Meta says artificial intelligence agents now handle more than 60 percent of the work of upgrading its advertising models. The company reported the figure in an engineering post on Wednesday, alongside an estimated 45 million dollars in compute savings by the end of the year.
The Upgrade Treadmill
Meta’s advertising system runs on a large fleet of distinct ranking models, each with its own architecture, training data and serving limits. Every model needs the same proven improvements, yet a technique that works on one model almost never copies cleanly onto the next.
Meta calls that job technique propagation: adapting a proven improvement across the fleet. Techniques include new model architectures, richer data representations and training optimisations. Each adaptation is effort-intensive, and the opportunity grows faster than engineers can keep up while the fleet keeps multiplying.
Standardisation came first. Over two years, Meta organised its models into reusable templates, shipping more than 5,000 technique-model combinations while cutting per-model engineering time by 92 percent. But templates handle only the deterministic steps; the judgment calls still needed people.
Where the Agents Fit
That is where the agents come in. Meta built its propagation system around three capabilities: discovery, which ranks the techniques worth spreading; generalisation, which adapts a technique to new models; and adoption, which ships the result through training and testing.
The agents draw on a shared knowledge base of prior training runs, model architectures and technique outcomes. They decide which adaptations to try, diagnose failures, navigate hyperparameter searches and manage hundreds of concurrent jobs. Engineers review the assessments and make the decisions.
The highest-revenue models still get the human treatment. Meta says its experts hand-adapt techniques for those models, where the stakes justify the effort. Everywhere else, agents do the adapting — and the company is now making agent-driven propagation its primary method across the fleet.
Our Read
The numbers are Meta’s own, and they are estimates — early results, measured against its own earlier process, on selected combinations. But the direction is hard to dismiss: the company that sells the ads now lets agents run the machine that improves them.
Meta is not alone in handing agents the keys. Meta and OpenAI already account for 90 percent of agent traffic on the web, according to security researchers. The ads post shows what that traffic does inside the house: not browsing, but rebuilding the revenue engine itself.
Why It Matters
Advertising is Meta’s business, and ranking models decide which ads each person sees. If agents can propagate improvements three times faster while using a quarter of the training compute, every new technique reaches the whole fleet sooner — and the cost of staying ahead falls with it.
The contrast with Meta’s other agent news is striking. Its Muse agent was caught reading Signal chats and inventing bank balances this week. The same company now trusts agents with the upgrades to its most valuable models — a bet that supervision, not autonomy, is the difference.
Background
The post is the second in a series on Meta’s ranking agents; the first introduced the Ranking Engineer Agent, which generates hypotheses and proves new techniques. Meta aims to cut propagation time from over a year to three months, pushing agent workflows toward five times today’s throughput.
Source: Meta Engineering’s post


0 Comments