# AI agents for design: letting an assistant make finished work

[Canonical HTML page](https://gradiently.design/guide/ai-agent-design-workflow)

A chatbot answers. An agent takes a job, works through the steps with real tools and hands back something finished. Design is a good place to start, if you set it up well.

## The short version

- AI agents for design are assistants that plan a design job, call real design tools to do it and check the result before handing it back.
- An agent needs three things to make finished work: tools it can use, a brand it can read and a clear brief with a definition of done.
- Agents are best at repeatable production such as resizing, weekly series and variations, and weakest at choosing a brand's look from nothing.
- A person should approve the first design in any series and every design before it is published.
- Gradiently connects to ChatGPT, Claude and other assistants through MCP, so an agent can create designs on your brand's Mark that open in the Studio.

**AI agents for design** are assistants that do more than reply. Given a job such as "make this week's four posts and their stories", an agent plans the steps, uses real design tools to carry them out, checks what it made and hands back finished, editable designs. The tools are what make this possible: connected to a design app through MCP, an assistant can read your brand, lay out a design and save it. Without tools, it can only describe what you should make.

## Assistant, agent and automation

The words get blurred, so here is how this page uses them. The difference is how much of the job the software carries without you.

| Aspect | Chat assistant | Agent | Automation |
| --- | --- | --- | --- |
| Input | A question | A goal and a brief | A trigger and fixed rules |
| Steps | One reply | Plans and runs several | The same steps every time |
| Tools | Optional | Essential | Hard coded |
| Judgement | Yours | Shared, within limits you set | None |
| Design example | Suggests a headline | Makes four posts and their stories | Posts a saved image at 9am |

Agents sit between a chat and a script: flexible like the first, persistent like the second.

## What a design agent needs to make finished work

- **Tools.** An MCP connection to a design app, so it can search assets, create designs and make other sizes. [MCP servers for designers](https://gradiently.design/guide/mcp-servers-for-designers) explains the plumbing.
- **A brand it can read.** Fonts, colours, logo, voice and audience stored where the tool can reach them, not retyped into every prompt.
- **A brief with a definition of done.** Formats, words, deadline facts and what counts as finished.
- **Limits.** Permissions that let it design but not publish, spend or claim anything you have not agreed.

The second item is the one people skip. An agent with no brand to read invents one each time, which is how a month of AI posts ends up looking like four different companies. [Brand guidelines for AI](https://gradiently.design/guide/brand-guidelines-for-ai) shows what to write down.

## The loop an agent runs

1. **Read the brief and the brand** It pulls your brand profile and any source material, such as an event page or a product list.
2. **Plan the set** Four posts at 1080×1350, four stories at 1080×1920, one headline and one detail line each.
3. **Make the first piece** It creates one design with the design tool, using your fonts, colours and background.
4. **Check against the brief** Are the words right, the date correct, the format exact? A good agent reads its own output before moving on.
5. **Repeat and resize** It makes the rest of the series in the same layout, then makes other sizes from each.
6. **Report back** It lists what it made, with links, and flags anything it was unsure about.

## A worked example: a weekly series

Say you run a bakery and post a "bake of the week" every Monday. You write a brief once and reuse it, changing only the facts.

```text
Brand: Crumb & Co (use our brand profile and Mark)
Job: this week's Bake of the week set
Facts: Cardamom knots, Monday to Wednesday, from 8am, while they last
Make: 1 Instagram post 1080x1350, 1 story 1080x1920
Words: headline under 6 words, one detail line, no hashtags in the design
Done means: both designs saved, facts checked, nothing published
Ask me before: changing the layout from last week's set
```

A reusable agent brief. The last two lines are what keep the agent safe and the series consistent.

A portrait bakery post announcing cardamom knots as bake of the week, on a warm gradient

The post the agent saves, on the bakery's own Mark.

The same bakery announcement recomposed as a vertical story on the same warm gradient

The story it makes from the post, laid out for the tall frame.

In Gradiently, the agent works through its MCP server: it reads the brand, creates the designs on the brand's **Mark**, the living gradient and material background every design wears, and saves them where you can open them in the Studio. The Mark keeps words readable by itself, moving its calmest region behind the text, so the agent never needs to add boxes or shadows. When the assistant lays the design out itself, it spends no Gradiently credits. [Connect an AI assistant](https://gradiently.design/guide/connect-ai-assistant) covers setup.

## What to hand an agent, and what to keep

### Hand it

- Resizing a finished design for every platform
- Weekly series with a fixed layout
- Variations of a headline to compare
- Filling a template from a list of facts
- First drafts of a campaign set

### Keep it

- Choosing your brand's look in the first place
- Approving the first design in any series
- Anything with prices, legal or health claims
- Publishing
- Claiming or spending on anything

The pattern is simple: agents are good at volume inside rules you have already approved, and poor at setting the rules. [Will AI replace designers](https://gradiently.design/guide/ai-and-designers) makes the same point from the designer's side.

## Common ways agent design goes wrong

- **Drift.** Each design is slightly different from the last. Fix it by pointing the agent at last week's design as the reference.
- **Confident wrong facts.** A date or price changed somewhere in the chain. Put facts in the brief exactly, and check them yourself.
- **Overcrowding.** Agents like to use every fact you give them. Ask for one headline and one detail line.
- **Too much access.** A key that can publish or claim will one day do it by mistake. Leave those permissions off.

> **Start with one design** Ask for a single piece first, approve it, then say "make the rest of the series like this one". Everything after inherits a look you have already signed off.

For marketing jobs beyond design, see [AI agents for marketing](https://gradiently.design/guide/ai-agents-for-marketing), and for writing briefs an assistant understands, [design by describing](https://gradiently.design/guide/voice-to-design).

## FAQ

### What is an AI design agent?

An assistant that takes a design job, plans the steps, uses real design tools to carry them out and checks the result. It hands back editable designs rather than advice.

### Can ChatGPT or Claude act as a design agent?

Yes, when they are connected to a design tool through MCP. Gradiently works inside ChatGPT, Claude and other assistants that support remote MCP servers.

### Should an AI agent publish designs automatically?

It is safer not to. Let the agent make and save designs, and keep approval and publishing with a person.

### How do I keep AI made designs on brand?

Store your brand where the tool can read it, reuse one brief, and approve the first design in each series before the agent makes the rest.