Like it or not, AI agents are cropping up everywhere.
Big tech companies are betting that in the not-so-distant future, we’ll all habitually rely on these systems as automated personal assistants, letting them handle our drudgework while we do something more worthwhile (sleep, maybe). Yes, agents have a somewhat tarnished reputation in the aftermath of the Hugging Face hack and similar incidents, and yes, they come with a pretty serious privacy trade-off (namely that in order to be useful they need access to your personal data). But fear not, because some of the most powerful agents on the market now come with cuddly-looking avatars that look like they were pulled straight from a cartoon for toddlers. And if there’s one thing the history of capitalism and corporate marketing has taught us, it’s that giving a potentially dangerous product a cartoon mascot eliminates all risk.
The most famous of the new fleet of agents is Meta’s Muse, personified by “Jolly,” a character that looks a like a mix between a Teletubby and a claymation yeti. The company really needs people to like Muse: While the majority of Meta’s revenue still comes from its ad business, it’s on track to spend as much as $145 billion this year on its AI buildout—a massive investment it still needs to justify to investors. At the company’s annual Meta Connect conference late last month, and with Muse at the top of the iOS App Store’s list of free apps in the U.S., CEO Mark Zuckerberg told the audience that “the centerpiece of what we’re building is Muse.” At its own developer conference less than one week later, OpenAI unveiled its own AI agents, called dots, which closely resemble Muse aesthetically and functionally.
Curious to see what all the hype was about, I gave Muse a spin to find out how useful and intelligent it really is. To put it briefly, it was not encouraging.
Housekeeping tasks with Jolly
The whole point of using an agent as opposed to a more limited chatbot is that it should be able to handle complex, long-running tasks without humans having to monitor and sign off on their every move. But the longer the leash they’re given, the higher the chances they’ll act in surprising ways; with great agency comes great liability. Meta reminds us of this on Muse’s sign-up page, which warns that the agent can occasionally make mistakes. Indeed, it already has.
I got set up with the free version, which comes with a weekly token usage limit. (There are also $20 and $100 monthly subscription options with higher allowances.) The set-up is straightforward: You choose a name for your agent—I was given the choice of Baxter, Jewel, and Muse—and, from the settings page, choose which apps you want it to connect to. For starters, I chose my personal Gmail and Google Calendar accounts. Enough to give it some control over my personal data without running much risk of catastrophe.
My first jaunt with Muse felt like a good chance to tackle some housekeeping chores I’d been putting off. First, an easy one. My girlfriend and I planned to see a play that evening, so I asked Muse to find a reasonably priced restaurant in Midtown Manhattan with a table for two at 6 pm. Its top choice (a Mediterranean spot) looked good, so I asked it to book a reservation and add it to my calendar, both of which it promptly did. And I have to admit, that part alone—just dashing off a quick command and watching the reservation confirmation land in my inbox—felt good; I was already starting to see why this was catching on with people. But when I opened my laptop and checked the restaurant’s menu, I immediately saw that its prices were considerably steeper than the budget range I’d asked Muse to stay within. After a little more back-and-forth, I eventually got frustrated with Muse’s slow response times and decided it would be easier to just walk in to one of my favorite restaurants. Old habits one, Jolly nil.
Next up, requesting refunds from airlines after some very delayed recent flights had forced me to pay for Lyft rides home in the wee hours of the morning. Muse did well for a while: I gave it screenshots of my receipts and information about my flights, and it drafted some decent email drafts—courteous, but firm. But within one of those drafts, it had buried an important message telling me that I’d have to copy/paste the drafts into the airlines’ websites, since they didn’t field refund requests via email. I ended up needing to take things into my own hands again, opening a browser and entering my information manually. (I did use the message it had originally drafted, though.)
But the main task I wanted to handle with Muse was booking my holiday travel, something I have an annual habit of putting off until the last possible minute but was hoping to get a head start on this year, given how ungodly expensive air travel has been. Travel-booking has been one of the major use-cases Meta has highlighted with Muse. But here too, I found the experience clunky. It would’ve taken me maybe twenty minutes to compare round-trip flight options on Expedia, find the best option, and move on to checkout. Instead, I fell into a long and rather disorganized conversation with Muse, reviewing the various flight itineraries it offered me and repeatedly scrolling back up to remember what those options were. (When I asked it to connect directly with Expedia, it told me it wasn’t able to connect to the site directly.)
It occurred to me to check the United Airlines website in case I had miles I could use towards my flights, which, as I had feared, looked like they were going to cost upwards of $1,000. Unfortunately for Meta, United prohibits bots from accessing its site. Numerous other companies have similar policies in place; last month, Amazon specifically banned Muse from using its online store. When it encounters roadblocks like these, or CAPTCHAs, Muse prompts the user to take manual control of the browser. But from the browser window within my chat with Muse, the United website would not accept my manually entered login information. Yet again, I ended up having to part ways with Muse and handle the task myself.
Not yet the age of agents
In all three cases, Muse felt less like a seamless assistant than a slow and superfluous middleman. Granted, this was just a first run, and I do not doubt that if I continued forking over my personal information to Muse and let it connect to more of the apps I already use, its performance would have improved. I haven’t given up on it, even though my first impression wasn’t excellent. Overall, the experience gave me a clearer sense of the direction AI agents are heading, but it also clarified for me just how far these tools have to go before they become a mainstay in most users’ lives. People are creatures of habit, and it’s going to take time before using and trusting an agent like Muse becomes second nature—no matter how adorable tech companies try to make them.
I have little doubt that at some point down the road agents will probably be handling many of the routine tasks that currently eat up so much of our time. Ten or twenty years from now, our descendants might look back on footage of people in the 2020s manually booking flights on Expedia or making a reservation on OpenTable the way that we read about newspapers being drafted with movable type, or early Disney animators assiduously sketching one frame at a time. Charming anecdotes from a bygone technological age, perhaps, but also a huge drain on time that people are arguably better off without.
The kind of fully autonomous systems that have taken software engineering by storm could eventually go mainstream for monotonous, everyday tasks. But that all remains hypothetical, and tech companies’ multibillion-dollar bets on agents are still just those: bets.