Automate your film production with AI - invideo Agent 2
Can AI agents actually help manage a bigger creative project? We're testing invideo Agent Two to build a roughly 3-minute short from the ground up.
I started by putting the characters, cars, locations and visual direction into Context, then used Briefs for individual scenes. From there I built a small crew of Expert Agents: a writer developed the story, a storyboard artist broke it into shots, and a DOP worked out the cinematography. I used Playbooks to give each agent persistent rules, project Memory to keep track of decisions, and Ultra for the heavier generation work.
Then I gave Agent Two the finished video material and had it check for continuity issues across the project.
Try invideo Agent Two:
https://invideo.io/i/SebastianKamph
Closed Caption:
AI agents have gone from a cool idea to something
people are actually using day-to-day in just the
last few months. I mean, I use them all the time.
Coding agents can now work directly with GitHub
and issues and pull requests. And with MCP, they
can connect into tools like Figma and actually
read or even modify real project files. It's
pretty cool. And now that same agent idea is
starting to move into creative work. And that's
what invideo Agent 2 is trying to do for video.
Not just generate something from a prompt, but
actually understand a project, remember it,
and work through it with you, which sounds
pretty cool. Now, invideo Agent 2 is four
times faster and half the price compared to the
previous version, but that's just comparing to
their own stuff. They are sponsoring this video,
so let's actually put it to work and see if it's
any good. Now, what I want to do today is actually
use it on a project rather than sit and read you
a list of features because individually a lot
of the pieces here aren't completely new. AI
that can read documents isn't new, memory isn't new,
generating video definitely isn't new. What's more
interesting to me is how all of those things
are being put together into something that's
supposed to behave like a team of people that you
can work with. You know, honestly, that's the part
I'm curious about. And is it any good? Would I
recommend it to you or would I use it myself?
That's kind of the big thing, right? So, for this
test, I'm making a short, roughly 3 minute film.
And if you want to join me, I have a link down
below to the platform. The idea is about a group
of misfits and their cars inside this underground
car culture. And visually, I want the whole thing
to have this stylized 3D concept art look. So,
rather than starting from nothing every time,
I'm going to give Agent 2 the basic world
of the film. The title, the characters,
the locations, the general subject, and the
visual direction. Right? And this is where
Agent 2 introduces something called context. Now
context is basically everything that should remain
true across the whole project. Then inside that
context, you can create individual briefs for
the specific things you're working on. So in this
case, the short itself and this world can live in
the context, while individual scenes or sequences
can have their own briefs. None of this is
super glamorous. It's project organization, which
is pretty boring. But on longer AI projects,
this is normally where things start getting
annoying. You establish a character, a location,
a visual rule, and then 20 generations
later, you're explaining all of it again to
AI in different ways. Now, Agent 2 is designed
to keep that information around instead. Now,
I'm going to give it the project information and
some material to work with, but I'm deliberately
not turning everything into one giant master
prompt first. I want to see how well it can
actually read what I give it and understand the
project from there. Now, Agent 2 can work with,
you know, scripts, PDFs, brand books, references,
links, and video, you know, the usual files. So,
the interesting bit isn't really that it accepts
different file types. The interesting bit is what
it does with that information once it becomes part
of the project and if it stays there.
So I give it something here and I just say take
a look at this and tell me what you understand
about the project so far. What feels established
and anything you think I should pay attention to
and go. Okay, so that's fairly quick and invideo
tells me they have been pushing the speed of Agent
2 and you know I can see why if I had to wait two
or three minutes every time I ask something small
I'd stop treating this like a conversation
very quickly. Now here I can ask something
get an answer react to it and keep moving. Now
whether the answer is good is obviously the more
important part right so let's see what it actually
understood about the characters the world and
the visual direction the cool stuff now the part
that I think makes Agent 2 much more interesting
is the expert agents instead of having one
giant AI conversation trying to write the film
design the shots think about cinematography and
keep track of everything else you can create
separate agents with specific jobs. So, I'm going to build a small crew
around this short. Sounds like a real short.
I'll probably want someone focused on the story
and script, like a writer, the storyboard
artist, a director of photography, DOP,
and someone keeping an eye on the characters
and the broader continuity of the project.
When you create one, Agent 2 basically
asks two things. Who am I? And who are you?
Which sounds very existential. Who am I even?
But really, it just wants a job description. So,
I can tell one, you're the writer on
this project. I'm the creative director.
Your job is to develop the story and script
for a 3 minute short using the characters,
location and world we've already established.
Then, the storyboard artist gets its own job.
The DP gets another one, and so on. Specialized
agents aren't really a new concept by themselves.
What I care about is whether these actually work
together instead of becoming four different chats
that I have to keep synchronized myself. That's I
mean that's what's the point of that. And because
right now we don't actually have a script.
That seems like the obvious place to start,
right? The project context already knows that
this is a 3 minute short about misfits and their
cars inside this underground car culture. It knows
the characters, the locations, and the stylized 3D
concept art direction. So, I'm going to go to
the writing agent and give it a fairly normal
piece of direction. Take what we've established
and develop a simple story for this 3 minute short.
Now, keep it focused enough that we can actually
tell it visually in that amount of time. That is
quick. Now, this is something I actually want to
spend all the time reading because I don't just
want it to give me a story. I just want to see
whether it understands what we're trying to make.
And of course, I don't think this is
the best story. Now, the big question,
best stories, are they still written by people
with feelings or robots who beep and boop? Well,
at this stage, AI can get you kind of
starting, but I think humans need at
least now kind of finish it. There's still that
fine touch with stories that uh people need to
uh finalize. But for this demo, this is perfectly
fine, but I can just give it that feedback the
same way I would with somebody I'm working with.
Simplify this section, keep the focus more on the
characters and the relationship with the cars
and give the ending a stronger visual payoff.
Then we iterate from there until we have something
I actually want to make. Now once we have a story,
now we finally have something the other
departments can work from. And this is where the
agent to agent stuff starts making much more sense.
Now I can go to the storyboard artist and say,
"Hey, take the script we just developed
and break the opening scene into shots.
I'm not pasting the script into this conversation.
I'm not explaining who the characters are again.
I'm not describing the visual style again. That
information already exists in the project. And
if I were to do any of that again, that's
kind of missing the point. So now it's picking
up what the writing agent just created and turning
that into something useful for its own department.
And that's the bit I want from the whole
crew idea. The writer does their part.
Then the storyboard artist can actually build
on it rather than becoming the messenger
carrying information between five different AI
chats. Now before I hand this over to the DP,
that's director of photography again. I want to
set up a playbook. Now, playbooks are
persistent rules for how one particular agent
should work. So, you're not teaching the DP
what cinematography is. You're teaching this
DP how you want your project handled. I might
add something like, "Hey, before developing any
shot, always check the established visual style,
the storyboard, and the story context. Keep the
cinematography grounded in the stylized 3D concept
artwork we established for the film. Save that.
Now, that rule should just live with this agent. I
don't have to keep pasting it into every request.
Your writer could have a completely different
playbook. Your storyboard artist could have
another one. Persistent instructions by themselves
aren't revolutionary, but having different
working rules attached to different specialists
inside the same project starts becoming much more.
And now the DP actually has something to do. Now
we got a story. We got a storyboarded scene. So I
can take one of those shots and say work out the
cinematography for this shot. Again, pretty normal
direction. The DP can see the story. It can see
what the storyboard artist decided. It knows which
character is involved, which location we're in,
and what visual style we established at the
beginning. And again, very fast. But speed aside,
this is probably the demonstration I care about
most. We started with characters, locations,
and an idea. The writer turned that into a story.
The storyboard artist turned part of that story
into shots. And now the DP is working out how one
of those shots should actually be photographed.
I didn't have to rebuild the film for every
department. That's the crew idea here. And for me,
it makes a lot more sense when you actually watch
the work move from one agent to another. Now,
I want to try the memory because this is one
of those features that sounds great in product
description, but only really matters when you stop
thinking about one prompt at a time. Now, we've
already made a bunch of decisions while developing
this film. So, I'm going to ask about something
from earlier without giving it the answer again.
What did we establish about the red Ford? And
there it is. Again, remembering one fact isn't
magic. The useful part is that a project like
this slowly builds up hundreds of little facts.
Which car belongs to which character, what a
location is supposed to feel like, what we changed
in the story, which visual direction we rejected.
Normally after enough generations you become the
database holding all of this together. Agent 2
is designed to keep that project memory around so
you don't have to reconstruct it every time you
come back. And this is probably a good point to
explain what invideo actually means by an agent.
There are obviously still AI models underneath
all of this. The model hasn't disappeared. But the
model is only one part of this system. The agent
is the layer around it. The memory, the project
structure, the specialists, how information moves
between them and how it decides what matters. Now,
invideo compares this to something like Claude
code and I think that comparison works as long as
we don't turn it into one of those um silly this
kills that things. Claude Code is built around
software development and it's amazing. Agent 2 is
built around creative production. I wouldn't ask
Agent 2 to code my accounting software and
I probably wouldn't ask Claude code to plan
the cinematography for this short but maybe it can
but probably different agents different jobs right
okay now we actually have enough to make something
we've gone from the original idea to a story
from the story to a storyboard and from the
storyboard to actual cinematography decisions.
So, I'm going to take one of those shots
and have Agent 2 start developing it.
For this, I'm using ultra. Agent 2 has light, pro,
and ultra. Light is more for drafts and tests,
planning. Pro is the general professional tier,
and ultra is meant for heavier work where there's
more context or more complicated sequences
involved. Now, invideo specifically recommends
Ultra for this kind of work. So, that's what I'm
using here. And yeah, it's pretty quick.
I'm going to sound like a broken record by the
end of this video, but it genuinely is one of the
things you notice. I don't expect every result to
suddenly be perfect. AI video is still AI video.
The harder question is whether this still feels
like the same film, same character, same car, same
world, same stylized 3D look. Making one cool
AI shot isn't really a difficult part anymore.
Keeping a 3 minute short from slowly
turning into five different films
much harder. Now, once I've got enough material
together, there's another thing I want to test.
Now, Agent 2 can actually read video or
look at video. So, I can give it a piece of
the work we've made and ask it to look at the film
rather than generate something new. Something like
go through this and flag anything that feels
inconsistent with the rest of the project or
anything you think deserves another look. And
this is a test where I genuinely hope it catches
something I could miss. Maybe a character detail
changes. Maybe something about a car drifts.
Maybe one shot just feels like it
belongs to a different visual world.
Now, I don't want an AI creative director making
every decision for me. I want something that can
watch the project with me and occasionally say,
"Hey, Seb, have another look at this. What do you
think?" The taste kind of still has to come from
you. Faster doesn't automatically mean better,
and smarter doesn't mean handing over the creative
decisions. There are a couple of other things
I haven't focused on because for me actually
making the film is more interesting. There is
multiplayer. So multiple people can work inside
the same project and talk to the same agents with
the same context. That could be useful if you got
an editor, a producer or another artist working on
different parts of the short. Now on the
enterprise side, invideo says that they have
enterprise-grade security support for teams and
external agencies working inside the same projects
and that customer data isn't used for training.
They also run a research desk and have research
grants for ambitious research problems with
grants up to 200k USD. And you know that's cool,
but I'm not going to pretend that's the reason
most of you are going to try this. For me,
the interesting question is still whether it makes
actually building the film easier. That and I bet
making AI social media slop will get even
easier as well. So after working through this,
I don't think the story is invideo invented
AI video again because they didn't.
And trying to sell it that way would
actually undersell the useful part.
We already have very good image and video models
and there's another one every five minutes. What
Agent 2 is trying to solve is the mess around those
models once you start making something bigger.
In this case, I didn't start with a finished
script or a storyboard, which I mean you could.
I started with some characters, locations,
a world, an idea for a three minute film.
From there, we developed the story, handed that to
another agent for the storyboard, handed that work
to the DP, established rules for how the different
agents should work, and then started actually
making the film. That combination is what I find
interesting. Now, not every part of it is new.
Not every suggestion is good. Not every generation
is going to work. And there are definitely things
I'd still like to see improve. And the speed
matters more than I expected because the whole
thing only really works if talking to these
agents feels like an ongoing conversation rather
than submitting another render job. So if you're
already working on larger AI video projects and
you've reached a point where keeping track of the
project is starting to take almost as much time as
making it, have a look at invideo Agent 2. Link
is down below and I'll see you in the next one.
Video Length: 16:57
Uploaded By: Sebastian Kamph
Published: 9/10/2026
View Count: 5,118