What Muse Is: An Agent With Its Own Computer

You have probably used a chat assistant: you ask a question, it writes an answer, and you go and do the work yourself. An agent is different. You give it a task, and it carries out the steps on your behalf: it opens files, reads email, builds documents and comes back to you when it needs a decision.

Compare two requests about the same job, a product shoot of an olive-oil bottle delivered in nine sizes:

What sizes does a web-shop hero image usually need?

Read the brief in my project folder, list every deliverable with size, format and colour space, and draft questions for the client about anything missing. Don't send anything.

The first request gets you an answer. The second sets Muse to work in your accounts. That is more useful, and it carries more risk, because the agent can now see your files and could act on them. That is why this course spends three lessons on setup before you give Muse any real work.

Muse is Meta's personal AI agent. It launched in the US on 8 September 2026 at muse.ai and was built by Meta Superintelligence Labs. When you sign in at muse.ai, you are passed through to a Meta sign-in.

Three things get mixed up with it:

• Muse Spark is the AI model Muse runs on. It is the engine, not the product you use.
• Meta AI is a separate Meta app with its own terms. Settings and promises from Meta AI don't apply to Muse.
• Other products called Muse, such as the Project MUSE journal database or a website builder, have nothing to do with Meta.

This matters when you look for help. If a search result, a tutorial or a set of terms refers to Meta AI or to another Muse, it is about a different product. In this course, trust Meta's own pages: the Help Center, muse.ai/terms and muse.ai/privacy.

Each Muse user gets a dedicated cloud computer called the Muse Secure VM (VM stands for virtual machine, a computer that runs on Meta's servers). It has a web browser, a file system and a terminal. Muse can write code and build small tools on it, and it keeps working there after you close the app.

What it produces on that computer are Artifacts: documents, PDFs, web pages and interactive outputs such as trackers and checklists. Meta's data help page also lists images, videos and podcasts.

In studio terms, picture a production assistant with a desk in another building. On the olive-oil job you could ask for a deliverables checklist Artifact listing all nine sizes, from the web-shop hero to the printed store card, with format, colour space and file name for each.

One warning now, and more detail later: an Artifact can be shared as a web page that anyone with the link can open. Never share an Artifact that contains client work.

Muse has many features. This course uses seven of them. Names in code style come from Meta's help pages. The app changes often, so your screen may look different.

• Chats: a tab that holds one long main chat for your ongoing work.
• Side chats: separate conversations, also in Chats. You'll keep one per client so their context doesn't mix with other jobs.
• Artifacts: the documents, trackers and checklists Muse builds.
• Goals: a tab for long-running aims and the plans Muse makes for them, such as keeping a client's asset log up to date.
• Scheduled tasks: jobs that run at set times, such as a Friday 4pm check of the export folder. You review them under the Assistant icon > Upcoming.
• Connectors: links to services such as Gmail or Dropbox, managed in Settings > Connectors.
• Memory: what Muse keeps about you, in a file called Memory.md under the Assistant icon > Identity > Memory. You can read, edit and delete it.

Flip the cards to test yourself.

An agent doesn't stop when it replies. A typical task goes like this. You give a task or goal. Muse plans the steps. It works in the background on its VM, reaching your services through Connectors. Meta's own lists name Gmail, Google Calendar, Dropbox and Box, among others. When something changes or it needs permission, it comes back with an approval card. Every action is recorded in the Activity log.

The Mac app can go further. With macOS permissions it can use local apps and files on your computer. That is powerful, and it reaches broadly, so a later lesson limits it to one project folder.

What Muse can't reach matters just as much. As of 28 October 2026, no Meta connector list includes Adobe: there's no Photoshop, Lightroom, Premiere Pro, Illustrator or Express connector. Muse can work on the paperwork around your files (briefs, emails, folders, specs) but not on your pixels. Don't expect it to mask, retouch or export for you.

Put the stages in order.

In this course Muse has one job: production coordinator. It logs, checks, drafts and flags. You make every image decision and every release.

In practice, Muse gives each incoming request an asset ID, notes its source and permission status, and turns a brief into a checklist. It compares an export folder against the spec and drafts replies to clients for you to approve. You judge the contact shadow and the reflection at 100 to 200 percent zoom, decide whether a missing permission blocks the job, and decide what leaves the studio under your name.

That split matches what Ely Greenfield, CTO of Adobe's creative business, has said about Adobe's professional customers: they want the repetitive production work automated without losing control. That's Adobe's view as a vendor, and it fits how a studio already works. The assistant keeps the job jacket. The art director signs it.

Match each task to who owns it.

Three ground rules for the rest of the course.

This course is independent. It isn't made or endorsed by Meta or Adobe, and none of the people quoted in it endorse it. Several speakers worked for Adobe when they were recorded, so read their points as the vendor's view.

What Meta says is a vendor claim. Meta itself says Muse "will sometimes make mistakes". It also calls prompt injection an unsolved problem across the industry. Prompt injection means hidden instructions in content the agent reads, such as a line in a client email or a web page, that try to steer it to do something you didn't ask. Under Muse's terms you are responsible for what it does on your behalf.

Your screen may differ. Muse is a young product that changes often. Menu names in these lessons come from Meta's help pages as of 28 October 2026. If something doesn't match, go by the official Help Center.