Netflix details the 'intelligence layer' behind its AI Discover TV tab
In a Netflix TechBlog post, engineers describe how AI Discover, a TV-app tab being tested with a small group of members, combines LLMs with existing recommendation systems using a plan-then-execute agent harness, asynchronous page generation and safety rules enforced in code.
Netflix has published an engineering account of AI Discover, a new tab in its TV app that the company says is available to a small group of members. Instead of relying only on recommendations inferred silently from viewing history, the tab lets members pick a theme, such as "Quick Watches," and see a row of matching titles, a short explanation and narrower themes to explore next. [1]
According to the post, the feature runs on what Netflix calls an intelligence layer, an agent harness that decides for each request how to combine large language models with the company's existing recommendation systems. The authors say they rejected the simplest design, handing the whole request to an LLM, because it would rely on probabilistic output where Netflix needs governed results. [1]
The layer has two parts, the company says. A controller gathers context, settles on a plan and validates it, and an executor runs it. Plans are graphs of approved steps called capabilities, and most of these wrap existing Netflix systems such as search, ranking and catalog metadata. Netflix describes the approach as plan-and-execute: an LLM may propose a plan, but code decides whether it runs, and if the planner is slow or its plan fails a check, the system falls back to a pre-tested plan. [1]
Because generation can take several seconds while a typical TV request returns in moments, Netflix says it made generative requests asynchronous. A button press is acknowledged immediately, the TV shows a thinking state, and the server notifies the device when the result is ready. [1]
Netflix says rules such as content and maturity standards and members' requests not to see certain titles are enforced in code rather than in prompts. The planner sees a stripped-down request and uses placeholders that are filled with request values only after validation, which the company says also protects against prompt injection. Each request leaves a structured record used for debugging and offline evaluation. [1]
The post also says all LLMs sit behind one common interface, so models can be swapped through server-side configuration without a TV app release. Member taste summaries are generated ahead of time by MemberLLM, an in-house model, and stored with member feedback in a memory store called Memory Hub. [1]
In this version there is no free-text input or open-ended chat; members select themes and use tuning options such as "don't recommend." Netflix says it ran safety evaluations on generated theme labels, row titles and explanations before launch. The company says it is still studying how cost and latency hold up as traffic ramps and how long members will wait for generated results. These claims come from Netflix's own blog and have not been independently verified.
Sources
- Netflix: Engineering AI DiscoverSource published · TechDots found it
Statements labelled as a company's claim have not been independently confirmed by TechDots.
Get Technology stories by email
Related coverage
Spotify Engineering questions the idea of a single 'AI product manager' role
A Nov. 9 post on Spotify's engineering blog examines AI product management from three angles, starting from the observation that most coverage treats 'the AI PM' as one job. Only an excerpt was available for review.
Stripe says it moved $1B in vendor payments onto its own Global Payouts product
In a company blog post, Stripe says it began shifting its vendor payments to Global Payouts in early 2026, cutting new-market setup from six weeks to one with a 99.9% payout success rate. The figures are Stripe's own and have not been independently verified.
Stripe says users can now query their Stripe data in natural language through an agent
In a 5 November blog post, Stripe said customers can use natural language with an agent to analyse their Stripe data. Its examples include building an investor slide deck and a custom revenue dashboard. Details beyond the announcement are not available.