Agents
Agents
Build, configure, and share AI agents that plan, build, and run your creative projects.
Agents are AI partners that help you plan, build, and run creative projects from start to finish, and the most powerful ones are the agents you design yourself. A custom agent has a name you choose, instructions you write, the tools you give it, and a knowledge base you attach.
It remembers your project preferences and the decisions your team has made, and it delivers its work as an editable Space, not a finished file you can't touch. This guide covers what agents are, the ones you start with, and how to create, configure, and share your own.
In this article
What is an agent?
An agent is an AI assistant with a job. Instead of writing a new prompt every time, you tell the agent once who it is and what it should do, and it carries that context across the whole project. A custom agent is made of four parts:
Part | What it does |
|---|---|
Name | What you call the agent, such as Brand Photographer or Client X Lead. |
Instructions | The prompt that defines how the agent behaves, its style, and its rules. |
Tools | The capabilities you grant it: image, video, audio, web search, and Spaces. |
Knowledge base | Reference files the agent reads to stay on-brand and on-context. |
Once it's built, you can share the agent with your teammates so everyone works the same way, without re-explaining the brand or the project each time.
The agents you start with
Every account comes with two ready-to-use agents. You can use them as-is, or as a starting point for your own.
Agent | What it's for |
|---|---|
Magnific Agent | Your default creative partner. It plans multi-step projects, builds and runs multi-asset workflows in Spaces, generates project status reports, searches your past chats across projects, keeps persistent project guidelines, and aligns outputs with the brand references you upload. |
Ad Creator | Shapes campaign concepts, hook variants, and production assets. It structures strategy through creative briefs, plans video ads shot by shot, and generates campaign-ready imagery and video. |
What makes custom agents different
Three things set custom agents apart from a standard AI assistant.
You design it. You define the name, the instructions, the tools, and the knowledge base. The agent works the way you work, not a generic default.
It remembers your work. Agents keep two layers of memory: your individual preferences and the shared context of the whole project. You never have to re-brief it from scratch.
It outputs editable workflows, not just images. The result lands in a Space your team can open, adjust, and iterate. You refine, you don't restart.
Navigating the agent
Everything you need to run an agent lives in one place: the conversation in the middle, a menu beside it, and your controls in a single row below.
The menu
Here you find your private and shared agents, along with your skills, separated into the ones Magnific provides and the ones you create. Your conversations live here too, so you can move between them without leaving what you're doing. There's also a link to Magnific Academy.
The agent opens in two ways. On its own page, the menu stays anchored to the left, and a creation you open appears beside the chat, so you can keep asking for changes while you look at it. When the agent opens alongside another tool, the hamburger button brings up the menu, and a creation fills the screen. On narrow screens and on mobile, the full page behaves the same way.
The clock button next to the input opens the record of that conversation, with two tabs. Activity is a step-by-step account of what the agent did, and each step takes you to the message where it happened. Creations gathers everything the conversation generated, with previews.
Agent settings open in a modal with three tabs: General, Agents, and Skills.
The composer
All the controls sit in a single row: attach a file, mention an agent, dictate, open settings, and send.
Control | What it does |
|---|---|
Model selector | Sets the chat model for this conversation. The chip shows which model is running and its credit consumption level. Your choice is remembered in that conversation, including when you come back from another device. |
Quick settings | Agent mode, autonomy level, and credit autonomy. |
Prompt history | The up and down arrows move through prompts you've already sent, the way a terminal works. They're saved to your account and stay there between sessions. |
Usage ring | Shows how much of your daily quota is left and when it renews, and updates live while the agent works. |
Following what it's doing
A single status line next to the agent's avatar tells you what's happening in real time. It names the tool it's running, says when something failed, and tells you when it's waiting on you.
Responses stream in as the agent writes them. The detail of the agent's thinking and its tool calls lives in the Activity tab, which keeps the chat itself clean.
Create a custom agent
Open the Agents panel
Select Create agent to start a new one.
Name your agent
Give it a name that makes its role obvious to your team.
Write the instructions
Define its personality, rules, and approach to work. The more specific the instructions, the more consistent the output.
Choose its tools
Grant only the capabilities its job needs: image, video, audio, web search, or Spaces.
Attach a knowledge base
Add brand, style, or project files for the agent to draw from.
Add sub-agents
Optional. Let the agent call specialized agents for specific tasks.
Share it
Make the agent available to your project teammates, or keep it private.
You can update an agent's name, instructions, tools, and knowledge base at any time. Changes apply the next time it runs.
Tip
Start small. Give the agent a clear, narrow job and a tight knowledge base, then expand it as you see how it performs. A focused agent is more reliable than one that tries to do everything.
Tools reference
Tools are the capabilities you grant an agent. Give it only what its job needs.
Tool | What the agent can do |
|---|---|
Image | Generate and edit images. |
Video | Generate and edit video. |
Audio | Generate voiceovers, sound effects, and music. |
Web search | Look up current information on the web while it works. |
Spaces | Run multi-step workflows on the canvas and deliver editable results. |
You can switch tools on or off at any time, even mid-conversation, without losing the context of what you're working on.
Credit autonomy, in the quick settings inside the composer, sets whether the agent asks you to confirm the cost before it generates. If you have an Unlimited plan, you can toggle it on or off.
Agent modes
The agent has two modes:
Create: generate and edit images and videos.
Brainstorm: conversation only. The agent sets aside its generative tools and focuses purely on developing ideas through chat.
You switch between them in the quick settings inside the composer, alongside two more controls.
Autonomy sets how much the agent checks with you before it moves on.
Level | What it does |
|---|---|
Low | Asks about everything before continuing. |
Medium | Asks about what matters and suggests the rest. This is the default. |
High | Leads the process and asks only when it's essential. |
Your autonomy level is stored in your browser and travels with each message, so it applies to what you send next.
Credit autonomy is a separate control. It sets whether the agent generates without asking you to confirm the cost first.
Knowledge base
The knowledge base is the set of reference files an agent reads before it works. It's how the agent stays on-brand without you pasting the same guidelines into every prompt. You can attach:
PDFs: brand books, style guides, briefs.
Text and documents: tone-of-voice notes, do's and don'ts, product details.
Images: logos, brand references, mood boards.
Video: reference clips or examples.
Keep the knowledge base focused on what the agent actually needs for its role. A tight, relevant knowledge base produces more consistent results than a large, mixed one.
Skills
Skills
A skill is a saved, reusable instruction: a document with a name, a description, and step-by-step instructions the agent follows when the task matches. You write it once and the agent repeats it the same way every time.
There are two kinds:
Kind | Who sees it | What you can do with it |
|---|---|---|
Platform skills | Everyone. Magnific provides them and they're the same for every account. | Read-only. Duplicate one to get your own copy and adapt it. |
Personal skills | Only you. | Create, edit, and delete. |
Create a skill
Go to Settings > Skills and select Add.
Name your skill. Give it a name that says what it does.
Describe what it does. A short summary of the job the skill handles.
Write the instructions. The steps the agent follows, including when to use the skill and when to skip it.
You can also have the agent write it with you. Settings closes, the chat opens, and you build the skill together.
Each account has a limit on how many skills it can hold.
Use a skill
Skills work two ways:
On their own. The agent recognizes that a skill fits what you asked and applies it. You don't have to do anything.
By hand. Type
/in the chat and pick one from the list. It stays in your message as a tag, and you can chain several in a single request.
How agent memory works
Agents remember context so you don't have to repeat yourself. There are two layers:
Memory layer | What it stores | Who it applies to |
|---|---|---|
User memory | Your individual preferences and the decisions you've made. | Just you. |
Project memory | The shared context of the project, including every interaction the team has had with the agent. | The whole team. |
Because project memory is shared, anyone on the team can ask the agent for a project summary at any time: what's been created, how teammates have interacted with it, and the current status of the work.
Sub-agents
Agents can call other agents to handle specialized tasks, like a director calling in specialists. The agent you're talking to acts as the lead and delegates parts of the job to its sub-agents.
You can activate, deactivate, or switch sub-agents in the middle of a conversation without losing the context you've built up. This lets one agent coordinate a whole workflow while specialized agents handle the parts they're best at.
Share an agent with your team
When you share an agent with your project teammates, everyone gets the same configured partner: the same instructions, tools, and knowledge base.
This means the person who built the agent doesn't become a bottleneck. A teammate who doesn't know the brand inside-out can still produce on-brand work, because the expertise lives inside the agent. The brand only has to be defined once.
What an agent produces
An agent doesn't hand you a flat, finished file. It produces an editable Space, the same canvas you'd build a workflow in yourself.
That means you and your team can open the result, adjust any step, and run it again. If something's off, you refine that part instead of starting over. The output is a workflow you own and can iterate, not a final render you're stuck with.
Inside a Space, the agent works with its own cursor. You watch it move across the canvas, select the nodes it's working on, and talk to you from a bubble next to the cursor, where its reply appears as it's written. When it's done, it points you to the canvas, where the work is.
From any finished creation, Recreate in chat opens the agent with that creation already attached and the right instruction in place, so you can ask for variations without uploading anything again.
Who custom agents are for
Custom agents are most useful for teams that produce a lot of on-brand creative work and want to scale it without re-briefing every time.
Marketing teams, in-house or agencies, building campaigns, social, e-commerce, and lifestyle content who want brand consistency at scale and a way to share one expert setup across the whole team.
Studios in cinema and advertising running client work who want one agent per client, continuity across projects, and outputs the team can shape before review.
If you re-explain the same brand or rebuild the same setup on every project, a custom agent removes that work.
FAQ
Can I edit an agent after I create it?
Yes. You can update its name, instructions, tools, and knowledge base at any time. Changes apply the next time the agent runs.
Can I switch agents in the middle of a conversation?
Yes. You can activate, deactivate, and switch agents and sub-agents mid-chat without losing the context you've built up.
Does the agent remember work across different projects?
Project memory is tied to its project. User memory, your individual preferences and decisions, follows you.
Do my teammates need to know the brand to use a shared agent?
No. That's the point of sharing. The brand and project context live inside the agent, so anyone on the team can produce consistent work.
What's the difference between the starter agents and a custom one?
The starter agents, Magnific Agent and Ad Creator, are ready to use out of the box. A custom agent is one you design, with your own name, instructions, tools, and knowledge base, and can share with your team.
What's the difference between a knowledge base and a skill?
A knowledge base is reference material the agent reads. A skill is a set of instructions the agent follows. Use the knowledge base for what the agent should know, and a skill for how it should work.