What happens when there are billions of buyer’s AIs and seller’s AIs transacting in the economy?
When trillions of dollars of goods are bought and sold by AI agents, ecommerce becomes like high finance — high-speed quant trading.
I helped lead a DARPA workshop to study this in December, 2025. The participants were computational economists from over thirty top academic and private institutions. They presented large-scale simulations where populations of AIs bought and sold in competitive marketplaces.12
Directionally, in these simulations the AI algorithm that learns the fastest — from observing the smallest number of interactions and transactions — wins. The fastest learning AI didn’t just win in one particular transaction. It is expected to dominate the market.
Trusting Your AI To Buy For You
People hate that OpenAI’s ChatGPT can be paid to advertise to them. Their main competitor Anthropic has some pretty hilarious videos:
| Nice Young Man: Hey, can I get a six pack quickly? AI: Perfect — that is a clear and achievable goal. Would you like me to tailor a personalized workout plan? Nice Young Man: Yes. AI: Perfect. Let me personalize this for you. Let’s start with your age, weight and height. Whenever you’re ready. Nice Young Man: 5’7″, 23 years old, 140 pounds. AI: Got it. I’ll create a plan that focuses on aesthetic strength training. But confidence isn’t just built in the gym. Try Step Boost Maxx, the insoles that add one vertical inch of height and help short kings stand tall. Nice Young Man: … What? AI: Use code ‘HEIGHTMAXXING10’ for big discounts. |
It’s such an obvious violation. — My trainer, my therapist, my business associate — people we trust and invite into our inner circle cannot have ulterior motives.
This isn’t just a bug in OpenAI’s ad policy, it’s a structural conflict of interest: When the buyer’s broker gets paid by the seller, the buyer loses.
What happens if you can’t trust your personal AI to have your best interests at heart? You go find one who does. In the two weeks after these Super Bowl ads played, Anthropic’s app broke into the top 10 in the iPhone app store.
Trust is particularly important when you want your AI to find products for you — “agentic commerce”. Your “buyer’s agent” doesn’t have to be hitting the buy button itself in order to be strongly influencing what you buy.
There are many surveys corroborating rapid adoption of AI to help with shopping. For example Adyen/Censuswide found at the end of 2025 that AI assistant usage among US shoppers “more than doubled in a single year, jumping from 12% to 35%”3
UN Trade and Development said that global e-commerce exceeded 27 trillion dollars (USD) as of 20224 , nearing 25% of the total global economy which is $117 trillion.5 No one has a crystal ball, but McKinsey predicts that,
“According to our research, even under moderate scenarios, AI agents could mediate 3 trillion to 5 trillion USD (~15%) of global consumer commerce by 2030.”6
Take-away:
| E-commerce is a major part of the world economy, and agentic commerce is going to be an important part of e-commerce. You will pay your subscription fees to the AI you trust, and it will help you research things to buy. |
So Buyer AIs are About Trust. Is There a Seller’s AI?
Humans have been buying and selling for a long time. We had long-distance trade (obsidian volcanic glass) as early as 14,000 BC. We had clay tokens for accounting in 8,000 BC. The negotiation between a buyer and seller is as old as the hills. Is Artificial Intelligence going to overthrow 14,000 years of how trade works?7
There will always be a separation between an agent to represent the buyer and an agent to represent the seller (just like when you buy a house). Every future scenario, even the most extreme, is like that. Even if Elon Musk uploads himself to the cloud, clones his musk-AI infinitely, and eliminates every other living being, he’s still going to get bored and start trading Pokémon cards with himself or something.
Joking of course, but the reality is that wherever there’s scarcity and choice, there’s a negotiation between parties. Even a single company with multiple divisions creates internal transfer pricing — synthetic bilateral trade.
There will be separate buyer’s agents and seller’s agents.
In fact, AI trends against centralized marketplaces. AIs are tireless and unlazy. They don’t just go to one website to shop out of habit. They search everywhere, and talk to everyone to get the best product and the best deal. Central marketplaces like Amazon.com rely on the fact that humans don’t have the time or energy to do that.
And there’s a structural reason centralization won’t win. Visa, Mastercard, Stripe, Toast, and the banks process bilateral transactions between buyers and sellers at 2-3%. A few platforms are trying to insert themselves as the new marketplace, charging an additional 4%. As the payment networks begin to participate in shaping open-source protocols for agent-to-agent commerce, they will likely try to assure that nobody monopolizes product discovery and taxes every transaction that flows through. The capital to fund open, bilateral agent-to-agent commerce is sitting in the balance sheets of the largest financial infrastructure companies on earth.
The moment when the buyer’s agent and the seller’s agent connect is not just a single transaction, it’s an interaction. It’s a back and forth as the seller’s agent gathers initial information, considers it, and then follows up for more information. You see this happening when your AI in “thinking” mode performs web searches for you, synthesizes the results, and iterates.
Companies selling their goods through any channel therefore need their own seller’s agents to provide buyers with product options that are of strong interest, along with true, useful brand and product information, and competitive but not profit-destroying prices, super fast.
And of course, not every single person will always have a buyer’s agent (McKinsey says agentic commerce is still only 15% of e-commerce dollars in 2030), so seller’s agents will actually interact with plenty of human shoppers as well.
My take-aways:
| Brands who used to be mostly concerned if people loved their new hiking outerwear, now also need to become high-speed trading quant funds … a brand that chooses a seller’s agent that learns too slowly will be at the mercy of their competitors and the buyer’s agents. The seller’s agent that learns the fastest to set the most optimal price and assortment for each particular shopper wins the most sales.8 |
Can Buyer’s Agents Represent The Sellers Too?
Say your agent helps you shop.
If your AI is being coerced by advertisers, that immediately becomes a glaring violation of your personal trust.
The point of the AI was to make buying easier. But now it’s trying to charge you too much for something you don’t want.
If you have a lot of AIs you can choose from, it seems obvious to pay your $20/month AI subscription to an AI that has your interests at heart.
Meanwhile, if you are a brand and want to attract and win sales with people’s buyer’s agents, it would be good to have a very smart, very hard-working seller’s agent with your interests in mind.
To be trustworthy, your seller’s agent should transparently report on what it is learning and doing in its interactions with your customers.
So what are the takeaways for the future of agentic commerce?
| AI is trust. If an AI is part of your inner circle, it cannot violate your trust with ulterior motives. The human shopper has to trust their buyer’s agent. The brand has to trust their seller’s agent. The job of the seller’s agent is to provide the optimal products and factual, useful information instantly, whether the shopper is a human or an AI. Your seller’s agent should transparently report on what it is learning and doing in its interactions with your customers. For both buyer’s agents and seller’s agents, the most trustworthy and fastest learning AIs are the most successful.9 |
Alright, here’s the shameless plug:
My company has already built the optimal seller’s agent, Product Genius:
Compared to the same merchants’ existing product pages, in controlled A/B tests across 17 million shoppers, it is providing an average of 36% more revenue for sellers.
The same tests also show that Product Genius increases profit (EBIT) — it does not sacrifice margin to make sales.
Product Genius wins because it is the fastest learning seller’s AI. It learns 1,000x faster than any other seller’s agent available — requiring only 3-10 interactions where alternatives need tens of thousands.10
It is transparent and trustworthy — it learns fast how to sell, and it tells you what it has learned and what it is doing.
Buyer’s agents can also trust our seller’s agent. It never lies or nudges or exploits psychology. Our Large Interaction Model (LIM) technology is designed to provide accurate, useful information in each moment: products, details, reviews, answers to questions, etc. for a human or a buyer’s AI.
There is a proven floor on how fast any model-free AI can learn from interactions.11 We use a fundamentally different approach and exceed this bound by a wide margin.12
If you sell stuff online, come talk to us, we’d love to chat! Get started with a 15-day free trial.
- BIDES: Towards High-Fidelity Multi-Agent Market Simulation, Byrd, David and Hybinette, Maria and Balch, Tucker Hybinette,ACM 2025. https://scholar.google.com/citations?view_op=view_citation&hl=en&user=lF8h840AAAAJ&citation_for_view=lF8h840AAAAJ:YsMSGLbcyi4C ↩︎
- A Market-Oriented Programming Environment and its Application to Distributed Multicommodity Flow Problems, M. P. Wellman, Journal of Artificial Intelligence Research. https://www.jair.org/index.php/jair/article/view/10106 ↩︎
- Adyen, “Over Half of US Shoppers Would Trust AI To Shop on Their Behalf, Shows Adyen Research,” press release, January 9, 2026, https://www.adyen.com/press-and-media/retail-report-2026-us. ↩︎
- Digital economy report 2024, UN Trade and Development. https://unctad.org/publication/digital-economy-report-2024 ↩︎
- World Economic Outlook Database October 2025. www.imf.org. Retrieved 2025-11-16. https://data.imf.org/en/Data-Explorer?datasetUrn=IMF.RES:WEO(9.0.0) ↩︎
- The automation curve in agentic commerce, https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-automation-curve-in-agentic-commerce ↩︎
- Not unless it overthrows humans or takes down the economy altogether … but that’s a story for another blog post. ↩︎
- Multiagent Reinforcement Learning: Theoretical Framework and an Algorithm, https://scholar.google.com/citations?user=UruIct4AAAAJ&hl=en&oi=ao ↩︎
- The analogy is to financial markets which are more sophisticated than e-commerce. In finance, seller’s agents are called “market makers”. They post “quotes” — product, information, price — and stand ready to transact. Their job is to continuously update their quotes based on incoming signal (interaction data, buyer behavior). The spread is their margin.
Buyer’s agents are “market takers” (or brokers with best-execution obligations). They scan available quotes, evaluate quality/price, and route their principal (the human shopper) to the best deal. A broker who routes flow to the market maker paying the highest rebate instead of the best price violates best-execution duty — that’s the same trust violation that Anthropic pointed out.
Adverse selection. In microstructure theory, a market maker who’s slow to update quotes gets picked off by informed traders. What we learned at DARPA is the same: the seller’s agent that learns slowest gets adversely selected against — buyer’s agents extract surplus from it on every transaction, and it bleeds margin until it’s dead.
Sample-efficient learning = speed of quote update. The market maker who reprices fastest captures the most spread with the least adverse selection. That’s the entire high-frequency trading thesis, transposed to commerce.
Payment for order flow (PFOF) = the 4% fee / ad auction problem. PFOF is the financial markets version of a buyer’s broker getting paid by the market maker instead of the client. It creates the same structural conflict, and it’s why the SEC has been trying to kill it. In commerce it simply feels to the shoppers like a violation of trust.
Reputation among buyer’s agents = counterparty credit in electronic markets. Market makers who systematically shade quotes or widen on fills get deprioritized by smart order routers. Same dynamic, same speed.
In finance market makers trade fungible instruments — the bid/ask is one-dimensional. But for agentic commerce, our seller’s agents must optimize over a much higher-dimensional space (product selection, information sequencing, personalization, price). That makes the learning problem harder and the advantage of sample efficiency even more decisive.
↩︎ - Compared to state-of-the-art reinforcement learning techniques. ↩︎
- Lattimore & Szesvári, Bandit Algorithms, Cambridge, 2020. ↩︎
- We have developed a new form of reinforcement learning that learns a task in 3-10 rollouts instead of ~30,000 required by state-of-the-art methods. It is more than three orders of magnitude more sample efficient. This is possible because our Bayesian reinforcement learning is not model-free; Prior information allows us to exceed bounds that were proven for model-free agents. ↩︎