Meta’s Intelligent Delivery rations ad compute per request, and agents now do much of the tuning

Meta’s engineering team detailed Intelligent Delivery, a system that decides where each unit of compute goes across its ad-serving stack: which region handles a request, which chip it lands on, and how much of a machine it gets. Meta says it has added about 10% to ads ranking performance, on top of other model and system gains, from 2017 through the first half of 2026.

The fleet is heterogeneous: CPUs from Intel, AMD and Arm, accelerators from Nvidia, AMD and Meta’s own MTIA chips. Large ranking models like the Generative Ads Model (GEM) increasingly run on GPUs, which requires new allocation methods. Meta says plainly that its ads ranking models are growing toward LLM scale faster than hardware can keep up. When compute is the constraint, the winners squeeze the most useful work from each chip. It fits a series of Meta disclosures: harvesting idle accelerators for optional ranking work (coverage), the billion-per-second Andromeda retrieval stage (coverage) and the TLX GPU compiler (coverage).

Meta says agents contributed to about 20% of Intelligent Delivery optimizations in the first half of 2026, and on its most automated mechanism about 80% of the engineering effort, with engineers approving every change. An agent reads source code, design documents and a decade of experiment logs, proposes a few changes worth testing and runs the experiments. Meta says that cuts the optimization loop from months to weeks while keeping people in control.

That is a credible, useful pattern for agents: narrow domain, rich history, human sign-off. Nvidia describes something similar for its own data center operations (AI factory agents).

What’s being optimized is ad delivery. Better ranking for Meta means more effective targeting of people using its apps, and the post measures success in ranking performance, not anything users would choose for themselves. Efficiency gains flow toward the advertising business first, a pattern critics have noted in Meta’s consumer agent too (Muse review).

When agents do most of the tuning on a system this complex, fewer humans fully understand why it behaves as it does. Approving each change is a good guardrail, but reviewing a proposed change isn’t the same as designing it. How Meta keeps that expertise in house as the agent share climbs is the long-term question. Next on Meta’s list is fusing request understanding and serving into a single real-time model. More in our Meta coverage and marketing coverage.