics of ML experimentation while engineers move toward hypothesis direction, architectural decision-making, and final approvals.[Source 1]