Netflix engineers AI Discover with a new intelligence layer for steerable TV recommendations
·2026.11.18 02:01
Key point
The system combines LLMs with existing recommendation engines to allow members to pick themes and tune results without typing.
Details
- Netflix introduced AI Discover, a new tab in its TV app that shifts from silent background ranking to steerable discovery where members select themes like "Quick Watches" to get matching titles and explanations.
- The feature relies on a new intelligence layer, or agent harness, that plans and executes requests by combining LLMs with existing recommendation systems to ensure speed, reliability, and safety on TV hardware.
- The architecture uses a plan-and-execute approach where a controller validates a graph of approved steps before an executor runs them, ensuring every action stays within code-defined limits and falls back to tested plans if LLM calls fail.
- Safety and personalization are enforced through code rather than prompts, using a persistent memory store called Memory Hub to store member taste summaries and explicit user preferences like "don't recommend."
- The system is designed to be model-agnostic, allowing Netflix to swap LLMs and capabilities via server-side configuration without requiring new TV app releases or complex migrations.
- Future work focuses on measuring how much members want to shape recommendations and optimizing cost and latency as traffic ramps up, with a long-term goal of a self-improving system that proposes and tests its own improvements.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.