||_||_|| ____|| ___ ||__||_ |_||__ _|__| |____|| __||||

When consumer agents took off several months ago, many treated mass adoption as inevitable, with the only open question being adoption timing. Agent orchestrators like OpenClaw and Hermes showed what an agent could do for you, from booking travel to running errands to buying things; if you didn’t have an OpenClaw set up, you were behind the times.

Cracks began to show in people’s love for the category a couple of months later. The excitement around OpenClaw died down as its limitations and poor UX became more obvious, and as Claude Code and Codex moved toward feature parity with the orchestrators. People started asking the hard questions around willingness to delegate purchase decisions, particularly large ones, to agents that can still hallucinate. They also questioned how often people know what they want before they start shopping; if most buying decisions take shape while people browse, there’s less room for an agent to shop for them.

Over the last couple of months, consumer agents have finally begun to achieve more mainstream product-market fit. Meta debuted its consumer agent called Muse in September. Within 40 days, more than 10 million people in the US had downloaded Muse, well ahead of ChatGPT’s 3.7M over the same stretch and a pace only 11 other apps have ever reached, and it hit 2.5M daily active users in mid-October. A series of others offering substantially similar products have launched as well. xAI unveiled Grok Bot, which passed 418K weekly users about a month after its August beta launch. OpenAI launched Dots in late September. An independent startup called Instinct, which opened an invite-only beta in late August, has passed 100K users and raised $1B at a $10B valuation from Sequoia, Benchmark and Coatue, and is reportedly raising again at $30B only a month later. Muse, Grok Bot, Dots, and Instinct heavily subsidize inference costs, operate in the cloud, and neatly package functionality like context management, managing multiple conversation threads, and having computer use that (usually) can circumvent bot blockers. This set of features has been a prerequisite to mass adoption that prior consumer agents like Hermes and OpenClaw failed to hit.

This is likely just the start. Meta is very good at pushing new apps through its existing ones, and within ten days of launch, Muse was getting the majority of Meta’s house ad promotions across its apps. That’s the same playbook that took Threads past X in daily mobile users (though it remains behind in total DAUs). And people are already using these agents to spend serious money, with more than a quarter of Muse’s users shopping through it at an average purchase of over $200, and Instinct is on pace for about $1B a year in transactions.

The consumer agent takeover

Further adoption of these agents will likely start small. But as with a new employee, trust builds with use. An Accenture study showed that only 9% of consumers would let an agent pay for something on its own today, versus 32% who would if they confirm the payment. In my personal use of Instinct, that shift happened within only a week or so. It planned and booked nearly all of my recent trip to Peru from a short list of sights I wanted to see and my general hotel and flight preferences, and it handled follow-up questions from vendors on its own. We also had some emergencies come up during the trip, like a last-minute hotel rebooking, which it handled with barely any input from me. I walked it through the specifics this time; next time I expect to be much more hands-off, given that the trip went smoothly. I’ve also started using it for things I wouldn’t have tried at first, like finding a scuba class for my brother, planning groceries and recipes, and watching prices on a pair of skis ahead of winter.

A relevant maxim is that technology simply makes things cheaper. All computers really did was make computation many orders of magnitude cheaper, which created many of the wonderful emergent technologies we have today that earlier generations couldn’t have imagined, even if they knew the overall premise that arithmetic would become infinitely cheap. Making intelligence orders of magnitude cheaper has a number of knock-on effects, many of them similarly unknowable at this stage, but one of these as it relates to consumers is that everyone is able to live their lives at least operationally much like the most well-resourced people opt to, and cut out all their annoying logistical headaches. Instead of booking their own flights or planning social outings, people will rely on agents, the way executives rely on assistants and chiefs of staff today.

Cheap intelligence means many people can afford help like this, but why would they want it? The primary reason is time savings. US travelers spend almost nine hours with travel content in the 45 days before booking a trip, viewing about 277 pages along the way. An agent compresses that into a short conversation, and because it can coordinate directly with other people’s agents, planning a group trip or dinner no longer takes a dozen texts. There’s also reduced mental load, since agents often act before you ask. They can monitor prices on the skis you’re looking to buy, flag the renewal you forgot about, and rebook your dinner reservation when your broader plans change.

A common argument against people adopting consumer agents is that human behavior is naturally emergent from the selection process. You might not know, for example, exactly what flight time you want until you’re forced to confront the Google Flights screen with the options, and think through live what the day you land looks like. Clothing shopping is a purer example; you often don’t know what you want until you’ve seen it. I would argue that emergent behavior is inherently predictable. Netflix reported that its recommendations drive about 80% of hours streamed, so what most people “discover” was actually pushed in their direction first. And agents will eventually figure out what your emergent behavior will be, as agents will get to know you and your preferences so well that they can put themselves in your headspace and fill in those details for you. A study found that a model reading someone’s Facebook likes judged their personality better than a friend with 70 likes, and better than their spouse with 300. Beyond this, I think it’s important never to underestimate human laziness; people don’t want to spend their time looking at flights at the end of the day, and if an agent they interact with through a unified, easy interface can figure out their preferences to a point where it’s good enough, I think they will opt to book through an agent.

What’s the business model?

Today, consumer agents are enormously costly. Muse gives every user a ton of tokens a week plus their own cloud computer, and one outside estimate puts the processors and memory behind those computers at roughly $28 per user in hardware cost. Running the models represents costs on top of that: ChatGPT costs OpenAI roughly $16 per user a year in inference, and agents burn about four times as many tokens as chat, which implies over $60 a year for each agent user. Meanwhile, the massive global compute bottleneck has reared its head to users, with Meta saying Muse users consumed 10x what its internal testers did and Instinct users complaining that the assistant is going silent for long periods. To keep up, Meta expects to spend $135B to $145B on capital expenditures this year, roughly double 2025.

Personal agent products are subsidizing the astronomical inference costs of every free user they let in, which is why Instinct has already raised $1.4B in the last five months. From a user-facing perspective, judging by how consumer products usually develop, they’ll likely stay completely free for most; at least Meta expects this for itself. Merchants aren’t paying yet either; Meta doesn’t charge them anything for Muse checkouts.

So personal agents are incinerating heaps of cash at the moment. But I think their path will parallel social media’s, which also famously had no strong monetization scheme for years until it reached critical mass, and then started monetizing at unprecedented scale. Twitter launched in 2006, didn’t sell its first ad until 2010, and didn’t post an annual profit until 2018. Facebook didn’t turn an annual profit until 2009, five years after launch, and as late as 2011, the year before its IPO, it made only about $7.70 annually per daily active user. Meta now makes about $57 a year per daily user across its apps. Google’s path was similar, launching AdWords in 2000 as its main monetization point after some initial resistance. Revenue went from $86M in 2001 to $1.47B in 2003 once Google switched to pay-per-click auctions in 2002.

There are several ways I think personal agents can monetize at massive scale:

• “Sponsored decisions”: Merchants pay the agent to steer decisions their way, both on what the user asked for and on things the agent anticipates they’ll want. This could be well positioned to absorb the hundreds of billions of dollars advertisers already spend trying to influence purchases. I think this will be the dominant and most lucrative path.
• Transaction fees: A cut of every purchase or booking the agent completes, paid by the merchant, like a card. This is what Zuckerberg has said Muse will charge, and it’s simple and familiar to merchants. Rates vary widely by category, from a few percent on most goods to travel, where Booking kept about 14.5% of the $186B in bookings it handled last year, and boutique hotels offer up to 30% per booking. I think this will be a strong, but secondary, monetization arm.
• Banking: The average savings account currently pays 0.38%, while Wealthfront pays 3.55%, meaning depositors are leaving roughly $330B a year on the table. Apollo has warned that agents like Muse could sweep that cash into higher-yield accounts, setting off an “agentic bank run,” and there’s some evidence that it’s starting; Plaid told Meta it hadn’t seen activity like Muse’s on any other service. The largest banks are already pushing back. Since mid-October, JPMorgan, Citi, Wells Fargo and Schwab have moved to block or limit outside AI agents, and five of the UK’s biggest lenders followed in early November, before Muse has even launched there. I think this could be a temporary roadblock, though, as it’s difficult to block a customer approving a transfer their agent instigates or sets up. If agents successfully direct money into new accounts, whoever directs that money gets paid for it, an example being Circle spending $1.7B on distribution last year, most of it paid to Coinbase for putting USDC in front of its users.
• Subscriptions: Charging heavy users for better models or extra usage, as OpenAI does with Dots and Meta plans to do with Muse. It could help cover the cost of some high-engagement yet low-value users, but it’s too small to be the main business and constricts top of funnel in the current customer acquisition war.

“Sponsored decisions” tops the list because it does what advertising does, only much more directly. Advertising at the end of the day is just nudging consumer behavior toward purchasing a given product or service, and each impression produces only a very small chance that someone actually buys. Advertisers on Meta pay about $14 per thousand impressions, and with roughly 2% of viewers clicking and under 2% of those buying, it takes close to 3,000 impressions, or roughly $40 in ad spend, to produce one purchase. Most advertisers now bid for purchases rather than impressions, but Meta still bills them per impression, so the price of every sale bakes in the cost of reaching everyone who saw the ad and didn’t buy. As people start delegating decisions to their agents, merchants can instead now buy access to a major part of the customer’s decision-making apparatus. An agent that has significant power to make the decision skips those 3,000 impressions entirely, so a merchant should be willing to pay a much larger share of the item’s price for each order an agent sends it. Maybe even more powerful than that, the agent can nudge people’s emergent behavior that it anticipates toward a given merchant, buying things for a user that are still relevant to their needs but that they never would have known they wanted in the first place. Merchants also won’t need new budgets to pay for any of this; it can come straight out of what they already spend on advertising.

“Sponsored decisions” is somewhat dystopian on its face, since if this gets out of hand, you could envision a world where agents are nudging user purchase behavior in strange directions and making subpar decisions in favor of “advertisers.” The risk isn’t all that different from what happens in practice today, though. A slightly worse hotel or a few extra dollars on a pair of skis is a cost most people already accept from booking sites and sponsored search results. For most purchases outside of purely commoditized products (of which there aren’t many), ideals like “best execution” are really fuzzy and hard to prove. Zuckerberg has ruled out the “sponsored decisions” direction for now (echoing Sergey and Larry’s initial hesitance toward Google being an ads business), stating that Muse will not run ads and will instead profit by taking a small fee from transactions. But I think there’s a strikeable balance here, and I wouldn’t be surprised if these companies are able to charge much more than just a flat Visa-like fee on transactions based on this approach of directly impacting user purchases.

Whether the business model ends up being “sponsored decisions” or pure transaction fees, because agents get paid on spending rather than attention, they don’t need Instagram-level daily engagement to be huge businesses. The top 10% of US earners account for about half of all consumer spending, so these households that run their travel and shopping through an agent would be worth more than a typical user of Meta’s apps, even if they use the agent only a few times a week.

The big platforms will heavily compete for consumer agent dominance. So where should startups compete?

It’s going to be very hard to compete as a new personal agent company, because the category comes with unusually strong lock-in and barriers to entry:

1. Compute costs: Subsidizing inference for millions of free users is a massive financial undertaking that few companies can afford.
2. Personal context: Agents get to know you extremely well, and providers will make it hard to take that context to a competitor.
3. Leverage with merchants: Large agents can push merchants to integrate natively and pay for the traffic, while small ones get blocked. Muse’s Connectors drew 1,500+ developer applications in under a week and already check out at Walmart, Best Buy, and Sephora. In early November, Amazon started letting customers text its own shopping agent (while blocking other agents), and its CEO expects Alexa for Shopping to be the primary way people shop Amazon.
4. Social network effects: Social features bind agents together and handle group coordination, especially in a closed system, as Instinct has done so far. Its Trusted Person Network lets members’ agents coordinate plans directly.
5. Platforms’ existing distribution: The platforms can push their agents to billions of people who already use their apps every day.

But there’s a lot of surface area that we believe will be lucrative around serving personal agents. Some examples we’ve been spending time on:

1. Spending limits, identity and chargebacks: Agents shouldn’t have carte blanche over our spend; they need limits, an identity tied to the person behind them, and a record of what that person actually authorized, so there’s a way to settle who pays when an agent buys the wrong thing. Our portfolio company @Catena’s Agent Commerce Kit links each agent to the entity behind it, and Catena received conditional OCC approval in September for a trust bank built for AI agents.
2. Agents for sellers: Most service businesses, like plumbers, clinics, and restaurants, have no way for an agent to natively book them, which is why agents call them on your behalf today. As this becomes more commonplace, businesses will want to process these agentic bookings with their own agents.
3. Insurance: Today, users eat their agents’ mistakes, and banks are starting to put that in writing. From November 24th, Wells Fargo will treat any instruction from a customer’s agent as “initiated and authorized directly by you, without further verification.” Underwriters don’t yet have the expertise to easily price this risk in-house.
4. “Sponsored decisions”: It’s possible that the new-age ad engine, or at least core pieces of it, for “sponsored decisions” will be built externally from any one agent, although I think it would be in the strong interest of a dominant personal agent company to own this tech itself. The acquisition of the ad exchange DoubleClick for $3.1B by Google in 2007 shows a route toward M&A for a company building toward this model.
5. Portable preferences: This could look like the Plaid of this market: a profile of everything important about your preferences that you own and can share with any agent.

There are countless others, like cybersecurity for agents, helping merchants get picked by agents, fraud detection, MCP setup services, and purchasing agents for businesses. And just as few people could have predicted Uber in 2007 when the iPhone was first launched, the biggest opportunities here are probably ones that are hard to conceptualize right now, before the agent-driven world takes shape at massive scale.

Looking ahead

Consumer agents’ detractors looked correct even as recently as a couple of months ago. Agents were a chore to set up, too token-hungry, unreliable, and routinely blocked by websites. This wave of newcomers, though, has fixed enough of those issues to reach millions of weekly users. Now that product-market fit is being shown, the art of the possible is beginning to reveal itself. I expect agents to go from booking the occasional trip to handling most routine purchases and managing where people keep their cash. These businesses may end up being some of the largest in the world, serving as the gatekeepers to, and even primary decision makers behind, much of consumer spending, and I think there’s going to be a ton of fertile ground around serving these companies.

If you’re building the infrastructure that world will need, I’d love to hear from you.

Disclaimer: The views expressed here are those of the individual CoinFund Management LLC (“CoinFund”) personnel quoted and are not the views of CoinFund or its affiliates. Certain information contained herein has been obtained from third-party sources, which may include portfolio companies of funds managed by CoinFund. While taken from sources believed to be reliable, CoinFund has not independently verified such information and makes no representations about the enduring accuracy of the information or its appropriateness for a given situation.

This content is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. You should consult your own advisers as to those matters. References to any securities or digital assets are for illustrative purposes only, and do not constitute an investment recommendation or offer to provide investment advisory services. Furthermore, this content is not directed at nor intended for use by any investors or prospective investors, and may not under any circumstances be relied upon when making a decision to invest in any fund managed by CoinFund. An offer to invest in a CoinFund fund will be made only by the private placement memorandum, subscription agreement, and other relevant documentation of any such fund and should be read in their entirety. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by CoinFund, and there can be no assurance that the investments will be profitable or that other investments made in the future will have similar characteristics or results. A list of investments made by funds managed by CoinFund (excluding investments for which the issuer has not provided permission for CoinFund to disclose publicly as well as unannounced investments in publicly traded digital assets) is available at https://www.coinfund.io/portfolio.