The Group Chat Might Be the Future of Work

Meta launched Muse this week, its new personal AI agent and a big piece of Zuckerberg's push toward "personal superintelligence."

I wanted to see what it could actually do, so I gave it a real task on my business website.

I Gave It the Keys

From my couch, I told Muse to sign into Wix, open the editor, make changes to the homepage, remove a section and publish them. When one of the changes went wrong, I had it go into the site history, find an earlier version, restore it and republish.

It worked, although not cleanly.

There were plenty of rough edges. Signing into Google was strangely painful. Verification codes failed. Sessions died in the middle of tasks. Editing inside Wix was awkward, especially when Muse needed to manipulate something precisely inside the editor.

A few things that would have taken me five minutes on a laptop stretched into much longer exercises.

The Interesting Part Was How Little I Had to Learn

I wasn't designing a workflow or building an automation. I wasn't thinking about APIs or prompt structure.

I was just telling it what I wanted done in normal language.

"Remove this section."

"Publish that."

"That broke something. Go back to the previous version."

Then I watched it use a computer and try to figure it out.

That feels materially different from the AI most people have used over the past few years.

I tried a similar experience with ChatGPT months ago when its agent could operate a virtual computer inside the chat. The progress since then is noticeable. Muse still gets stuck, especially inside systems like Wix, but you can see where this is headed.

We May Need to Build Differently for Agents

The experience also changed how I think about some of the things I'm building.

For example, I probably don't want my business website living inside a highly constrained website builder forever. A custom-coded site gives an agent much more room to work.

If agents are going to increasingly maintain, modify and operate software for us, the systems we build may need to become easier for agents to work inside too.

That's probably another project.

The Group Chat Is the Part That Matters

This week it was just me in a chat telling an AI agent what to do.

Now imagine a small company where several people are in the same group chat with the same agent.

Marketing asks for a landing page. Someone asks it to pull customer feedback. Another person asks it to update the CRM. Someone else wants the numbers checked before tomorrow's meeting.

The agent handles the work in the middle of all of it.

I think that could become a very normal way companies operate.

Consumer AI Is About to Get More Interesting

For the past few years, most of the serious AI conversation has centered on enterprise software. Consumer AI has mostly meant opening a chatbot and asking it questions.

Agents push that much further.

Meta is especially interesting because it already owns the places where billions of people communicate. If an agent eventually lives inside WhatsApp, Messenger, Instagram or a pair of glasses, people may never think much about the technology behind it.

They'll just message it like they message everyone else.

There's also something slightly ridiculous about the fact that this post is part of the experiment.

I used an AI agent to help work through the experience, shape the writing, open my website, create the blog post, format it and publish it.

So the article you're reading about people managing work through conversations with AI was itself produced and published that way.

That's why I think people are underestimating what the interface for AI work may actually look like.

It may just be a group chat with an AI agent sitting in it, doing the work.