BlogBuild on

Why a canvas, when an agent can make the video?

4 min read

When Layer launched its Motion Designer Agent on 2 November, the announcement led with three negatives: "No video editing skills. No node-based workflows. No motion design software to learn." You brief the agent the way you'd brief a production team. It writes the script, plans the shots, adds voice, music and kinetic type, re-cuts the video for three ratios and makes hook variants, and you ask for changes "in the same thread, without starting over" (PocketGamer.biz).

It's a fair challenge to anyone building on a canvas. If an agent can make the video from a conversation, why should a team look at steps at all?

The two answers the industry gives

One camp says the steps are the problem. Sett tells UA teams that their job is becoming "managing a team of AI agents", where you look through the ideas, "approve most of them, reject what doesn't look right" and go for coffee, and it advertises that its agents build every playable "from scratch" because "templates are for amateurs" (Sett). Krea Agent sums up the same model: "the agent operates the tools, you direct and approve" (Krea).

The other camp says the steps are the work. Figma Weave's pitch is "build once, run at scale", and its co-founder Itay Schiff wrote this month that "generation isn't authorship": the model makes material, and people still have to shape it. In his view, workflows "are most powerful when they preserve the choices a creator has made, while leaving room to reopen those choices when the work demands it" (Figma).

Then the two camps merged

Look closely at this year's launches and the divide is mostly gone. Krea's Node Agent "turns a sentence into a runnable creative workflow": it shows a plan and a cost per step, and when you change something it reruns only what comes after the edit (Krea). Krea Agent, the chat-first one, can now save a workflow "as a skill and run it again later". FLORA's FAUNA is described as an AI creative agent built on a node canvas (AI Unfiltered). Layer sells both an agent and a Workflows product.

Engineers settled this earlier. Anthropic separates workflows, where models and tools follow "predefined code paths", from agents, which "dynamically direct their own processes", and advises using the simplest one that does the job (Anthropic). Andrej Karpathy's "autonomy slider" makes the same point from the user's side: how much the AI decides should depend on what's at stake (summary). A weekly batch of ads with prices, legal lines and approved footage is high-stakes, repeatable work, which is the textbook case for a workflow, with the AI doing the work inside it.

So the question that matters isn't agent or workflow. It's where your team's decisions live once the ad is made.

Three places the decisions can live

In one person's session. The prompt that got the product shot right, the model that rendered the character best and the reason the second take won are somewhere in a chat thread. Krea notes that its agent sessions are private, with no share link yet. The colleague who picks up the next brief starts a new session.

In the vendor's system. With a managed service, the agents learn from your results, and the learning stays with them. Brainlabs' Liz DeAngelis described the same trade with Meta's own automation: brands may see performance improve "without knowing what actually improved it". "I'm just not ready to outsource the learning" (PPC Land).

In a workflow the team owns. Each step keeps its prompt, its model, the material it started from and the take the team chose, and anyone on the team can open it. When the brand lead fixes the logo in the 9:16 step, it stays fixed for every batch that runs through it.

What a team needs that a session doesn't give

Everyone who makes a call can see the work. One ad has six people behind it: the designer, the junior, the UA manager, the brand lead, legal and the product marketer. In a chat one of them drives and the others review exports. On a canvas, legal's note sits on the disclaimer, the brand lead's on the logo and the UA manager's on the opening.

A fix changes one thing. Ask an agent for a change and you usually get a new version, which someone has to check for whatever else moved. On a canvas the change happens in the step where the problem is. Since 30 October that matters more, because Meta's revised terms make advertisers responsible for AI-generated output (ContentGrip).

Next week starts from this week. A session is finished when the video is. A workflow isn't. Next week's gameplay, offer or event art goes into the same steps, the winning opening leads, and every fix comes along.

The agent inside the workflow

Motn is built on that last answer, with the AI doing what agents do well. When new material comes in, Motn rewrites the prompts for it, sets the steps up again and generates several takes for each one, and Claude writes the text and motion graphics as code. What it doesn't do is hide the steps. Everyone on the plan sees them, can change any of them, and every change stays for the next run.

If you make one video at a time, a chat may be all you need. If your team ships ads every week and wants each week to start where the last one ended, an agent can make the ad, but only a workflow makes next week's ad better. The full case is in Don't rebuild. Build on. The side-by-side comparisons are in Motn vs Layer and Motn vs Sett. Or start with the $10 Try pack.

Make yours on Motn.