Meta's Muse, one week in: the model isn't the best, and the integrations make up for it

"Afraid to announce that Muse from Meta is currently the best agent implementation to date. Stupidly easy, intuitive, powerful, and free." That's Paul Couvert on Threads, and judging by the replies, a lot of people agree.

Last week I recorded my first reaction to the launch and asked what Meta's Muse will mean for brands. Muse has now been live for a week, so I went through what real users are posting about it. Here's what I found:

The reactions are surprisingly positive

I expected more scepticism for a Meta product that wants access to your inbox and your accounts. Instead, most of what I read was genuinely enthusiastic. Meta's own Muse account describes Muse as always on, fast, connected to your apps and designed to be secure, and free for most of what people need, with subscription plans for people who want to do more.

What people actually use it for

The use cases were the part I was most interested in. One user, who has given their agent its own name, Mambo, said Muse is pretty modest when you ask it what it can do, and then listed what it had already done: booked a flight, signed up for a marathon, and built a financial report for a restaurant from the reports each location sends in.

Others have generated thousands of images and videos. Someone learning a new language uses it for real-life practice: tell it your level, A1 or B2, and it corrects you gently or harshly, with a daily quiz if you want one. One person's first-week list read like a to-do list most of us never get through: inbox to zero, unsubscribed from dozens of spam senders, researched and bought a scooter for the ride to the train station, removed their details from third-party tracking sites, and contacted insurers and repair people after a storm damaged their house.

Consumers first, professionals second

What stands out is what's missing. I didn't see many people using Muse to write code or build workflows. It's mostly normal, everyday stuff that matters to the person doing it. That's where Muse is strongest, and it shows how well Meta understands what people actually need.

Some posts are bolder. One user wrote that having a Muse agent feels like the employee they knew they needed but never pushed themselves to hire. It's a big claim, and a bit general. I'm still looking for the concrete ideas behind statements like that. Another called it a Claude Code level product. Meta has been big in AI for a long time, so maybe that shouldn't surprise anyone.

The model isn't the best. The integrations make up for it

This is the line from the week that I keep coming back to. One user wrote that the model probably isn't quite as good as the best ones from OpenAI or Anthropic, but the integrations more than make up for it. Plus it's free.

That's the whole story in two sentences. A slightly weaker model that can actually reach your inbox, your bookings and your messages beats a smarter one that can only talk. What "free" means here we'll find out over the next few years. But for now, access is winning over raw intelligence.

My verdict after a week: it's very early days. Nobody knows exactly what it's capable of yet, and people are still poking at it to find out. It's only available in the US for now, so from mainland Europe I'm watching closely.

None of this is new. The packaging is

Here's the thing. Almost everything people are describing, I've been doing with Claude for a long time, using my own skills and routines. Anthropic just folded Cowork into the regular chat, so even the distinction between talking and doing is disappearing there. What Muse does differently is package it so that someone who'd never set up a skill gets the same result.

I think we'll see a lot more of these one-stop shops: you add your connections, give it the right context, and it works with you on everything you're working on. Which raises the question that actually matters.

Your context has to be tool-agnostic

At some point you have to decide where your information lives and how your data gets used. And that decision should not depend on which assistant is winning this month. Your context has to be tool-agnostic.

Every one of these agents needs the right context to help you well. Access to the right data for whatever you're working on, to what happened in the past, to what you've tried and what you haven't. I learned that the hard way when a scheduled routine of mine failed three weeks running because it didn't know something about its own environment. And it's why I think of a context window as working memory, not a hard drive: what the agent knows has to come from somewhere you control.

So start organising that data, and document the journey, not just the output. You can't judge a book by its cover, and an LLM can't either. If all you keep is the final deliverable, the agent gets the cover. Give it the chapters: what happened, what you tried, why you decided what you decided. You don't have to write it out yourself. These tools can document the process for you. But the earlier you start, the better it gets.

"Being 'good at prompting' is a temporary state of affairs. The current AI systems are already very good at figuring out your intent, and they are getting better." Ethan Mollick, Co-Intelligence

Mollick's point holds for models too. Prompting skill fades, and so does any one model's lead. The documented history of your work doesn't.

Everyone gets an assistant

I find it fascinating where this is going. Every single person can now have an AI assistant that works alongside them, answers questions and gets things done with them. I think we're underestimating what that does to how we make decisions and how we see the world around us. Admin that took days or weeks, dealing with public offices, finding the best price on something you've wanted for ages: an assistant can handle all of it, as long as it has the right context and access to your data.

I've been working in digital since 2005. I've watched how we use data change, seen marketing automation arrive, and lived through GDPR. AI touches all of those at once. The model will keep changing. Your context is the part you get to keep, so start writing the chapters now.

Sources & further reading

External
Paul Couvert (@itspaulai) on Threads: "the best agent implementation to date"
Meta: Introducing Muse, the world's first personal AI agent built for everyone
Ethan Mollick, Co-Intelligence: Living and Working with AI (Portfolio, 2024)

Related posts
Meta's Muse will book things for your customers. The hard part is the first booking.
Claude just merged Chat and Cowork: the decision that disappeared, and the caveat that didn't
My scheduled Claude Code routine failed three weeks running: what agentic workflows need that a written lesson doesn't give them
A context window is working memory, not a hard drive. Feel that, and you manage AI differently.