Design Agents Need Creative Direction, Not Just Prompts
2026-05-30 · 4 min read · Janaina Maia
The most useful part of a design agent is not that it can make something quickly. It is whether it can help a human designer make better decisions.
The Verge tested Adobe’s Firefly AI Assistant, a conversational agent that can operate creative tools, explain its steps, and make iterative edits through a chat-style interface. The review was mixed: the assistant was more engaging than a basic image generator, but the visual results still felt like work from a mediocre design intern. I think that is exactly the right comparison, and it exposes the design challenge ahead.
A junior designer can help, but only if the direction is clear. They need a brief, constraints, references, feedback, standards, and someone with judgement deciding what good looks like. An AI design agent is no different.
Chat is not creative direction.
Many AI tools still behave as if the user’s job is to type a wish and wait for output. That is a shallow model of creative work. Good design is not only production. It is framing the problem, deciding what matters, making trade-offs, recognising patterns, removing noise, and knowing when something is almost right but still wrong.
When a design agent turns every interaction into a prompt, it quietly transfers the hard part to the user. The user has to describe taste, context, brand, risk, quality, audience, and intent in a tiny text box. That may work for play. It is not enough for serious product work.
The agent should expose its judgement.
What I find promising in Adobe’s approach is that the assistant explains what it is doing. That matters. A creative agent should not feel like a slot machine that returns images. It should show the moves it made, the assumptions it used, and the options it considered.
But explanation is only the starting point. The better product pattern is critique. If the agent changes composition, colour, contrast, hierarchy, or tone, it should help the user understand why. If it produces three directions, it should explain the trade-off between them. If something is weak, it should say what is weak in plain language.
Enterprise AI needs the same lesson.
This is not only a creative-tools problem. Enterprise AI products are heading in the same direction. Agents will draft reports, prepare workflows, make recommendations, generate diagrams, analyse documents, and create artefacts that people may treat as finished work.
If those agents only produce output, they will create review burden. If they can expose assumptions, compare options, flag uncertainty, and invite targeted human judgement, they become much more useful.
The product question is not, “Can the agent make the thing?” The better question is, “Can the agent help the human understand whether this thing is good enough to use?”
Design implications.
- Give agents briefs, not vibes: make goals, audience, constraints, and quality criteria explicit.
- Design critique surfaces: show why an option works, where it is weak, and what trade-off it makes.
- Preserve human taste: let users build reusable preferences instead of rewriting the same prompt every time.
- Separate exploration from approval: make it clear when the agent is sketching, recommending, or preparing something for final use.
- Make iteration visible: show what changed between versions so review is faster and less mysterious.
My take.
Calling these tools agents raises the bar. If the product behaves like a chat box attached to a generator, it is not a collaborator. It is a faster production assistant with unclear taste.
The next step for AI design tools is not more magic. It is better direction, better critique, and better ways for humans to keep judgement in the loop.
A good design agent should not replace creative direction. It should make creative direction easier to express, test, and refine.