If you run dozens of production models and are being pitched agentic ML tooling, Meta's latest numbers say to do the boring part first. Meta estimates its AI agents now handle over 60% of all technique propagation across its ads ranking models, the work of carrying a proven ML improvement from one model to all the others. That result sits on two years of standardization that cut per-model engineering time by 92% before agents took over execution. Copy the order of operations: templates and evals first, agents second.
Meta Engineering published both posts on November 18, 2026. The headline claims are Meta's own and unaudited: agents have driven execution since March 2026, each engineer now handles "3x more technique-model adaptations," training compute fell a median 4x "across select combinations," and Meta projects "an estimated $45M+ in compute savings by the end of the year." The model behind the agents is not named. Below is what the system does, what had to exist before it, and which parts of the design an ML platform team without Meta's budget can borrow.
What Is Technique Propagation, and Why Did Meta Hand It to Agents?
Technique propagation is the work of taking a modeling improvement that has already won on one model (a new architecture layer, a feature-interaction module, a training setting) and adapting it to every other model where it might also win. At Meta that backlog is large. The November post says propagation "can take over a year" today and sets a goal of three months, which needs about 5x higher end-to-end throughput. The techniques it names as examples include Wukong, a stacked factorization-machine architecture built to give recommendation models a scaling law, and InterFormer, a heterogeneous interaction-learning module for click-through prediction.
The agents split the job into three stages:
- Discovery keeps a live view of proven techniques, estimates how well each would transfer, and produces a ranked list of technique-model candidates. Engineers review the agents' assessment and decide.
- Generalization proposes hypotheses for adapting a technique, launches experiments on a representative set of models, diagnoses failures and iterates. Search starts from configurations likely to work, drawing on what the post calls "millions of prior training runs," and kills unpromising experiments early.
- Adoption launches the experiments for each target model, manages multi-day asynchronous training runs, tunes each model, recovers from failures and assembles a validated proposal for a human to approve.
Agents also schedule their own capacity. They rank concurrent jobs by revenue importance, technique potential and deadline, and reclaim GPUs from stalled or low-value runs, across what the post calls "hundreds of concurrent jobs."
This is the latest published piece on Meta's Ranking Engineer Agent (REA). The March 2026 REA post covered experimentation: three engineers delivered launch proposals for eight models, work that "historically required two engineers per model." The April KernelEvolve post covered kernel tuning, where Meta reports more than 60% higher inference throughput on its Andromeda ads model on Nvidia GPUs and over 25% higher training throughput on an ads model running on its own MTIA chips.
Standard Model Templates Did the Heavy Lifting First
Standard Model Templates (SMT) are Meta's set of reusable, opinionated model setups (shared architecture choices, feature logic, training settings and serving configs) from which production ranking models are built. Once a technique is validated at the template level, it reaches every model built on that template. The companion SMT post credits two years of standardization and automation with 5,000+ technique-model combinations shipped and the 92% cut in per-model engineering time. The agent post counts that result separately from what the agents added.
The SMT paper on arXiv states the math plainly. With n models and k techniques, bespoke models make propagation an O(n·k) problem: every technique adapted to every model by hand. Templates turn it into O(n+k). The paper reports a 6.3x increase in technique-model adoption throughput over five global development cycles in Meta's ads ranking system.
The templates decide how much work is left for the agents. The November post describes two paths. For template models, agents generalize a technique once per template and then run an "auto-research loop" to tune each model on top of it. For the bespoke models outside the templates, agents first learn alongside ML engineers on the first few adaptations and only then propose changes for the rest, with the engineer "reviewing and steering." The SMT post says the same about automation in general: AutoML and optimization work best in a structured search space, and fragmentation limits which models they can reach.
If your fleet looks like most enterprise fleets (each model built by a different team, on a different feature pipeline, with its own training script), an agent has to work the expensive, engineer-supervised path on every model.
Where Meta Keeps Humans in the Loop
Meta's design puts humans at three points: plan approval, judgment calls during the run, and the decision to ship. In REA's original form, the agent "proposes a detailed exploration strategy, estimates total GPU compute cost, and confirms the approach with an engineer" before running anything, according to the March post. It confirms compute budgets up front and halts or pauses runs when thresholds are reached. It works only inside Meta's ads ranking codebase, consults a runbook of known failure patterns, and hibernates during long training jobs, waking when they finish.
The November post keeps that structure. Engineers approve plans and set priorities, review key decisions, and approve what ships. The ship decision rests on gates that exist independently of the agent:
- Offline, a change must improve normalized entropy without regressing training or inference cost.
- Online, it needs a statistically significant A/B win.
- A generalized technique is screened on scaled-down models first, then has to improve representative full-size models and meet every guardrail.
Meta also names the metrics it uses to judge the agents: precision and recall for discovery (recall weighted higher), and for adoption, "the number of human interventions needed to complete an experiment" and "the percentage of issues agents diagnose and mitigate on their own." Those two adoption metrics are the most portable thing in the post. They measure whether the agent is reducing supervision, which is the cost you are paying to remove.
There is one carve-out. For the highest-revenue models, the post says expert engineers adapting each model by hand "is still the process." Agents assist there but do not drive. Meta trusts the agents with the long tail and keeps people on the models where a regression costs the most. That is a sensible default for anyone else. It also suggests the 60% figure leans toward the models where agent mistakes are cheapest.
What Meta's Numbers Leave Out
Every figure here is self-reported, and several come without the baseline you would need to reuse them. Before quoting them in a business case, read the caveats Meta prints alongside them:
- The $45M is a projection. The post calls it "an estimated $45M+ in compute savings by the end of the year" and gives no baseline spend, so you cannot turn it into a percentage.
- The 4x compute cut is a median over "select" combinations. The post does not say which ones or how many.
- The 60% is an estimate of propagation work. The post gives no model count.
- The adoption metrics have no values. Meta names interventions per experiment and self-diagnosed issues as its yardsticks and publishes no numbers for either.
- The model is unnamed. The agents "combine LLM-based reasoning with specialized ML tooling." The March post says REA runs on Meta's internal Confucius agent framework, a scaffold whose own paper reports a 59% Resolve@1 on SWE-Bench-Pro. It does not say which base model a buyer would need.
- The SMT figures differ between the blog and the paper. The blog reports a 0.63% average offline quality gain and a 0.86% cumulative online lift; the arXiv abstract reports 0.77% and 1.52%. The paper is on its fourth version, revised most recently on November 15, 2026. Both are positive, and both are small per model, which is the point at Meta's scale and may not be at yours.
- REA's "2x accuracy" has no defined metric. The March post says REA-driven iterations "doubled average model accuracy over baseline across six models" without saying which accuracy measure.
The system is running in production. Its numbers are sized to an ads business that the March post says reaches billions of people, so treat them as Meta's results and build your own baseline before you forecast anything.
The Last Time ML Teams Learned This
ML teams have paid for skipping the platform work before. The 2015 NIPS paper "Hidden Technical Debt in Machine Learning Systems" by D. Sculley and colleagues argued that real-world ML systems commonly incur massive ongoing maintenance costs, and listed the causes: entanglement, hidden feedback loops, undeclared consumers, data dependencies and configuration problems. Feature stores, model registries and pipeline templates have spent the decade since working on that debt.
Agentic ML tooling is now arriving on the same fleets. An agent that can launch experiments on a fragmented fleet will launch more of them, faster, on top of the same entanglement. Meta's sequence was to standardize the models, build the run history and define the gates, and only then let agents execute. Meta's agents had 5,000+ shipped combinations and millions of logged runs to learn from. Your history will be smaller, which is a reason to start logging it in one place now.
The same lesson shows up in other production agent work we have covered. Meta's data-column onboarding agent worked after the team replaced one general agent with a thin orchestrator and deterministic routing. AT&T keeps human review on nearly 500 agentic workflows until they clear an accuracy threshold. In both cases the structure around the agent came first.
What to Do Before You Buy Agentic MLOps Tooling
Start with an inventory and a run history. An agent pilot comes after both.
This Week:
- Count your model families. List every production model with its owner, architecture, feature pipeline and training script. If more than half are one-offs, you are on Meta's expensive path, and the first investment is consolidation.
- Find your run history. Check whether past training runs, configs and outcomes live in one queryable place, such as an experiment tracker like Weights & Biases or the tracking layer in Databricks Mosaic AI. If they live in notebooks and Slack threads, no agent can start "from configurations likely to work."
- Write down your ship gate. Name the offline metric, the regression guardrails (cost, latency) and the online test a change must pass. If it is not written, an agent cannot be held to it.
This Month:
- Pick two or three model templates. Choose the largest group of similar models and agree a shared architecture, feature logic and training config for it. Meta's SMT post warns that a template that tries to keep "every historical edge case" will recreate the fragmentation it was meant to fix.
- Pilot propagation by hand once. Take one technique that already won on one model and push it through a template group manually, timing each step. That gives you the baseline Meta did not publish.
Before Next Budget Cycle:
- Scope the agent pilot to the long tail. Keep your highest-revenue models with experts, as Meta does, and point the agent at template models.
- Require plan approval with a compute estimate. Copy REA's rule: the agent proposes a plan and a GPU budget, an engineer approves, and runs halt at the threshold.
- Measure the pilot on Meta's two adoption metrics. Track human interventions per completed experiment and the share of failures the agent diagnosed and fixed itself. If the first number is not falling by the end of the pilot, the agent is adding supervision work. Our MLOps platform guide for regulated industries covers the sign-off and registry layer you will need when the agent's proposals reach model risk review.
The Bottom Line
Meta's posts give a detailed public account of agents doing ML engineering in production. Humans approve plans and ship decisions, agents do the multi-day execution between them, and the highest-revenue models stay with experts. The 60%, the 3x and the $45M all rest on a fleet that was standardized before an agent arrived. For an ML platform leader, the purchase order for agentic tooling belongs after the template work and the run history.
Count your model families this week; that number tells you which of Meta's two paths you are on.
Continue Reading
- Meta Cuts Data-Column Prep to Under an Hour With 14 Subagents
- AT&T Keeps Human Review on Nearly 500 AI Workflows
- Best MLOps for Regulated AI: Domino, Then a Sign-Off Layer
- Amazon Study Finds One Agent Is the Efficient Default
- Northern Trust Cuts Onboarding Data Prep 83% With AI Agent
- Why Meta Killed Its AI Leaderboard in 48 Hours
