12/32proof of conceptMulti-Agent Creative Production
Design StudioAn AI Creative Team With Its Own Critic
Six specialist agents take a brief to three art directions, and a separate critic judges the result.
What it solves
One prompt to an image model gives one generic picture and nobody judges it. This works like an agency team: specialists with their own skills propose three real alternatives, and a critic who did not make the work reviews it before anyone can approve it.
How it runs
Seven steps, input to output
Goes inA brief, the brand rules, any reference images and the facts the work may use
- 01
Brief
The brief, brand rules and supplied facts are read, and any claim without a source is flagged before copy is written.
- 02
Three directions
The creative director, visual designer and photo editor agents develop exactly three, each a layered spec. A set that repeats an idea is rejected.
- 03
Imagery
The studio producer, the only agent allowed to spend, sends image requests to FLUX on fal.ai against a live-priced budget cap.
- 04
Local GPU
ComfyUI runs BiRefNet to cut out backgrounds and a 2x upscaler, on the laptop, at no cost per image.
- 05
Render and check
Type, logo and images stay separate layers, rendered to PNG and checked for margins, contrast, fonts and disclaimers.
- 06
Critic
A separate agent judges the rendered file at full and phone size on six criteria and names the layers that fail.
- 07
Revise
At most two targeted revisions of the failing layers, then a person records approval.
Comes outThree art directions as layered specs, rendered and reviewed, with an optional motion cut
What it does
A local creative studio run by six specialist agents: creative director, visual designer, photo editor, motion designer, studio producer and an independent critic, each with its own skills.
A brief must become exactly three distinct art directions, each a layered design spec. fal.ai supplies FLUX imagery under a live-priced spend cap, and ComfyUI on a laptop GPU cuts out backgrounds and upscales.
Text and logos are never generated.
Automated checks run first, then the critic judges the rendered file at full and phone size, and only failing layers are revised.
Three decisions
What shaped the build
The maker is not the judge
The critic sees the brief and the rendered image, not the makers’ reasoning, and approval stays locked until the current version has passed both the checks and the critique.
Paid calls are capped and never retried blindly
Each request reserves its estimated cost against a cumulative budget, priced live from fal.ai. A failed or interrupted call is checked by its request ID before anything is sent again.
Text is never generated
Headlines and logos are renderer layers, every image prompt asks for no text, and a spec marked factual cannot use generated imagery at all.
Built with
8 parts
- Six agents · 10 skills
- director to critic, each with a brief
- fal.ai · FLUX
- imagery, spend-capped, never retried blind
- ComfyUI · BiRefNet
- local cutouts + 2x upscale on the GPU
- Claude Code · Codex
- the agents run in the session, no API
- Zod specs
- every direction a layered spec
- Playwright · Sharp
- layers rendered to PNG
- Remotion
- optional motion cut
- Express · Node 24
- local API, 127.0.0.1 only