Meta AI agents now handle more than 60% of the work of moving machine learning improvements across the company’s ads ranking models. Since agents began driving execution in March 2026, Meta estimates the shift has tripled engineering productivity and will save more than $45 million in compute by the end of the year.
The company calls the process technique propagation, and it has become a bottleneck worth automating.
The problem with spreading a good idea
Meta’s ads system runs many distinct ranking models, each serving a different product surface or advertiser goal. Each has its own architecture, training data and serving limits. A “technique” is anything that makes a model better, such as Wukong factorization layers, Interformer interaction networks, richer data representations or cheaper training and inference.
A technique that works on one model almost never copy-pastes onto the next. Speed matters as well, because techniques have a shelf life. A gain today may be superseded within a couple of model development cycles, so the sooner a proven idea reaches the whole fleet, the more it is worth.
Two limits slow things down. Adapting a technique by hand takes a domain expert who knows both the idea and the target model, so throughput is capped by engineers. And every attempt needs a training run, so compute is finite.

What came before the agents
Meta first standardized its model design with templates built from reusable blocks, so a technique validated once at the template level reaches every model made from that template. Over two years, this approach shipped more than 5,000 technique-model combinations and cut per-model engineering time by 92%.
Judgment calls were left over: which technique deserves which model, how wide a hyperparameter search should be, why a model isn’t improving, and how to handle failures. Those still needed people. For its highest-revenue models, engineers still adapt techniques by hand.

Where the agents step in
Meta’s agents work in three stages and share one knowledge base of past training runs, model architectures and technique outcomes.
- Discovery: agents track which techniques have been proven, estimate how well each might transfer, and rank technique-model candidates by expected impact. Engineers review the list instead of building it.
- Generalization: agents rework a proven technique for models with different architectures or features. For template-based models they improve the templates themselves. For bespoke models they learn from an engineer’s first few adaptations, then propose the rest.
- Adoption: agents launch experiments, manage multi-day training runs, monitor convergence, diagnose failures and assemble proposals ready to land.
Agents also decide where compute goes. They rank hundreds of concurrent jobs by revenue importance, technique potential and deadline, and pull capacity from stalled or low-value runs so GPUs stay on the most valuable work.
The reported results
Drawing on millions of prior training runs, the agents start from the configurations most likely to work, narrow the hyperparameter search after each round and stop weak experiments early. Across selected technique-model combinations, Meta saw a median 4x reduction in training compute. Each engineer can now handle three times as many adaptations.
All figures are Meta’s own estimates, measured against the same work done before agents.
How Meta checks the agents
Agent-propagated techniques face the same bar as human ones. They must improve offline predictive metrics such as normalized entropy without regressing guardrails like training or inference cost, then win a statistically significant online A/B test before shipping.
Because a final check comes too late to steer work that runs for days, each stage has its own measure. For discovery, Meta tracks precision and recall of the candidates, and weighs recall more, since a model missed there never gets another look. Generalized techniques are first tested on scaled-down models, then on a representative set of full-size ones.
What happens next
Meta says its current investment is in global compute optimization, the question of which runs deserve a fixed budget. The stated goal is for every proven technique to reach every applicable model within a few weeks. If the numbers hold, Meta AI agents will keep taking over the routine parts of ads model development, while engineers approve plans and set priorities.
Source: Engineering at Meta