Tomorrow Tuesday — September 2026 (Issue #10)
Welcome to another issue of Tomorrow Tuesday, where we cover the revolution running through film, video games, music and the creative industries.
Last issue of September. Is it too early to start wishing people a happy new year?
A newsletter by SOUTS, where we explore what happens when artificial intelligence stops being just a tool and starts changing the way we create and consume stories.
Today we’re doing something different: I’m going to take one topic and go a bit deeper into what it is and what it means. Today: AI agents in Hollywood.
What an agent is and how it’s used in video
We can sort the evolution of creative AI into three moments:
2021–2023: we generate images. Models like Stable Diffusion show up. This is the starting point, where the whole idea of prompting and generating an image appears.
2023–2025: video generation and workflows arrive. Generating video, extending it, modifying it, connecting shots. The boom of n8n, ComfyUI and the like. We combine tools to extend what we can do.
2025–today: agents. We delegate processes. Systems we give a goal and some room to decide how to reach it. The agent makes decisions and uses tools to get there.
Agents are LLMs placed in an environment where they can use tools, receive results and decide what to do next.
The key difference isn’t that they have a chat or are connected to tools, it’s that the system can adapt its path to what it finds along the way. In a fixed workflow the sequence is defined in advance. In an agent, the model has room to decide what the next step is.
For example: an automation could generate ten images and animate all of them. An agent could review those images, choose which ones to animate, redo one that doesn’t hold up and, if it comes out right, animate it.
The pieces of the puzzle
We care about agents for the audiovisual world. For that we need several pieces:
An LLM that organizes the work. The brain. It interprets what you ask for, plans how to get there and decides which tools to use.
The ability to observe.Being able to see images, hear music, watch videos, read transcripts and so on. Reviewing what someone says is not the same as reviewing how a character moves. Some LLMs can “see” images, others can’t, and a few (for now) can “see” video.
Tools. In our agents these can be image generators, but also tools to search the internet, run code or modify a project. The model chooses the action; the connected tool is what carries it out.
Instructions and context. The general instructions tell the agent how you want it to work; the context gives it what it needs for this particular job. In our case a brief, a script and references are things we can pass as context, along with guides on how to use certain tools.
Memory and project state.A record of what has happened so it doesn’t make the same mistake twice. It can be kept around for every step or stored and pulled in when it’s needed.
How do we use it in video?
Say we ask the agent:
“I want a thirty-second scene where someone is waiting for a phone call. It has to start uncomfortable and end relieved. Here’s the character, here are the references and here’s the budget. Show me the storyboard before generating the shots.”
An agent will break down the scene, propose framings and prepare a storyboard. At that point it could stop and ask for your approval, as we asked it to.
Then it calls the video generators and puts together a first cut. Depending on its tools and its ability to evaluate, it could spot a problem, attempt a fix or come back to you with a question.
For example: “This shot works visually, but the expression doesn’t convey relief. Should we try a different performance or change the framing?”
This is the kind of process we could build with these pieces.
The shift is going from asking for an image to delegating part of the work.
The present: what works and what doesn’t
What happened since last year is that LLMs got better, so agents got better, so every platform rushed out to build its own agent.
Runway. Its Agent is pitched as a tool for creating finished material. It also has an MCPthat lets you use its capabilities from external agents like Claude Code. It’s designed more for going start to finish inside this system.
Higgsfield. Its Production Skills bundle includes workflows for building scenes in Blender, organizing footage in Premiere and compositing shots in After Effects. The goal seems to be a project that someone can keep working on.
Adobe. Its agents live inside Creative Cloud, where they handle tasks that require juggling several Adobe programs, so creatives don’t have to leave the environment they know.
Pre-production is the new production
If you’re going to delegate part of the execution, what you cancontrol becomes much more important. That’s why everything that happens before the first shot is generated matters more and more.
Pre-production here means having a very clear script and well-established rules about what goes and what doesn’t. But it also means creating images of the character so it’s exactly how you imagined it, and doing the same with the space, the sound and the light.
Pre-production also includes planning how you’re going to tackle the content, and you can do that together with the agent.
Once pre-production is locked, the agent can start generating the videos with whatever degree of autonomy you want to give it.
Evaluating quality is another problem. Models can check that a file has a certain resolution or duration. When it comes to looking at an image or video with even a minimum of judgment, they slip badly.
Today this is the big bottleneck. Agents do things, but then they don’t know whether the thing is good or not.
And what about costs?
Less manual intervention doesn’t necessarily mean a lower total cost. Without guidance, agents can generate videos that end up being useless, or use more expensive models than you had in mind.
Runway has a smart feature where you tell it whether to prioritize speed, quality or cost. Based on that it decides which models to use, and with that, the cost.
Today, agents in the audiovisual world are still more of a bet than something that delivers value, although at SOUTS we have our own agents that we use quite a bit for specific tasks. But in the short term, the knowledge you can build here is probably what will give you the most value in this industry.
The future: does everything change, or nothing?
The unimaginative promise is that agents will take on more and more autonomy and start generating content with fewer humans, seeing what works, adapting what doesn’t and improving all the time. At Tomorrow Tuesday we find that vision pretty boring.
On one hand, the future we see is more assisted control: more like a creative partner and less like a roulette wheel. Yes, I want an agent to generate shots and prompts and run them, but I also want to be able to tell it “raise the camera a little and have the character exit more slowly.”
On the other hand, they will also get more autonomy, and that will make it easier for people to tell their ideas. This is especially powerful for fan communities. I see a future where IP owners provide guides for agents on what each character can and can’t do, so fans can create content that follows the creator’s vision.
The important thing is that not all content is the same. Many different ways of working will coexist. There will be projects where it makes sense to delegate almost an entire first draft. Others will need composition, performance and editing approved separately. And others will use agents only for organization, research or technical prep.
In the end, the distinction is which decisions we want to delegate and which ones we want to keep control over.
That was Tomorrow Tuesday, the weekly newsletter. If you want to subscribe, subscribe!
At SOUTSwe try to always stay on top of what’s happening, and every time there’s something relevant we’ll bring it to you in this newsletter.
See you next Tomorrow Tuesday.
Yves
