Scrubbing a video looks trivial, but the real problem is showing correct frames, frequently, while the user keeps moving the target across compositions of overlapping tracks and effects, under hard mobile device limits.
A scrub is a continuous stream of seek requests of different shapes, fast or slow, forward or backward, each resolving the same decisions differently: distance to the nearest keyframe, which frame to show and how soon, when to abandon work already in flight; across device tiers, content, and composition shape, no fixed policy covers that space.
In Edits we engineered an engine that serves frames fast, smoothly and accurately while requests keep arriving, interrupting stale work to render at every transition, seeking GOP-aware off a keyframe list, tracking the finger on backward scrubs while decoding runs the other way, and trading precision for speed where the user cannot tell; it delivered a 10x improvement on seek latency and frames shown per second.
No single policy covers the breadth, so we are going AI-native: refactoring into a portable system that plugs into Meta’s Family of Apps, developing agentic just-in-time observability into production behavior and skills that help coding agents extend the engine and validate their own work.
The vision is an agentic system that profiles production data at scale, generates algorithmic variations, and experiments to to evaluate what can ship, with a human on the loop.
