Building an Agentic Video Editor for Mass Consumer

How many people recorded a photo or video at the conference—and how many actually posted a video? Ekaterina Deyneka opens with those two questions. Fewer hands remain for the second. She recognizes the pattern from her own recordings of conferences, events, trips, and meetups: capturing material is easy enough; turning it into something worth sharing takes work.

Video editing is tedious, largely manual, and dependent on artistic judgment. Reelful, where Deyneka is founder and CEO, approaches that gap as an agentic editing problem: let the user supply the material while the agent does the editing work that otherwise keeps it in the camera roll.

The starting interface is simple: drop in photos and videos, then provide context about what happened or directions for the edit. A request might ask for captions, music, or a voiceover. The agent must understand the media, find the right moments, assemble them, and produce the requested supporting material—including B-roll—before returning a ready-to-share clip.

A speaking-to-camera recording makes the editing task especially concrete. The source may contain long pauses and unsuccessful takes. The desired output keeps the useful speech and removes those interruptions, without requiring the user to locate every cut. Deyneka follows this example with an initial edited-video preview: the product’s output is a finished clip, rather than instructions for making one.

The infrastructure resembles an agentic app builder. Both begin with a user request and start a remote machine—a sandbox—where an agent works with tools and skills. In the video editor, the request also includes media. Deyneka corrects a duplicated heading on the comparison slide: its second column should read “agentic video editor.”

The useful architectural connection is the working artifact: an agent modifies a structured project inside a remote environment, and the system turns that project into something the user can inspect.

The shared infrastructure does not make the creative problems identical. Reelful focuses on editing personal, real-life footage and generates relatively little of the content. Deyneka considers this harder than working from a blank canvas: the agent has to choose the best available moments, decide what to omit, and organize what remains into a coherent sequence.

The footage may also be messy or incomplete. The quality target remains a polished, professionally edited result—ideally one whose viewers cannot tell whether a person or an AI edited it. That is the aspiration guiding selection and assembly, not a measured result presented in the talk.

The workflow separates deciding what to make from executing the edit:

1. Understand the media. Identify what happens in the supplied clips and photos. Transcribe speech when the input includes speaking-to-camera footage.
2. Propose a creative plan. Let the user approve it, request changes, or regenerate it before editing begins.
3. Start the execution environment. Once the user approves the plan, spin up the sandbox in which the agent will work.

The approval step gives the user a chance to correct the direction before the agent begins assembling the composition.

Inside the sandbox, the agent brings skills that encode editing judgment. Deyneka’s examples locate taste and craft in specific decisions:

• Cut rules: how to select the best moments.
• Font pairs: which combinations suit a particular use case, and which do not.
• B-roll instructions: how to generate supplementary footage for the edit.

These skills guide how the available material becomes a composition.

The agent can also initiate subprocesses to generate music that fits the exact composition, create a voiceover, add sounds, or animate images. Supplied photos can therefore become moving elements rather than remaining static throughout the video. These assets feed into the next stage: the Remotion composition.

Remotion makes video composition programmable in React. Deyneka describes the representation in terms of assets, tracks, and their order. More precisely, a composition combines a React component with video metadata such as frame rate, dimensions, and duration. Although she calls the framework open source, its published source currently uses a special license with eligibility and commercial-use conditions.

A code-based composition turns editing decisions into a task a coding agent can perform. For example, a small React composition can express two supplied photos appearing in sequence. Here, from and durationInFrames define the timeline, while the composition supplies the output metadata:

import {
  AbsoluteFill,
  Composition,
  Img,
  Sequence,
  registerRoot,
  staticFile,
} from 'remotion';

const PhotoEdit = () => (
  <AbsoluteFill style={{backgroundColor: 'black'}}>
    <Sequence from={0} durationInFrames={60}>
      <Img
        src={staticFile('arrival.jpg')}
        style={{width: '100%', height: '100%', objectFit: 'cover'}}
      />
    </Sequence>
    <Sequence from={60} durationInFrames={90}>
      <Img
        src={staticFile('gathering.jpg')}
        style={{width: '100%', height: '100%', objectFit: 'cover'}}
      />
    </Sequence>
  </AbsoluteFill>
);

const Root = () => (
  <Composition
    id="PhotoEdit"
    component={PhotoEdit}
    durationInFrames={150}
    fps={30}
    width={1080}
    height={1920}
  />
);

registerRoot(Root);

The filenames and timing illustrate the representation; Reelful’s agent must derive the actual asset choices and arrangement from the media and approved plan. Deyneka’s reason for choosing this approach is that agents are already good at writing code.

Code generation is followed by a verification layer. It checks that the composition is clean, well-defined, and renderable. If it finds problems, the agent iterates on the composition before producing the polished clip. The described loop checks the generated artifact rather than assuming the first composition will work; the talk does not specify the individual checks or their implementation.

Media understanding, planning, sandbox execution, asset generation, composition, and verification form a complex workflow. The consumer should not need to understand all of it. Reelful’s first interface choice is mobile-first creation, intended to fit into everyday activity. Deyneka mentions walking and lifting weights, and also names driving as a usage scenario; interaction with an editor should wait until safely parked.

The next two choices address different sources of friction:

• Directional templates reduce prompting work. Users can select a speaking-to-camera, B-roll, or voiceover direction and drop in their media. Deyneka says this can work without an explicit prompt.
• A built-in editor preserves familiar control. After agentic generation, users can make small manual changes, such as removing a second or correcting a word in the captions.

The agent produces the initial video, while the familiar editor handles precise adjustments. Deyneka then turns to videos she recently created with Reelful.

The played example begins with an invitation to an exclusive Cursor Creators Dinner. Playback pauses briefly while Deyneka checks whether the audience can hear it. The narration then describes a custom event space transformed into a tropical sunset setting, establishing the venue before moving through the experience.

Next come the food and atmosphere: private chefs from Los Angeles, lobster, wagyu, crab, and strawberry ice cream. The recap moves on to conversations about people’s work and the ease of connecting with the community, then reaches gifts before playback cuts off. It is a concrete example of the desired output: an event recording assembled into a narrated social recap.

Deyneka says these videos were assembled using only the agent, without a regular video editor, and that she was already posting them on social media. That connects the demonstration to the opening problem: recordings that previously remained unshared had become material she was publishing.

At the conference, she offers access to a beta of Reelful’s second version. The product is still early and under active development, and she invites feedback, contact by email, and accounts of how people actually use it. She also announces recent funding from a16z Speedrun.

Then she asks to film a video with the audience. The talk ends where it began—with someone capturing an event, now with a workflow for turning that recording into something shareable.