Designing an AI agent that understands how creative teams work.

Overview

From an image generator to an agent that remembers the brief.

As co-founder and founding product designer at CotonAI, I owned product strategy as well as design. The question I kept bringing back to the team wasn't “how do we generate better images?” but “why does every prompt start from zero?”

This is how that question moved CotonAI from a visual workspace toward an agent that carries creative intent across a whole collection.

• Problem: Every prompt started over.

  The AI forgot why a colour was rejected or which silhouette was approved.

• Solution: The canvas is the context.

  An agent that plans, executes and returns work to the canvas, with designers approving each step.

• Outcome: 71% first-pass

  Designs accepted without a revision in the Steele pilot.

Problem

We didn't need another AI image generator.

Fashion teams already had plenty of tools to generate images. Yet designers were still spending hours moving between moodboards, spreadsheets, image generators, review threads and production files.

The real problem wasn't creating an image. It was everything that happened before and after it. A designer could generate 20 variations in minutes, but the AI wouldn't remember why a color was rejected, which silhouette was approved, or how the collection needed to evolve. Every new prompt felt like starting over.

• No memory of decisions

  Rejected colors and approved silhouettes were lost between prompts.

• Context scattered across tools

  References, material constraints, brand identity and feedback lived in different places.

Direction

What if AI understood the whole creative process, not just the next prompt?

Working with fashion brands and production teams, we saw how much creative work depended on invisible context: reference images, material constraints, brand identity, previous decisions and feedback.

We started CotonAI as a collaborative visual workspace where designers could sketch, generate, edit, explore colorways and review work together on a shared canvas. The bigger opportunity emerged from those workflows: the canvas shouldn't just hold the work. It should become the context through which an agent understands the work.

Solution

Three product decisions for the agent experience.

The proposed workflow became a loop the agent runs with the designer. The design challenge was making complex execution feel understandable, without exposing unnecessary technical complexity.

How it works

• 01

  Understand context

• 02

  Plan

• 03

  Execute

• 04

  Review

• 05

  Refine

• 06

  Remember

Decision 1: Context before conversation

• Before

  Designers had to explain everything through prompts.

• Change

  An image, sketch, annotation, garment reference or selected canvas region can communicate intent.

• Result

  From “describe exactly what you want” to “show the agent what you're working on.”

Decision 2: Execution, not just generation

• Before

  A creative task rarely ends with one image, but the tool did.

• Change

  A harness for managing context, planning steps, selecting tools, generating alternatives and returning outputs to the canvas.

• Result

  Multi-step work that stays understandable.

Decision 3: Humans remain creative directors

• Risk

  Autonomy is useful until an agent makes the wrong creative decision.

• Change

  Visible progress, contextual feedback, version comparison and approval checkpoints.

• Result

  The agent handles repetitive execution. The designer keeps creative judgment.

Impact

Early pilot outcomes validated the collaborative workflow.

These are early workspace and pilot outcomes, not evidence that the proposed autonomous agent architecture has already achieved the same results.

• AI-assisted designs: 140+

  Steele pilot

• Review comments: 188

  Left by teams on the shared canvas

• First-pass acceptance: 71%

  Designs accepted without a revision, Steele pilot

• Revisions per design: 3–5 → 1–2 (Typical)

  Before 3–5 vs With CotonAI 1–2. Typical revisions in the Steele pilot. Bars sit at the midpoint of each range.

• Visual assets per month: ~5,000

  Workflows supported through CreateNet

Learning

Make intelligence understandable, controllable and useful inside real work.

The hardest part of designing AI products isn't making intelligence more powerful. It's making it understandable, controllable and useful inside real work.

The live workspace established the foundation. Persistent creative memory, multi-step orchestration and richer agent controls are the next product direction: moving from AI that generates creative outputs toward AI that helps teams carry creative intent across an entire workflow.

“That's the design problem I'm most interested in solving.”