A new ML technique that boosts an ads ranking model’s performance is only as valuable as the speed and scale at which it reaches production. At Meta, the ads recommendation system runs on a fleet of distinct ranking models—each tailored to different products, data, and constraints. Rolling out a proven innovation across all of them used to be slow, manual, and expensive. Since deploying AI agents to drive this process, Meta now credits agents with over 60% of all technique propagation, delivering a 3x gain in engineering productivity and a projected $45M+ in compute savings this year.
The challenge is combinatorial: every new architectural advance or training optimization—like Wukong factorization layers or Interformer interaction networks—must be adapted to dozens or hundreds of model variants. Historically, this meant domain experts hand-tuning each adaptation, a process that capped throughput and burned engineering bandwidth. Compute was another bottleneck: every adaptation attempt required a full training run, and the space of possible combinations grows faster than the available budget.
Standardizing Model Templates and Automating Propagation
Meta’s first step was to standardise model design with templates. Models are now organised into a set of reusable building blocks. When a new technique is validated at the template level, it can propagate automatically to every model built from that template. Over two years, this approach shipped over 5,000 technique-model combinations and cut per-model engineering time by 92%.
But automation only handles the deterministic parts. The real bottleneck was still human judgment: deciding which techniques to propagate where, how to tune hyperparameters, and how to debug failures. These steps kept the process slow and resource-intensive for the long tail of models.
AI Agents Drive Discovery, Generalisation, and Adoption
The current system centres on AI agents with three core capabilities:
- Discovery: Agents maintain a live inventory of proven techniques and estimate their transferability, ranking technique-model candidates by expected impact.




