Meta hands ad compute allocation to AI agents: about 10% ranking gain, optimization loops cut from months to weeks

Ads ranking models are growing toward LLM scale faster than the hardware to run them, and the same unit of compute returns very different value across requests, models and chips, so something has to decide where compute goes across the whole serving stack. The earlier approach grouped requests into coarse segments by surface, context and ad creative representation, with engineers writing allocation policies for each group. That segmented approach multiplied rules with every new signal and could not capture nonlinear interactions among thousands of signals, which limited accuracy.

Intelligent Delivery uses a shared ML intelligence layer to estimate each request's marginal utility of compute, then layered decision engines turn those estimates into placement decisions. At its center is the Resource Core Model, a deep learning model with 6B sparse parameters and 20M dense parameters that runs once before candidate retrieval and is shared across the delivery funnel. On offline benchmarks across major surfaces, identifying the ad requests that benefit most from additional compute improved by about 20% relative to the previous production model.

Allocation is decomposed into layers: shifting traffic between regions, matching workloads to the best-fitting hardware, and dividing each host's capacity under a latency budget. The fleet spans Intel, AMD and Arm CPUs plus NVIDIA, AMD and Meta's own MTIA accelerators across several generations, and different stages are bound by different resources. Retrieval is bound by memory capacity, ranking by compute and model inference by accelerator throughput, so the system benchmarks efficiency for every workload-SKU pair and solves the assignment so scarce hardware goes to the workloads that convert it into the most useful work.

The optimization loop adds AI agents that read mechanism source code, design documents and a decade of experiment logs in a long context window and narrow the space to a handful of changes worth testing. Previously engineers relied on prior expertise to pick configurations, or used a constrained optimization approach that covered only a few key knobs and generalized poorly. In the first half of 2026 agents contributed to about 20% of Intelligent Delivery optimizations, and on the most automated mechanism they contribute to about 80% of the engineering effort, shortening the loop from months to weeks while engineers review and approve the final change before it ships.

The report covers deployment inside Meta's ads serving system: request-identification gains on offline benchmarks, allocation mechanisms across regions and hardware, and the share of configuration optimizations involving agents. These findings apply to that company's serving stack and hardware fleet, and no reproducible public experimental setup is given.

The roughly 10% ranking gain is a cumulative contribution from 2017 through the first half of 2026 on top of model and system advancements, not the effect of a single change. Agent-proposed changes still require engineer approval, and end-to-end real-time optimization and serving frontier ranking models on GPUs remain open directions.