Watching what’s happening to video production right now feels like rewatching a movie I’ve already seen. A few years ago, agents came for software engineering and everyone argued about whether “agentic” meant anything or was just a word consultants used to raise their rates. Then, quietly, the job changed shape. The same thing is starting in filmmaking, and this time I get to watch from the front row: I’m using it to make marketing videos for clients and, less defensibly, a series of short YouTube documentaries about trains.
I’ve started calling this agentic filmmaking, because as far as I can tell nobody has named it yet, and I’d like to plant the flag before someone at a conference does.
The slot machine and the crew
When most people hear “AI video,” they picture the generators. Type a prompt, pull the lever, get eight seconds of footage. Maybe it’s stunning. Maybe the train has eleven wheels and merges with a cow at the four-second mark. You pull the lever again.
That’s a slot machine. A very impressive slot machine, but the defining feature is that you have no idea what comes out until it comes out, and your only recourse is to pull again.
Agentic filmmaking is a different arrangement entirely. You hand an agent actual production work: here’s forty minutes of raw footage, here’s what the piece is about, here’s roughly how it should feel. The agent transcribes the audio, finds the usable takes, plans a cut, builds the title cards, syncs the captions, grades the color. Then, and this is the part that matters, it watches its own output, notices the lower-third is covering the interview subject’s face, and fixes it before showing you anything.
You review the cut. You give notes. It does another pass.
That’s not a slot machine. That’s a crew. A weird, tireless, occasionally confused crew that works at 4am and never asks about lunch, but a crew. Your job quietly changes from operator to director.
Why trains
I do video work for clients now, mostly marketing pieces. But you should never test a new workflow on client work alone, for the same reason you don’t test a new deploy process on Black Friday. So I picked a testbed I actually care about: trains.
I love trains. I have always loved trains. There is no business case. I’m producing a series of short documentaries about them because the subject makes me happy and because a passion project is the perfect proving ground: I know exactly what good looks like, and if an episode fails, the only stakeholder who gets hurt is me.
The workflow surprised me more than any demo ever has. I sit in a terminal and talk. Transcribe this interview. Cut everything before the second question, the first answer was me rambling about gauge widths. Build a title card in the style of the last one. Tighten the pacing in the middle, it sags. The agent does the work, renders a preview, looks at the preview, and comes back with the result plus notes on what it changed.
At some point I realized I had produced a finished, captioned, color-graded short doc and had never once opened a timeline. The part of my life where I scrub footage at 2x while my eyes glaze over just didn’t happen. The craft decisions were all still mine. The labor wasn’t.
The part the SFX world won’t enjoy
I watched this exact movie play out in software, which is why I recognize the plot.
When agents arrived for code, the panic was about replacement: the machines are coming for the programmers. That’s not quite what happened. What happened is the crew size collapsed. Work that took a team of five now takes one person with taste and an agent, and the person with taste became more valuable, not less.
Video production is structured the same way, which is what makes it so exposed. A five-person post house is mostly a pipeline: someone doing rough cuts, someone doing rotoscoping and cleanup, someone syncing and conforming, someone comping the shots the client will never look at closely. Almost none of that is the creative core of the work. It’s the labor around the creative core, and labor around a creative core is precisely what agents eat first.
The directors, the editors with a genuine point of view, the VFX supervisors who know why a shot feels wrong: I think they end up like the senior engineers did, doing more of the interesting work with less grinding underneath it. The rough-cut-and-cleanup middle of the pipeline is the part that should be nervous.
Does “agentic” deserve the gravitas?
I asked this same question about agentic engineering, so it’s only fair to ask it here: agentic filmmaking sounds like it has gravitas, but does it really?
Half of what gets called “agentic” in any field is a cron job with a marketing budget. So here’s the threshold I’d hold the term to: the loop. Does the system look at its own output, judge it against the goal, and iterate without you prompting each step? If you’re approving every individual operation, you don’t have an agent. You have a render button with a chat window attached, and there’s no shame in that, but it’s not the thing I’m describing.
The loop is what changed my work. Not the generation, the review. The agent watching its own render and catching the awkward cut before I do is the moment this stopped feeling like a tool and started feeling like a collaborator.
Plant your own flag
I don’t think agentic filmmaking is a 2030 prediction. I made a train documentary with it last month, and this morning I gave notes to a crew that doesn’t exist.
If you make video for a living, my honest advice is the same advice I gave developers a few years ago: don’t wait for the industry to settle the vocabulary. Pick a project you love, something with zero stakes and maximum joy, and run it through an agent-driven pipeline end to end. You’ll learn where it faceplants, and what your taste is actually worth once the labor gets cheap.
Mine was trains. Yours will be better. It’s hard to be better than trains, though.
Shameless plug: At Victoria Garland we build Shopify infrastructure for a living, but clients kept asking for marketing video, so now the crew that doesn’t exist works for them too.