Meta said artificial intelligence agents reduced median training compute fourfold across selected combinations of advertising models and techniques, describing the results in an engineering article published on Nov. 18.
The company said engineers could also handle three times as many model adaptations with agent assistance. The article, “How Meta Uses AI Agents to Scale ML Innovations Across Ads Ranking Models,” describes work on the models that select ads across Meta’s products.
Meta, which sells advertising, said its recommendation system uses many models with different designs, training data and operating constraints. A technique that improves one model usually requires changes before it can be used on another, the company said.
The article did not specify the period covered by the agent results or identify the selected model and technique combinations used to measure training compute. Meta described its results as early and said it was expanding agent-driven workflows.
The company said agents draw on records of millions of earlier training runs and the properties of each model to choose configurations to test. They use successive training results to narrow the search and decide which experiments to continue or stop.
Separately, Meta said standardizing and automating model design had allowed it to ship more than 5,000 combinations of techniques and models over the past two years. The company said that work reduced engineering time for each model by 92%.
That earlier work organized models around reusable templates, allowing a technique checked at the template level to pass automatically to models built from it. Meta said engineers still had to make decisions about which techniques to apply and how to resolve failures.
The company said agents now assess techniques across its models and rank candidates for engineers to review. They also help adapt techniques to different models, run experiments, monitor training and diagnose failures.
Agents can oversee hundreds of training jobs at once and rank them by factors including revenue importance and deadlines, Meta said. Engineers review decisions, set priorities and approve what is shipped.
Meta said agent-assisted techniques face the same tests as those developed by engineers. They must improve measures of predictive performance without breaching limits such as training or operating costs, then show an improvement that meets a test for statistical significance in a live comparison test.
The company said manual adaptation by specialist engineers produces the strongest results and remains the method for its highest-revenue models. It also said getting every proven technique to every suitable model can currently take more than a year.


