Netflix has published an engineering account of AI Discover, a new tab in its TV app where a small group of members can try new ways to find something to watch. The post, attributed to Varun Khaitan, Peng Ren, Gary Tang, Itay Teller, Atif Faheem, Nipun Kumar, Donnie DeBoer, Eduardo Lundgren, Matteo Rinaldi and Raghav Bakshi, describes the intelligence layer behind the experience and the decisions that shaped it.
AI Discover moves from silent ranking toward steerable discovery. A member picks a theme, such as "Quick Watches," and receives a row of matching titles, a short note on why they fit, and a few narrower themes to try next. The engineering question, per the Netflix TechBlog, was how to deliver a better member experience with generative AI on TV without sacrificing speed, reliability or safety.
Why it matters
Every Netflix page combines recommendations with rules. Eligibility, maturity and safety rules decide what can appear. An experience built on LLMs needs the same rules plus code that decides when to call an LLM and how to handle its output. Netflix calls that code an agent harness, and it is the core of the intelligence layer.
The post argues against the simplest implementation, which hands the entire request to an LLM and lets it pick titles, write labels and decide what happens next. That approach relies on probabilistic output where governed results are needed, because every turn in AI Discover ends on a screen in someone's living room. At Netflix's scale, even rare failures happen many times a day, and a member sees every slow call, unavailable title or off-brand line.
Technical details
Netflix lists seven challenges that shaped the design: fitting LLMs into a product and architecture designed without them; keeping recommendations grounded in a real, available catalog; working with slow, unpredictable LLM calls; adapting to fast-changing models, tools and prompts; building for TV input and hardware; holding a safety and reliability bar at scale; and learning from what members say alongside viewing behavior.
When a member picks a theme, the TV app sends the request to the Page Server, which builds pages for all Netflix apps and hands the request asynchronously to the intelligence layer. That layer has two components: a controller and an executor.
The controller gathers context, including real-time signals such as the time of day, and settles on a plan, either an established one or a new one drafted by its planner. It validates the plan and retrieves only the member details the plan requires from Memory Hub, the per-member memory store. Those details may include member taste summaries generated by MemberLLM, Netflix's in-house language model for personalized recommendations, or feedback shared by the member.
A plan is a graph of approved steps, each running a capability. A capability is a named unit of work that can have multiple versions and belongs to a family such as retrieval, ranking, generation or packaging. It declares its inputs and outputs, the context it may see and what happens if it fails. Before execution, the controller validates the graph: every step is approved, every input has a producer, nothing loops and the plan ends in exactly one page. Afterward, the controller checks the result before it is served.
The executor runs the plan. Steps pass typed outputs through declared contracts, independent steps run in parallel with their own deadlines and fallback behaviors, and some steps are optimized LLM calls with one task each, such as writing row titles or suggesting next themes. Netflix describes the combination of pre-run checks and in-run contracts as governed execution.
What to watch
Netflix says LLM calls are the slowest and most expensive part of the system, with tail latencies that can spike and output that may not match the request. Its response is to use LLMs selectively, treat their output as untrusted input, and test performance and quality before shipping. The post also notes that TV apps update slowly, so changes must ship without an app release, and that members must always get a page even when an LLM call fails or yields poor results, with structured records showing why something broke.




