Internal Agents Map

Ranking Engineer Agent (REA)

Company
Meta
Approach type
Agent
Work
Data, Research
Human involvement
Drafts reviewed
Invocation
Interactive, Background, Event-driven
Deployment stage
Deployed
Evidence strength
Detailed primary
Entry reviewed

Purpose

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.

Check the reviewed sources and claim details.

How it works

Representative workflow: Hypothesis generation through an engineer-confirmed plan, multiround training experiments, and a model-improvement proposal returned for approval. Research details

Generate hypotheses from two sources

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

Agree the exploration strategy and compute budget with an engineer

Before executing any plan REA proposes a detailed exploration strategy, estimates total GPU compute cost, and confirms the approach with an engineer

Run the three-phase plan

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

Hibernate during a training run and wake when the job completes

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

Log each experiment back into the insight database

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

Where people stay involved

Each scope pairs its normal attention boundary with supporting evidence. See the supervision definitions for the level mapping and limits.

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.

  • engineer-approved experiment plan → multiround training experiments → model-improvement proposal reviewed for launch

    Outcome review · Level 4

Catalog interpretation: Level 4 for engineer-approved experiment plan → multiround training experiments → model-improvement proposal reviewed for launch; human attention boundary: outcome-review.

Observed in March 2026

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
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; the legacy unknown claim stays in research details.
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
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
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
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

REA Planner and REA Executor over a shared skill, knowledge and tool system

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

Validation and failure handling

Screen failed runs and adapt within predefined guardrails

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

Reported observations

The catalog records what the sources report, with the scope and the denominator of every figure. A qualification below limits the figure it sits under.

Effectiveness · Reported measurement · Average model accuracy of REA-driven iterations against baseline approaches across six ads ranking models in the first production validation

Headline claim

“REA-driven iterations doubled average model accuracy over baseline approaches across six ads ranking models in the first production validation (self-reported, March 2026).”

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

Adoption output · Estimate · Engineer staffing for model-improvement proposals: three engineers across eight models against a stated historical norm of two engineers per model

Key observation

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.

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

Adoption output · Reported measurement · Model-improvement proposals per early adopter, reported as a rise from one to five in the same time frame

Key observation

Meta reports that early adopters increased their model-improvement proposals from one to five in the same time frame.

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

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.

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.

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.

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.

Sources and research details

Citations link to the original publisher. Each source also keeps a preserved copy in the repository, so a changed or removed page stays checkable.

  1. Ranking Engineer Agent (REA): The Autonomous AI Agent Accelerating Meta's Ads Ranking Innovationhttps://engineering.fb.com/2026/03/17/developer-tools/ranking-engineer-agent-rea-autonomous-ai-system-accelerating-meta-ads-ranking-innovation/Engineering blog · First party · Last source verification: 2026-09-21
    Preserved copy
Question coverage and scope
purpose
Reported
workflow
Reported
human involvement
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.
implementation
Reported
validation
Reported
observations
Reported
lessons
Reported
Research details for every claim on this page
  1. Summary
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    High
    Confidence reason
    A linked first-party source states the claim.
  2. Headline claim
    Statement type
    Metric
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    Meta reports the result in its own engineering blog with a stated denominator of six models and a stated comparison against baseline approaches, but does not define the accuracy measure, give absolute values, identify the baseline approaches, or cite an independent review.
    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
    Observation date
    2026-03
  3. Sandbox

    unknown

    Statement type
    Inference
    Provenance
    Catalog judgment
    Confidence
    Medium
    Confidence reason
    The post names Meta internal job schedulers, access controls granted by preflight checklist review, and a codebase restriction, but no execution isolation boundary for the agent; unknown does not mean absent.
  4. Harness
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    High
    Confidence reason
    A linked first-party source states the claim.
  5. Tool access
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    High
    Confidence reason
    A linked first-party source states the claim.
  6. Knowledge
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    High
    Confidence reason
    A linked first-party source states the claim.
  7. Context management
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    High
    Confidence reason
    A linked first-party source states the claim.
  8. REA Planner and REA Executor over a shared skill, knowledge and tool system
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    High
    Confidence reason
    A linked first-party source states the claim.
  9. Generate hypotheses from two sources
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    High
    Confidence reason
    A linked first-party source states the claim.
  10. Agree the exploration strategy and compute budget with an engineer
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    High
    Confidence reason
    A linked first-party source states the claim.
  11. Run the three-phase plan
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    High
    Confidence reason
    A linked first-party source states the claim.
  12. Hibernate during a training run and wake when the job completes
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    High
    Confidence reason
    A linked first-party source states the claim.
  13. Log each experiment back into the insight database
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    High
    Confidence reason
    A linked first-party source states the claim.
  14. Screen failed runs and adapt within predefined guardrails
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    High
    Confidence reason
    A linked first-party source states the claim.
  15. Key observation
    Statement type
    Metric
    Provenance
    Reported
    Confidence
    Low
    Confidence reason
    The 5x label is a ratio derived from a staffing comparison rather than a measured throughput: Meta sets three engineers across eight models against a stated historical norm of two engineers per model. No time period, matched baseline measurement, or control for model difficulty is reported, and the output counted is proposals to launch improvements rather than launched improvements.
    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
    Observation date
    2026-03
  16. Key observation
    Statement type
    Metric
    Provenance
    Reported
    Confidence
    Low
    Confidence reason
    Meta reports a rise from one to five model-improvement proposals per early adopter in its own engineering blog. The interval is given only as the same time frame, the number of early adopters is not stated, and proposal quality or acceptance is not assessed.
    Reported by
    Meta
    Scope
    Model-improvement proposals per early adopter using REA
    Denominator
    Not reported
    Method
    Self-reported before-and-after proposal count over an interval described only as the same time frame
    Observation date
    2026-03
  17. Lesson
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    The resilience section states that REA consults a runbook of common failure patterns, makes prioritization decisions, and debugs preliminary infrastructure failures instead of waiting for human intervention, and that engineers provide periodic oversight. The post reports the design; it gives no escalation rate or measurement of how often the guardrails held.
  18. Lesson
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    The safeguards paragraph names the codebase restriction, the preflight checklist access grant, and the up-front compute budget confirmation with halting or pausing at the threshold. These are stated controls; the post does not describe how they are enforced or audited.
  19. Lesson
    Statement type
    Fact
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    The hypothesis-generation section states that synthesizing the historical insights database with the ML research agent surfaces configurations unlikely to emerge from any single approach, and that the most impactful improvements combined architectural optimizations with training-efficiency techniques. The attribution is Meta's own and rests on no reported comparison against a single-source baseline, so the lesson stays bounded to the reported ads ranking case.
  20. Lesson
    Statement type
    Opinion
    Provenance
    Reported
    Confidence
    Medium
    Confidence reason
    The closing section states the team's own position on how ML engineering should be structured, with agents handling iterative mechanics and humans making strategic decisions and final approvals. It is an attributed preference about the direction Meta is taking, not a measured outcome or a general recommendation.
  21. Operating model assessment
    Statement type
    Inference
    Provenance
    Catalog judgment
    Confidence
    Medium
    Confidence reason
    The post places human attention at two points of a normal run: an engineer confirms the exploration strategy and GPU compute cost before any plan executes, and humans make strategic decisions and final approvals on the resulting proposals. Engineers are described as giving periodic oversight rather than continuous monitoring, and routine failures are handled inside the guardrails, so attention does not return at each experiment round or at the level of individual training configurations. The post never says humans are involved only when the agent raises an exception, so the boundary does not reach exception-only. The documented unattended runtime of days to weeks is a separate property and does not move the boundary. The post also states it covers only the ML experimentation capability, so this assessment does not extend to other REA capabilities.
    Observation date
    2026-03
Back to agents
Know an internal agent?

Share a resource or public mention, suggest an addition, or correct an existing entry. Pull requests are also welcome.

Suggest a resource or update