Judging by the last few months of AI-econ discourse, apparently the firm is dead.
It’s over––companies have no reason to exist. Agents run the economy now.
The narrative is not coming from AI anime pfp schizos, but from top labs, crypto or fintech pumpers, and economists.
The narrative goes––the firm will implode birthing billions of agents who transact.
How long until agentic payments overtake human payments?
Coinbase’s Brian Armstrong asked.
Ramp is shilling agentic payments, as is Visa and of course Stripe.
DeepMind’s Nenad Tomasev wrote in September
The rapid adoption of autonomous AI agents is giving rise to a new economic layer where agents transact and coordinate at scales and speeds beyond direct human oversight.
with expansions by Deepmind’s Séb Krier.
The hype’s intellectual forerunner is a paper by MIT and Harvard economists who last year foretold the coming Coasean Singularity.
Basically they argue, if you actually had competent, cheap AI agents doing search, negotiation, and contracting, like your own daemon, then a ton of Coasean reasons firms exist disappear, and a whole market design frontier reopens.
summarized Rohit Krishnan.
Non-standard contracts, custom pricing, emails, phone calls, etc… can and will be done by AIs talking to AIs.
wrote one of the paper’s authors Andrey Fradkin.
This singularity is a direct (and I’ll argue, naive) application of Ronald Coase’s Nature of the Firm. Ronald Coase famously argued the firm exists to minimize transaction costs. It grows larger when it’s cheaper to hire than to transact externally. It grows smaller when contracting externally is cheaper.
“Transaction costs –> 0 ==> Firm does not exist.”
Some of this hype isn’t shocking – it’s no surprise payments people would like to hype up a new payment modality on rails they own. Allegations abound that labs talk AI disruption as a distribution tactic to juice their own equity or feelings of importance.
But are they right? Will we soon see Billions Of Agents Transacting? Are we entering the Coasean Singularity?
Intro to Capitalism Link to heading
These days San Francisco appears more communist than China: China embracing open models for decentralized deployment while American labs cozy up to the state.
I should be forgiven, then, for wondering whether the West Coast has forgotten what capitalism is about.
A16Z’s Alex Danco gives the clearest pronouncement – I’ve committed his form to memory, paraphrasing
Capitalism is the delivery of shareholder value via the abundant provision of scarcity
Capitalism is about scarcity.
AI is an apparent technology of abundance. What is scarce in our new ASI future?
Certainly not intelligence. Soon AI will be too cheap to meter.
It’s my continued estimation that the only remaining sources of scarcity are
- trust
- human attention
- and context
Everything else – and we’ve heard a lot of handwringing from private market SaaS investors - workflow lock-in, domain knowledge, data moats, probably others, is cope. It’s a rejection of the moment. This is not AGI.
So, taking this humble proposal of scarcity, we should unpack the claims of “trillions of agents” coordinating through market protocols in a decade.
The Economy as Competitive Claude Prompting Link to heading
Everyone uses Claude now.
The economy basically is reducing to a game of competitive Claude prompting.
So if we’re all using K̶i̶m̶i̶ ̷C̷u̷r̷s̷o̷r̷ Claude Code, what differentiates any of us?
What’s differentiating? Link to heading
Julien Bek at Sequoia says it’s judgment. That human judgment is a durable wall.
You can both recognize that the METR thinking time horizon is noisy, but also be suspicious that
Judgment is different.
Returning to the conservative form of scarcity, you might find heterogeneity or differentiation among firms or their respective agents based on
- trusted distribution (e.g., private access to actuators)
- human attention
- private access to context
Holding the first two constant for a moment, the game then starts to look like that of quant funds, each firm competing on who has the best context to competitively power Claude.
Would Citadel expose their context to Millennium?
AI makes transaction costs explode Link to heading
Today’s simple application of Coase is economically naive.
Per Christensen’s Law of Conservation of Attractive Profits––and as I’ve written previously––AI commodifying execution sends scarcity to adjacent parts of the value chain, namely trusted operations and distribution. Distribution and trusted operations afford access to context flows and commercial venues to monetize this context-powered intelligence. Indeed, it’s ownership of assets that deliver the best access to this context.
While it’s technically true that agents can, per Seb,
negotiate, calculate, compare, coordinate, verify, monitor, and much more in a split second
for a tenth of a penny, doing so typically requires context revelation, a now newly expensive activity. Even indicating that you’re open to transacting leaks signal about preferences, strategy, plans, and observations that––in a world of competitive Claude prompting––you want to keep private. Opening up an RFP tells your competitors what you think is important and even how much it’s worth to you.
The agent discourse jumped to
Agents Conducting Trillions of Transactions
without answering why anyone would bother to transact. This is an important question as transacting in this new era faces new opportunity costs. Ostensibly, an agent would only transact if another captured a resource it needs. Everyone has access to the same models, so heterogeneity only arises from capture of these remaining sources of scarcity. Directionally, it’s interesting that in early experiments of agent economies by @krishnanrohit, agents tend to choose autarky––they decide specifically not to transact and to build everything themselves.
In choosing to transact beyond the walls of the firm or an agent, the entity takes on the role of an information monopolist. There’s something outside its walls that it might need, and these needs are private to the market. Concretely, an entity might externally contract
- AI outcomes (e.g., AI customer service)
- A part for a supply chain
- Market research into customers for a new product
The more competitive the market, or the more differentiating the service is to the business, the more valuable this exclusive information becomes. Firms may transact with platforms (e.g., Sierra or Harvey) that are specifically incentivized to aggregate customer context for improvement of their own service, distributing any alpha from the contracting firm’s context to all its competitors. This misalignment increases the costs of context revelation and reduces total economic surplus.
MIT economist Alessandro Bonatti studied the information monopolist’s problem in 2022. The paper showed the information monopolist makes the most profit by vertically integrating, not exposing the information to the market which brings negative externalities to all.
Alpha is valuable because no one else knows about it. A trillion agents freely transacting sends the value of alpha to zero. [I am suspicious zk solutions could actually help.]
In contrast, of course, quant funds routinely execute their trades off market to obscure pre-trade intentions – supposing that AI agents Are Different Though feels underdeveloped.
It’s a barbell here too Link to heading
It seems this all leads to a barbell of very large firms and many small ones.
Thanks to AI, it’s easier than ever for ‘regular people’ to build software to power business. Mechanical engineers with little technical background make custom software giving Claude screenshots, and asking it to explain things to them like they’re 5. These young small businesses are not context-maxxing, just using new powers of AI to power businesses that have nothing to do with tokens. They’re not aiming for venture scale, just to create software that suits their and their community’s needs.
They don’t need to be a design partner to a Y-Combinator company to play an expensive game of telephone with Claude.
But this will also lead to very very large firms, who build to maximize their context and commercial access, and vertically integrate to shut everyone else out. They don’t want anyone watching or listening to their activities because doing so would give a leg up to a competitor, and they work hard to mine the context they privately commercialize.
The best and most complete example of this is Waymo. Waymo outfits cars others make with a brain it exclusively owns and operates. OEMs reportedly get no access to Waymo data. Waymo is context-maxxing. It accepts no intermediaries––software or embodied ones––to come between it and its context advantage turned trusted service.
Indeed, vertical integration internalizes context. As context becomes scarce, the firm exists not to reduce transaction costs but to prevent the extraction of its scarce context!
What about consumer agents? Link to heading
I’ve so far only attended to candidate agents representing firms. There is a whole possibility of agents representing humans. So far, this medium hasn’t gotten very far. Today’s largest retail platforms are shutting independent consumer agents out, preferring to build AI experiences on platform that maintain the customer relationship and accrue context back to the platform.
The canonical story of consumer agents is helping people book a trip, find coffee (Stripe’s original motivating example from 8 months ago that hasn’t gone anywhere) or buy a super specific kind of shoes that requires a grand aggregation that Google shopping, or Amazon, Shopify or TikTok could not provide. These still feel like a stretch.
A worthy consumer agent will very likely need to be embodied, and perform actual context-mining labor to stand up to data rich platforms that’d otherwise prefer to own the customer. This is the failure of imagination of current commerce analysts
If consumers demand to shop via agents with natural language, then on-platform agents will be the superior way to do that, and the largest eCommerce retailers already support that natively.
wrote Eric Seufert. This is trivially true––an independent agent has nothing to offer platforms who, indeed, are also competitively wrapping Claude. That this needs to be stated feels more a reflection of today’s breathless hype than any deep argument. Rather, it’d only be via building its own data gravity––clearly through embodiment though perhaps there are other avenues––that an independent agent could accrue power to stand up to existing platforms. This changes the value proposition of an independent agent from “chat to shop” to “the agent knows what I like better than anyone.” This is a variant of The Diff’s Edge Router thesis.
Empty agent economics Link to heading
It’s true that early examples of technologies often look like toys, but it’s telling that today’s flagship examples of “agentic payments” are API calls that already accessible through user accounts and sell no scarce resource. They accrue no network effects nor accumulating customer information on interactions. You just must hope that Claude always chooses you.
In Stripe’s example, while it may have been Parallel was the first to ship machine payments, the demo agent may not have optimally served the user, choosing more expensive Parallel over Cloudflare or Bright Data. It’s unclear how these mediating companies selling to agents are not directly a race to the bottom when they’ll ultimately advertise a menu of services that Claude will directly price compare.
So before we breathlessly predict the rise of trillions of agents transacting guiding us to a Coasean Singularity that destroys the firm, I’d ask, what market incentives actually induce or allow this to happen?
I’m working on a paper to formalize this –– writing institutional form surplus and dominance as a function of information economics.
Will ship the paper soon.