The development
What happened
Meta has described Intelligent Delivery, its machine-learning system for directing limited serving capacity toward ad-ranking work expected to benefit most. Predictive models guide allocation across requests, hardware and ranking stages, while AI agents help investigate changes that engineers review before deployment. This is an account of an existing production system, not a new public service.
Key facts
A Resource Core Model uses six billion sparse and 20 million dense parameters before candidate retrieval. Specialized models estimate how resource settings affect expected outcomes, compute use and latency.
Decision engines coordinate serving regions, workload–hardware matching and per-request resources within a host’s latency budget. Meta says it tests their coupled effects in an end-to-end, closed-loop hardware environment.
Meta reports about 10% additional ads-ranking improvement across major surfaces from 2017 through the first half of 2026, beyond model and system advances. It is not a new 10% gain measured solely from adding agents.
A separate roughly 20% result concerns identifying high-impact requests in offline benchmarks against the previous production model. That metric is different from the system-wide ranking result.
Agents read code, design documents and experiment logs, propose changes and run experiments. Meta attributes about 20% of first-half-2026 optimizations to agents; engineers approve the final change before it ships.
The reported 80% engineering contribution applies only to Meta’s most automated mechanism, not the whole system. Combining request understanding and serving into one real-time model remains future work.
The context
Why it matters
The AI connection is operational: learned predictions decide where ranking models spend compute, and agents assist the optimization process. The account shows why faster individual models are only one part of serving efficiency. Its results remain Meta’s own measurements; they do not establish equivalent gains, lower total energy use or autonomous deployment elsewhere.