Mark Zuckerberg's Muse: an uncut gem?

Welcome to Product Cocktail, where the takes are as polarizing as a shot of Fernet—but the insights come together like a perfectly crafted daiquiri.

9.21.2026. 8:45 PM EDT: "I like what I've seen from Muse so far."

When a close friend (who also works in this Stockholm Syndrome profession—as a Microsoft AI PM) sent that in a group message, that's when I started taking Meta's personal AI agent seriously.

I had ignored the initial announcement in early September, having been focused on Proof, interviews, and the shenanigans involved in parenting.

During that time, Muse had gone viral, hitting #1 in both the App Store and Play Store within 10 days. As of yesterday, the app has been downloaded over 5 million times (in 22 days). By comparison, ChatGPT took 56 days to reach the same milestone, with Claude at 103 days and Grok at 492 days. Muse has maintained a 4.9 across 107K App Store ratings and a 4.7 on the Play Store with 38K reviews. Note that downloads aren't adoption, so this doesn't really tell us how sticky the product will end up being after launch hype dies down.

Meta's stock is up nearly 20% since launch, even after a decline following OpenAI releasing a competing agent product (Dots, September 29th). (Disclosure: I've been a teeny Facebook/Meta shareholder since their IPO day).

NY Times reporter, Eli Tan, called it "the most useful A.I. app [he] had ever used" after using it for two weeks. In the same breath, the intrepid technology reporter mentions "one clarifying moment came after I connected my credit cards to Muse..." Evidently the wins start when the access does.

Claire Vo—AI founder, OpenClaw power user, and host of the How I AI podcast—reviewed Muse two weeks ago, saying it "might be [her] favorite designed agent [she] played with in a long time." Her review emphasized the strong consumer positioning, controls and permissions, novel and well designed features, and the avatar.

With all the hype around Muse, I spent the last week using the product and talking to other users to give you an early read.

First off, Muse is actually a compelling product. It's Meta's first consumer-facing AI product that I actually give a shit about and one of their most interesting product launches in the last 10 years—I'd put Facebook Marketplace and Threads in that bucket too. (FYI - Claire Vo's review goes deeper than I will here.)

Muse has a thoughtful, well-built landing page, positioning the product as "Muse, AI that gets it done for you" while flipping through a number of use cases: sorts your email, books reservations, shops for you, saves you money, manages calendars, and builds fitness plans.

The one area it veers off this path (oddly) is explaining the virtual machine sandboxing "Muse Secure VM is a persistent, isolated Linux computer with a full browser that you and your agent can use to complete tasks, allowing you to intervene at any point."

Meta brings in the greatest hits from OpenClaw, like IDENTITY.md, SOUL.md, and MEMORY.md. Each is a markdown file that the agent loads into memory to understand who it is, its personality, and the long-term memory built over time. (For new readers, I was an early-ish adopter in building a now-deprecated OpenClaw agent.)

They don't just stop there, however. This isn't a Snapchat-to-Instagram-Stories type beat, they remixed the agent experience into a genuinely better customer experience with a handful of novel features.

Activity: A chronological, timestamped log of all activities the agent has performed for you, with an audit trail of all of the tool calls and steps for each activity. While this level of detail won't be used by most customers, it's incredible from a product trust standpoint.

Feed: We know Meta loves a good feed, but unlike your opaque Instagram feed and your hellscape of a Facebook feed, this one has a customizable prompt that allows you to tailor what you see.

Ideas: A continuously updated, personalized list of tasks that Muse can do for you. Thoughtful idea in a product category that requires a fair bit of re-education and unlearning from what the median consumer has used ChatGPT to do.

Goals: A long-running goal tracking feature that walks you through setting goals, then checks in on them with you periodically.

Let's dive into how this thing looks in practice.

My biggest hesitation around Muse is that I haven't really found anything to do with it that I can't accomplish elsewhere.

I gave Muse the old college try with a variety of things: finding a deal on bar stools (I found a better one myself), help freezing my daughter's credit (an incredibly manual task it couldn't help much with), ideation on an upcoming newsletter idea (fine but practically annoying compared to Claude Code), among others. Running into these stumbling blocks while experimenting with OpenClaw would be annoying but excusable—for a consumer app, it's a retention problem.

In fairness, it does make a strong attempt to identify and personalize use cases—even the cross-app ads for Muse are tailored (perhaps unsurprisingly, since Meta has an ungodly amount of data on me.) I just haven't found a killer use case that's making me shout "Muse" from the rooftops, while blasting Supermassive Black Hole.

Admittedly, this isn't exactly new news. I've struggled to find obvious "automation" opportunities in my personal life for a dedicated AI agent in the past. I already use Claude Code extensively, albeit for more hands-on work. I'm not deep enough into my parenting journey that I'm battling to keep up with tens of weekly school and kids sports reminders like a game of Beat Saber. My wife and I enjoy finding new restaurants and bars, so I'm not going to delegate that—when I tried anyway, it did a "just okay" job and didn't successfully book the reservation.

In recognition of my own creative/intellectual limitations, I did some customer research. Of about 20 people I asked (mostly people in tech), three copped to using Muse: two Meta PMs and the same Microsoft AI PM.

When I asked how they were actually using it, I got an interesting array of answers: activities to do with their kid, concert alerts ("Songkick, basically" they admitted), restaurant bookings.

My initial reaction to most of this was... "aren't these already solved problems?" Activities? ChatGPT or Google. Concerts? Songkick or Spotify. Restaurant booking? OpenTable or Resy.

I'm not the only one with this perspective. While interviewed on TBPN, Ben Thompson (Stratechery) argued that tech companies keep demoing tasks people enjoy or won't give up, like shopping.

Thompson lands the "solution in search of a problem" thesis by arguing that agentic AI companies should take a "Marketing 101" approach by making the problem felt and then showing how the product solves it. For example, helping people process returns and get money back in their pocket solves a real pain point, as opposed to the vagaries of "being more productive."

One thing worth acknowledging: when I dug further into the positive reviews, the pattern that emerged was among parents. Claire Vo successfully used it to build a quick morning newsletter for her family (niche, but clever). One of the Meta PMs I spoke to has used it for school email follow-ups, family birthday gift ideas, and researching and (unsuccessfully) attempting to get quotes from carpet cleaners. Each of these represents real pain for an especially time-poor demographic.

I'm intrigued, but unfortunately, these are the exact type of tasks Muse failed at when I tried them.

My second biggest hesitation with Muse is that the use cases worth having require deep access to your most personal accounts—email, bank accounts, health data, etc. For a company that generates nearly 98% of its revenue by monetizing your data to show you ads, this gave me more than a bit of pause.

(Full disclosure: As part of the September 8 press release, Meta announced that "Muse Confidential VM" will ship later this year, enabling users to encrypt the entire virtual machine, including data and conversations, with a key only the user holds.)

The area where agents can be most effective—talking to other apps—is also where they're most likely to break down, in my experience. My OpenClaw had repeated failures trying to add invites directly to my Google Calendar, and I ran into similar agent-to-external system issues here.

One of the first ideas I threw at Muse made me positively giddy: grab all of my Facebook friends' birthdays, let me select the relevant ones, and add them as recurring GCal invites. Unfortunately, this broke down during the Facebook authentication step. Muse suggested I manually provide the birthdays from Facebook. No thanks, chief. That's exactly why I haven't done this exercise before. According to a Meta PM I spoke to, this should work. Hell, even their own ads suggest this.

The other legitimately annoying task I asked Muse—to request a damaged goods refund on Waterloo I had ordered from Amazon—failed because Amazon blocks Muse and other agents.

In spite of Meta's insistence that you can "stay in control of what your agent does," the controls don't always work as expected. When YouTuber Matt Robb asked Muse to handle his Marketplace listings, it asked him whether he would "Allow One Time" or "Allow Always."

He tapped Always, thinking it would still ask before accepting offers. It didn't, and shared his address to buyers, prompting one to show up at his apartment building when he wasn't home. (It turned out okay this time—the buyer returned to buy something else.)

David Singleton, of Meta Superintelligence Labs, concluded that there was "no breach of privacy controls." The controls worked as designed, which makes this an even scarier outcome.

Meta tries to make users feel safe, build trust, and feel in control. Unfortunately, the tasks it currently seems to excel at, reliably, are already things you're comfortable with entrusting an AI chatbot with. The bigger things that require trust (connecting accounts) are where the seams show up, and a bad failure can end the relationship. I experienced multiple breakages and I'm still using it—but I definitely don't feel comfortable giving it much more access, and most people are probably not so forgiving.

My third major concern with Muse is another broader agent paradigm problem that it hasn't quite solved. An agent doesn't take work off of your plate, it turns you into a manager.

It takes a lot of work and thoughtfulness to find things for an agent to do and onboard them to processes in the correct manner. I have managed folks directly and indirectly for most of my career. I still found it challenging to "manage" my OpenClaw agent.

And most people don't know the first thing about management. (I see y'all nodding.)

Some of the best examples of people using "personal agents" effectively are from people with deep management experience. Claire Vo, former CPO at Optimizely and Color, CTPO at LaunchDarkly. Expert at wielding swaths of product and engineering teams. Hilary Gridley, former Head of Core Product at WHOOP. Over half a decade managing product teams. (If you want to get better at this, Hilary is the 🐐.)

Muse's answer to this is to "manage up" instead, suggesting Ideas that your cute little agent buddy can take care of on your behalf. This is a great product primitive and the personalization seems to get better over time.

My concern is the "over time" part. After a few days of using the app, these were pretty generic or follow ups on tasks Muse had already failed on.

I'm not writing Muse off yet. In fact, I want it to succeed. Agents aren't necessarily new news—they've been around for a while on the enterprise side—but they feel like the next evolution of AI beyond the chatbots we've come to know and love (some of us, at least). There are real trust concerns here, but there's potentially real value.

If Meta can figure out a way to deliver meaningful wins as part of the product onramp, close some of the handoff/trust seams, and build on the Ideas primitive, I think they might have something here.

Until then, I'm still experimenting with Muse in privacy-cautious manner (no banking info for you, little buddy), but it's not going to be an easy retention story based on my OpenClaw experience.

My last message: "What are you doing each day? I haven't heard from you since July 24 and I'm seeing $12.40 in token spend since beginning of the month."

Send me questions, feedback, and cocktail recipes:
[email protected].

_{Icons made by Icongeek26 from www.flaticon.com.}