Skip to content
AI.info

The Pulse

Meta says AI agents handle most ads-ranking technique transfers

Meta estimates that AI agents now handle more than 60% of the work of adapting proven techniques across its ads-ranking models. The company says early results include lower training compute on selected combinations and projects more than $45 million in compute savings by year-end.

Meta says AI agents handle most ads-ranking technique transfers

AI agents now handle more than 60% of the work of adapting proven improvements across Meta's ads-ranking models, the company estimates. The agents began driving that work in March 2026, according to an engineering post published on November 18.

Source engineering.fb.com

The company estimates more than $45 million in compute savings by the end of the year. Compared with the same work before agents were used, Meta reports early results of three times greater engineering productivity and a median fourfold reduction in training compute across selected technique-model combinations.

Moving improvements between models

Meta's ads recommendation system uses distinct ranking models for different product surfaces and advertiser objectives. Their architectures, training data and serving constraints differ, so an improvement proven on one model generally needs adaptation before it can be used on another.

Those improvements, which Meta calls techniques, include changes to model architecture, data representations, training and inference efficiency. Engineers have traditionally studied how a technique interacts with each target model, then adapted and tuned it for that model.

The company has also standardized model designs around reusable templates. Over the last two years, that approach and automation shipped more than 5,000 technique-model combinations across its model fleet, Meta says. The company says it reduced engineering time per model by 92% over that period.

The newer system uses agents for decisions that remain after the deterministic steps are automated. Those include selecting which techniques to try, diagnosing failures and deciding how broadly to search possible settings.

From discovery to adoption

In the discovery stage, agents draw on a shared knowledge base of previous training runs, model architectures and technique outcomes. They identify proven techniques, estimate how well each may transfer and rank candidates by expected impact; engineers review the assessments and decide what to pursue.

Generalization is the work of adapting a technique to models beyond the one for which it was developed. For models built from shared templates, agents propose optimization hypotheses, run experiments on representative models, evaluate results and diagnose failures. A technique validated at the template level can then reach models built from that template.

For its highest-revenue models, Meta still customizes techniques for each model. Agents learn alongside engineers during initial adaptations, recording configuration changes and outcomes, then propose and test adaptations for remaining models for engineers to review and steer.

Agents also draw on millions of prior training runs to select starting configurations, narrow the search for model settings after each round and decide which experiments to continue or stop early. The reported median reduction in training compute applies to selected technique-model combinations, rather than the entire fleet.

Running the experiments

At the adoption stage, agents launch experiments, manage training runs lasting multiple days, monitor convergence and assemble validated proposals. They also diagnose infrastructure, training and data-pipeline failures and apply fixes or retry jobs.

When many workflows run concurrently, agents rank jobs using factors including revenue importance, technique potential and deadline proximity. They can reclaim capacity from stalled or low-value runs, according to the company.

The work follows Meta's earlier development of the Ranking Engineer Agent, or REA, which generates hypotheses and runs experiments on ads-ranking models. Its hypothesis process draws on a database of past experiments and a research component that examines model configurations and proposes optimization strategies.

Before a plan runs, REA estimates its compute cost and confirms the approach with an engineer. It can wait while a training job runs and resume when the job finishes, while engineers retain oversight at strategic decision points.

Sources

  1. Engineering at Meta
  2. Engineering at Meta

Companies in this story

Related stories

Explore

More articles