
How agentic video editing turned a two-day ad edit into a few hours
Key takeaways
An agentic video editor is an AI agent that plans and carries out an edit with real editing tools, then takes your notes in plain language and revises. I run ours in Claude, with Magnific generating the footage, voice and music through its MCP connection. A three-minute, fully AI-generated video used to take one of our editors about two days. It now takes two to three hours, plus about an hour of notes, and I can run four edits at once in separate chats. It removes the editing bottleneck, but you still need a clear brief and someone who can judge a cut.
Editing is where ad production gets stuck
In most creative pipelines, editing slows things down far more than ideas do. Copy, scripts and static ads now come out of AI tools almost instantly, so video is where the queue builds up. A few months ago we relied on a team of video editors just to keep our pipeline healthy, and every new concept still waited in someone's queue.
When I asked one of our editors to cut a three-minute video built entirely from AI-generated footage, it took him about two days to deliver that one video. Generating the clips was quick. Assembling them, timing the voiceover, adding captions, finding music and exporting every size took the rest of the time.
I'm not the only one who sees it this way. In January 2026, a16z partner Justine Moore argued that editing is the bottleneck that limits how much good video gets made, and that agents could do for editing what Cursor did for coding. Her essay is called It's time for agentic video editing. That matches what we see on ad accounts. When a brand needs fresh concepts every week, editing capacity decides how many of them actually go live.
What an agentic video editor is
An agentic video editor is an AI agent that takes a goal in plain language, plans the edit and carries it out with real editing tools. It cuts the footage, places the captions, mixes the sound and exports the file. Then it takes your notes and revises, the way a human editor would. You stop dragging clips around a timeline and start describing the edit and reviewing the result.
That makes it different from the AI video tools most people already know. A video generator makes new footage from a prompt. An AI feature inside an editing app handles one task, like captions or removing silences, while you still run the timeline. An agent runs the whole job across many tools and comes back with a finished cut.
| Tool | What it does | What you still do |
|---|---|---|
| Video generator | Creates new clips from a prompt or an image, with models like Veo, Kling or Seedance | Assemble the clips, add captions, mix the sound and export the ad |
| AI features in an editing app | Handles one task at a time, such as captions or silence removal | Run the timeline and make every edit decision |
| Agentic video editor | Plans and runs the full edit across many tools, then revises from your notes | Write the brief, review each cut and give notes |
You'll also hear the phrase AI-native video editing, which describes editing built around an agent from the start rather than a traditional timeline with a few AI buttons added on. Some people on X have started calling it vibe editing, borrowing from vibe coding. The established tools are heading the same way. Adobe opened a public beta of its Creative Agent on June 18, 2026, including a Premiere assistant that organizes footage and assembles rough cuts from plain-language requests, according to Adobe's community team. OpusClip now offers AI Producer, an editing agent that installs as a plugin in Claude Code.
The stack I use
My agentic video editor is Claude, currently Opus 5.5. I run it in Claude Code on my Mac, where it works directly with the files in a project folder. It uses ffmpeg, a free command-line video tool, to cut, caption, mix, resize and export. It also writes the whole edit as a script, so every revision gets rebuilt from the same instructions instead of starting from scratch.
The editing itself doesn't need an MCP or a plugin, because Claude Code runs ffmpeg directly on the files. MCP is an open standard that lets an AI assistant plug into other services, and I only need it when Claude has to reach one. For me, that service is Magnific, the company formerly called Freepik. Through its MCP server, Claude can generate a clip, a voiceover or a music track without me leaving the chat. Magnific's MCP covers video generation, text to speech, music and sound effects, and it draws on the same credit balance as the app, according to Magnific's documentation.
Higgsfield also has an MCP connector that works with Claude and gives access to models like Veo, Kling and Seedance. I chose Magnific for two reasons. Its credit packages work out to better value for the volume we produce, and it gives me the Freepik stock library of images, video and music. Generated and stock material come through the same connection and end up in the same edit.
It helps to know how Claude sees the footage. It doesn't watch a video the way a person does. It reads a transcript of the voiceover and still frames it pulls from each clip. That is why it cuts dialogue cleanly on whole words, and also why I check every B-roll choice it makes.
How an edit runs from brief to export
Every edit follows the same five steps. Claude does most of the work in the second, third and fifth steps, and I own the brief and the notes.
- Brief
The concept, script, length, sizes, brand fonts and colors, references and anything that must appear on screen.
- Source
Claude generates clips, voiceover and music through Magnific and pulls stock where it fits. I approve the material before any cutting starts.
- First cut
Claude assembles the edit, cuts to the voiceover and the beat, adds captions and mixes the sound.
- Notes
I watch the cut on my phone and reply with specific changes tied to timestamps.
- Versions
Claude applies the notes and exports every size the placements need, with a cover frame for each.
Write the brief for an editor
Write the brief exactly as you would for a human editor, because the agent can't guess what the ad is for. Give it the concept, the script or voiceover, the target length, the sizes you need, your fonts and colors, a reference ad or two, and anything that has to appear, like the product name or a legal line.
Approve the source material first
Approve the clips before the cut starts. Regenerating one clip is cheap, but rebuilding a finished edit around a bad clip is not. Generate only what the edit needs, too. A folder of near-identical takes makes the edit slower, because someone still has to choose between them.
Let the agent make the first cut
Claude transcribes the voiceover so it can cut on whole words, times the cuts to the music, keeps captions inside the safe zones for Reels and Stories, sets the voice above the music and evens out the loudness. For a three-minute AI video, getting to a cut I would launch takes two to three hours.
Give notes with timestamps
This is where knowing a little about video editing pays off. Notes like "the opening is slow", "hold the product shot one more second" or "the caption covers her face at 0:12" get fixed in minutes, while a note like "make it pop" gets you nowhere. Because the edit is a script, a client note about the brand font or loud sound effects is a small change that gets rebuilt, not a fresh roll of the dice. Plan for about an hour of back and forth.
Export every version you need
Once the 9:16 cut is approved, Claude exports the 1:1 or 4:5 feed versions, reframes each shot so nothing important gets cropped and pulls a cover frame for each one.
Running four edits at once
The speed of one edit matters less than how many I can run at the same time. I can work on four edits in parallel, each in its own Claude chat.
Each chat gets its own folder, brief and source files, so the edits never mix. While one renders, I review another. My job has moved from operating one timeline to reviewing four.
| Measure | With a video editor | With an agentic video editor |
|---|---|---|
| Time to deliver | About two days | Two to three hours, plus about an hour of notes |
| Edits one person can run at once | One | Four, each in its own chat |
| What the person does | Cuts the video | Writes the brief, reviews each cut and gives notes |
What still needs a person
The most common complaint about AI editing tools is that you end up as the AI's intern, spending an hour fixing a one-click edit. In my experience that happens when the brief is thin and nobody approved the source material. An agentic editor edits what you give it. It won't rescue a weak concept, and it can't tell you whether an ad will persuade your customer. I never hand over these parts of the job.
- The concept and the script come first. The edit can only carry the message it's given, so that work belongs in creative strategy before anyone generates a frame.
- Taste decides the pacing, which B-roll supports a line, when to hold a shot and when to cut away. You need someone who can see what's wrong and say it clearly.
- Generated footage can warp hands, labels and packaging, so check the product in every shot and make sure the claims match what the brand can support. Transcription tools also mishear product names, so listen yourself before the captions go live.
- Watch every export on a phone at full speed, once with sound and once without.
When you want more control, use CapCut
If you want a timeline you can open and adjust by hand, you can ask the agent to build the edit in CapCut instead. I'd wait for the official integration before relying on it.
CapCut has no public editing API, so the community-built MCP servers that connect it to Claude work by writing CapCut's project files directly. One of them, capcut-mcp, tells you to close CapCut before it saves, because CapCut's autosave can overwrite the changes.
CapCut announced an official ChatGPT plugin at its CapCut World event in Los Angeles in October 2026. You describe the video in chat, CapCut builds an editable draft, and you keep refining it in the conversation or by hand in the editor. CapCut's own examples of chat notes are to hold a shot longer, make captions easier to read or keep music underneath a voiceover, which is exactly how we brief Claude. The company also says a more agentic experience is coming to CapCut itself. Its product page says the connection runs over MCP and is built for the ChatGPT desktop app, with availability that varies by region. OnlineMarketing.de reports an open beta planned for November 2026, though CapCut hasn't set a date. Once it's widely available, it will be the safer way to give an agent a timeline you can still open yourself.
Build it in-house or hand it off
If you have someone who can write a brief and give clear notes, start small. Pick one concept, write the brief, approve the source clips and run one edit in one chat before you try four. Judge the result by the cost of each approved ad, counting generation credits, rerolls and the time spent on notes, rather than by how good one demo looks.
If editing is your bottleneck and you'd rather receive finished ads, that is what our AI Creative Production service does. You bring the approved concept and script. We handle the visual production, product detail and editing for your Meta campaigns, and we cover the AI platform costs. You can see finished work in our ad portfolio and in the Madbury Road case study, where we made AI video ads for a furniture brand with no existing footage.
If you need the concepts as well, Creative Strategy covers the research, concepts and briefs behind each ad.
Putting it into practice
- Start with one concept in one chat, with a written brief and approved source clips, before you run edits in parallel.
- Give notes the way you would to a human editor, with timestamps and specific changes.
- Keep someone who understands video in the review seat. The agent saves the editing hours, but a person still has to judge whether the cut works.
Book a call

