Just in time for Halloween, Wall Street has a new hobgoblin. It's called the persistent agent, and on the scariness scale, it falls somewhere between a wet gremlin and a bond vigilante.
By now you have surely heard of Muse, launched Sept. 8 by Meta Platforms, the company behind Facebook, Instagram, and creepy bionic eyeglasses. Another early contender is privately held Instinct, whose agent is invitation-only for now. This past Tuesday, ChatGPT owner OpenAI launched an agent called Dots. Anthropic has Claude Cowork and Alphabet has Gemini Spark. Elon Musk's Space Exploration Technologies, or SpaceX for short, has Grok Bot. Apple has, well, not much, but maybe enough-we'll come to that.
Musephobia has helped send shares of Planet Fitness and Sirius XM Holdings down 19% and 13%, respectively, since Sept. 7, on fears that agents will make it easy for busy and distracted consumers to cancel forgotten subscriptions. Airbnb and Booking Holdings are down 13% and 17%, respectively, because Muse can chisel down travel costs through tireless robo-shopping. And Charles Schwab is off 11% because the agent might coax customer balances out of checking accounts paying 0.01%.
So far, there have been more stock tricks than treats, but Meta itself is up 18% since Muse's launch. This past week, a Morgan Stanley travel analyst theorized that winners in Muse-assisted travel will be companies with differentiated goods, relatively limited shopper attention, and high distribution costs. Cruise lines check all three boxes, and, coincidentally or not, Royal Caribbean Group and Carnival are lately running slightly ahead of the market.
Two new reports are more jarring. One from BofA Securities explores whether Muse is a threat to Apple. Unlikely, it concludes, but the question is more telling than the answer. Why would a month-old app pose problems for a hardware company with 2.5 billion installed devices worldwide? Barclays Capital, meanwhile, wins on financial audacity. It estimates that the prize in today's persistent agent turf war could be $10 trillion in stock market value. That's nearly two Nvidias.
When the stakes are that high, the fear is rampant, and the technical details are complicated enough to make a cyborg cry in binary, it's time for another installment of my semiregular How to Pretend to Understand series. Past topics have included Brexit and quantum computing. The idea is to quickly arm the layperson with just enough background on a topic to survive an office conversation about it, lord a couple of facts over less-informed friends, and maybe form some educated investing guesses.
Personable as your favorite chatbot might seem, it doesn't really chat. It converts questions into numbers representing words or word fragments (tokens), then uses a probability engine to determine the next best number to add, again and again, until it finishes its "response" and converts its added numbers back into words. A persistent agent, on the other hand, doesn't simply ask itself which token comes next in the answer, but rather which tool it should use or sub-task it should perform to get closer to an assigned goal.
The "persistent" part means that these agents can keep working even after you've logged off, and learn about you as they go, like a human assistant. For that, they need their own computing sandbox. Pioneer OpenClaw, launched late last year, can run locally on user machines. Muse uses a virtual machine, or pretend computer made from software running loads of them on an actual computer, like tiny apartments inside a massive building. Agents also use "headless" tools, meaning they can browse websites, fill out forms, place orders, send messages, and even make synthesized voice calls, entirely in code, without any of the windows or graphical front ends that humans need to make sense of the internet.
Two other important concepts to consider are connectors and payment rails. Connectors are things like email, calendars, account logins, and other sources of personal information. The more you tell an agent about yourself, the more it can do to (hopefully) help. A Barron's colleague used Muse to find cheap pants and book a toenail-related doctor's appointment, but only by providing his credit card statement did he find and kill a forgotten subscription. Payment rails allow agents to execute transactions. Muse uses a third-party payment company called Stripe, for customers who've made peace with Mark Zuckerberg snooping through their social media but aren't ready to hand him their credit cards.
Trust is a big deal here. Apple has plenty of it, because it makes money by charging dearly for its wares, not selling customer information. It has its own payment rail, Apple Pay, but Siri isn't a persistent agent, or even much of a chatbot. Apple's artificial-intelligence trump card was supposed to be that tasks would increasingly run on local or "edge" devices like iPhones, but it's unclear how equipped today's phones are for staying up all night to perform a persistent agent's duties.
Meta has less trust. But it has an early-mover advantage, and an excellent handle on how and where to get paid on services that don't have customer fees. That's important, too: Barclays reckons the cost to host a persistent agent user is around $4 a month, or 10 times the cost of a chatbot user. Amazon.com has banned Muse as a shopper on the belief that if anyone should get paid to steer customers, it should be the store.
Expect today's agent hopefuls to compete on cost, point of access, third-party application tie-ins, consumer versus commercial use, data sovereignty, and permission-Muse likes to ask before doing things, while Instinct prefers a freer hand.
Barclays says the first agents to lure customers to share connectors hold the advantage, and that Muse and OpenAI Dots are well-positioned. BofA says that Apple can eventually win on trust, but that it will be difficult to do so while releasing major software updates only annually, because persistent agents grow more powerful every day.