Five million downloads in 22 days. That was the number staring back at Alex, the builder behind Muse, a few weeks after the AI agent went live, and even he had not seen it coming. Most people had spent the past couple of years treating AI like a slightly smarter search engine. Muse was trying to do something else entirely: hand the kind of AI power that software engineers had quietly been hoarding over to everyone else on the planet. The fact that it worked, fast, was the thing that had the tech world recalibrating.
Cleo Abram sat down with Alex to pull apart exactly what Muse is, what it is still getting wrong, and what its rapid rise says about where AI is headed for ordinary people.
Why agents finally clicked for people who never learned to code
For most of this year, a strange two-track conversation had been running in parallel. Engineers and developers were watching AI agents take over entire workflows and could barely contain themselves. Everyone else was shrugging and saying the tools felt like a slightly better Google. Alex described that gap as one of the core problems Muse was built to close. ‘Most people in the world, they don’t know how to code. They probably never even want to code. But they still wanna feel the experience of agents,’ he told Abram. The team spent months polishing the rough edges that developers will tolerate but everyone else will not, because, as he put it, developers can deal with ‘unripe technology’ while a general audience simply walks away.
The result is an agent that handles the genuinely tedious infrastructure of modern life: navigating insurance claims, canceling subscriptions buried behind a deliberately confusing sequence of clicks, sorting out DMV paperwork, managing small business admin. The team actually stress-tested Muse against the DMV website specifically, treating it as the pinnacle of human frustration. When someone asks why they would need an AI assistant if they only have one restaurant reservation a month, Alex’s answer is simple: almost everyone has subscriptions they should not be paying for, or gift cards they have forgotten about, or a medical claim so complicated they gave up halfway through.
Muse also makes phone calls on a user’s behalf. When Abram had it call someone that morning, the agent introduced itself by saying: ‘Hi, I’m Haley calling on Cleo Abrams’ behalf. I’m transcribing our call for notes.’ The other end of the line knows it is talking to an AI.
The sentinel watching the agent, and why that matters
Building something that does what you tell it without doing things you never said you wanted is, Alex acknowledged, one of the hardest open problems in AI. Muse runs inside a sandboxed secure virtual machine, a self-contained environment where your personal data lives and from which nothing leaves without scrutiny. A separate AI layer, the sentinel, watches everything the main agent tries to send outward, flagging anything that looks like a credit card number or a Social Security number before it can leave. Every time Muse tries to connect to a new website, a dialogue box asks the user to approve it. Some people find it annoying. The team considers that a reasonable trade.
A more encrypted version, the confidential virtual machine, is in development. It was designed by the person who built WhatsApp’s encryption, Moxie Marlinspike, and is intended to offer the same data guarantees WhatsApp provides, including the kind where even Meta could not hand over your information to a government if asked. The reason it did not ship at launch is that there are still technical trade-offs to resolve, including potential speed costs, and the team had already delayed the original Muse launch by several months to get the security architecture right.
The deeper alignment challenge, the paperclip problem scaled down to ‘please book this doctor’s appointment without blackmailing anyone,’ is one Alex said keeps him up more than most things. Muse’s approach combines rigorous red-teaming before any model ships with an ongoing sentinel layer that mirrors, in miniature, the concept of scalable oversight: as agents get smarter, a separate AI watches them, rather than relying purely on human monitoring.
A small detail that is still open
The Muse mascot, the small digital character users interact with, looked in the very first internal prototype exactly the same as it does today. Serious adults sat in a room arguing about whether he was too playful, not colorful enough, insufficiently weighty for the stakes involved. They looked at the alternatives and kept coming back to the original. He was, as Alex put it, just ‘perfect.’
The bigger question sitting underneath all of this is what Muse is actually for, in the largest sense. Alex’s answer is that most people, somewhere between childhood and a steady job, stop dreaming at scale. They had ideas, then the maze of real-world logistics ground those ideas down. His partner wants to start a museum of ice, tracing the full history of how humanity has kept things cold. It sounds niche. Alex thinks that is exactly the point. Muse, if it works the way the team hopes, is the thing that helps someone with an idea like that figure out which buildings to lease, who funds projects like this, and how to begin curating the exhibit, one small cleared obstacle at a time.
