Why agent-made videos don't break: Ozor's three scene checks

When an AI assistant builds a video with Ozor, it writes real code: one React component per scene. Models write good code most of the time, and confidently wrong code some of the time: an icon that does not exist, an import from the wrong place, a property read on data that is not there. In a web app that is a bug report. In a video it is a black frame in the middle of your launch. So every scene an agent sends to Ozor goes through three checks before it is saved, voiced or charged, and none of them involve AI.

A scene is a function of the frame

An Ozor scene is a React component that receives the current frame number, the canvas size, its duration and its data, and returns what that frame looks like. It is rendered once per frame at 30 fps: in the editor preview, in the frame previews an agent asks for, and in the export that produces the MP4, possibly in a different order and on a different machine each time.

That one design choice explains most of the rules. If anything depends on wall-clock time, a random number or a CSS transition, frame 120 in the preview and frame 120 in the export will not match, and the video stutters or flickers. So everything that moves has to be computed from the frame number alone:

import { React, getProgress, ease, lerp } from '/shared-lib';

export const Scene = ({ width, height, frame, durationSec = 3, data }) => {
  const p = getProgress(frame, 6, 18, ease.outExpo);   // 0 → 1 over frames 6–24
  const drift = Math.sin(frame / 40) * 12;              // ambient motion, still pure
  return (
    <div style={{ width, height, background: data?.brand?.colors?.background || '#0E1116' }}>
      <h1 style={{ opacity: p, transform: `translateY(${lerp(40, 0, p) + drift}px)` }}>
        {data?.headline || 'Ship the story'}
      </h1>
    </div>
  );
};

Agents learn this from the scene guide they read at the start of every conversation. The checks make sure they actually followed it.

The three checks

They run in order, and the render check only runs when the code could possibly work.

The render check is the one that catches what code review misses. It renders the scene with the real data the agent passed and your real brand kit, at the real aspect ratio, so a headline that only exists in the agent's head or a brand color read without a fallback shows up as a crash now instead of a black frame later. It also catches the quiet failure: a scene that runs fine but draws nothing visible.

What happens when a check fails

Nothing is saved, nothing is voiced, and no credit is spent. The agent gets back a short, specific message it can act on:

Not saved: fix this and call add_scene again.
  failed check: rules
  ✗ [unknown-icon] `ChartGrowth` is not an icon in the renderer's lucide bundle,
    so `Icons.ChartGrowth` is undefined and the scene fails at first render.

Because videos are built one scene at a time, only the broken scene goes back and forth; the agent does not resend the whole video, which keeps your assistant's token use down. From your side, a failed check usually looks like nothing at all: the progress line for that scene arrives a few seconds later.

There is one escape hatch. For quality warnings the agent has genuinely tried to fix, like a frame that is blank on purpose, it may pass force=true. Anything that would crash the render or break the editor is never accepted, forced or not.

Beyond "it renders": does it look right?

Passing the checks means a scene works. Whether it is good is a design question, and Ozor gives the agent two tools for that. preview_scene renders the entrance, middle and exit frames (or up to four frames the agent picks) and returns them as images, so a model that can see can look at its own work and fix a cramped layout or unreadable contrast. validate_scene runs the three checks on draft code without saving anything, which is handy while iterating on one hard scene.

The scene guide carries the taste: motion that never freezes, staggered entrances, movement still going at the cut, text at least 22px for anything that matters, real content as the hero instead of icons. And the final judge is you, in the editor, where any scene can be changed by asking.

Why no AI in the checks

It would be easy to have a second model review every scene. We chose not to. A compiler, a rule checker and a real render give the same answer every time, finish in seconds, and cost nothing in tokens. They also answer the questions that matter for a video file (will it compile, will it render, will the export match the preview) with certainty instead of an opinion.

It is also what keeps the connector cheap. Ozor does not run a language model to build these videos at all; your assistant writes the code on your AI plan, and Ozor charges one credit per video (up to ten narrations) for the parts it adds. More on that split in Videos for agents.

Frequently asked questions