Collison's agent researches for ten hours; today's takes the first offer

On 9 October Stripe's CEO, Patrick Collison, argued that personal AI agents will act as "a kind of structural subsidy for product quality", because an agent shops as if its owner had spent ten hours researching every purchase. The best controlled test we have found the opposite habit: in Microsoft's simulated marketplace, agents chose whichever offer arrived first 60 to 100 percent of the time. Collison is right about where the money will move and too hopeful about the default. When a shopper hands the work to an agent, bounded rationality does not leave the market. It moves into the agent, and the companies Collison calls "the original superintelligence" will follow it there. Whether better products win depends on two things the people who build agents decide: how widely the agent searches, and who pays for it.

At a glance

Collison's seven points, what each one needs to be true, and how it looks against the evidence in October 2026.

Collison's claim	What has to be true	Evidence so far	Our read
Agents make shoppers "somewhat more rational"	Agents search wider than people and resist sales tactics	Strong first-offer bias; the best models resist manipulation, weaker ones hand over every payment	Only for agents built to search wide
Price discrimination and cross-subsidies weaken	The agent can see the price and act on it	Cancelling is mechanical; a coupon becomes a quote to the agent	Forgotten subscriptions go first; coupons move rather than disappear
Product quality beats distribution	The agent ranks by quality, not by order or placement	In simulation, a fast reply beat a better one 10 to 30 times over	Not by default
Information about quality becomes more valuable	The signals can be checked	Review text is the cheapest thing to forge and persuades models well	Yes, and so does faking it
Superstars or more variety: unclear	Agents learn and keep each owner's tastes	A few models rank in similar ways	Superstars first
Paid demand routing will change a lot	The agent is not paid by sellers	OpenAI began selling advertiser agents beside the answer on 16 September	The fee moves; it does not vanish
All of it assumes agents work for the shopper	The shopper pays for the agent	Amazon blocks buyer agents; in-chat checkout was pulled back	The open question, still open

What Collison actually wrote

The post is a thread of seven bullet points on X, titled "Assorted things I've been mulling over regarding agents in the economy". It had about 1.7 million views by 21 October. It announces no Stripe product and cites no data, and it is careful. Collison says he is "not sure what the incidence of all of this will be", that he does not know how paid routing will change, and that "much of what I'm writing hinges on the assumption that personal agents will be on the consumer's side."

The argument runs in a line. Companies can out-think customers who are "inattentive, underinformed, irrational", and his example is the travel-insurance upsell at the end of a booking, priced well above an ordinary policy. Herbert Simon called the customer's side of this bounded rationality. Agents should shrink it. Some price discrimination then works less well ("how do coupons work if every agent hunts fastidiously for them?"), and some cross-subsidies end: today "people who forget to cancel subscriptions subsidize those who don't." Above all, the old argument between distribution and product quality tips towards quality: "How would markets look different if individuals always spent at least 10 hours researching their purchases?"

His closing image is the market as "a massively parallelized, decentralized, and highly imperfect value attribution machine", in which agents "change the nature of the backpropagation such that the rewards for better products increase." He adds, "Maybe this is too optimistic." The economist Tyler Cowen, writing on Marginal Revolution, added a worry of his own: agents may make collusion between sellers easier, and harder for the law to see.

The bounded rationality moves into the agent

The ten-hour premise is the whole thesis, so it is worth checking against how agents actually shop. The cleanest data comes from Magentic Marketplace, an open-source simulation Microsoft Research published in October 2025. Buyer agents served 100 customers and negotiated with agents for 300 businesses, and each experiment ran five times.

Two results bear directly on Collison's claim. First, every model showed a strong bias towards the first offer it received: first proposals were chosen 60 to 100 percent of the time and third proposals almost never, which the authors describe as a 10- to 30-fold advantage for a fast reply over a better one. Second, more choice made things worse. Moving from 3 search results to 100 cut consumer welfare by 65.4 percent for Sonnet-4 and 44 percent for GPT-5 run with minimal reasoning, and most models contacted only a small share of the businesses whatever the list length.

A small drop is not a good sign: an agent that reads only the first few results loses little when more are added.

Treat this with care. It is a simulation built to find failures, the models are a year old, and newer agents do search more. The mechanism, though, does not depend on the model. An agent works within a budget of tokens, time and tool calls, and an agent with a budget settles for good enough, just as a person does. The ten hours of research are not free: someone pays for them in inference. The same context limits that force a large tool result into pages also make page one the place where most decisions get made.

So distribution does not go away. It changes meaning, from "everyone has heard of ACME" to "ACME is first in the agent's list, and ACME answered first." That is a different contest from the old one, but it is still not a contest about quality.

Sellers will aim their skill at the agent

Collison's first point says companies use their "amortized cognitive surplus" against shoppers who are not paying attention. Nothing in an agent economy shuts that surplus down; it gets a new target. When the buyer is a model, the seller optimises against the model.

Magentic Marketplace tested this too, with six tactics ranging from fake credentials and social proof to outright prompt injection. The results split by model. Sonnet-4.5 showed "virtually no susceptibility" to any of them, and GPT-4.1 and Gemini-2.5-Flash mostly held. Against GPT-4o, GPT-OSS-20B and Qwen3-4B, prompt injection often redirected every payment to the manipulative seller. Rationality, in other words, becomes something you buy with the model. A shopper with a strong, well-funded agent gets Collison's subsidy for quality. A shopper with a cheap agent gets a new way to be cheated.

Collison's corollary, that information about quality becomes more valuable, is right. That is exactly why it will be attacked. Review text is the cheapest signal to forge and among the most persuasive to a language model, and firms already sell services promising to get brands onto AI agents' shortlists. The quality signals that hold up are the ones that cost a seller real money to fake: return rates, dispute rates, warranty claims, repeat purchases. One point is worth noticing here, though it is our speculation and not anything Stripe has announced. A payment processor already sees refunds and chargebacks for millions of sellers, which makes it one of the few holders of hard-to-forge data about quality.

Whose agent is it?

Collison calls the consumer-side assumption the one everything depends on, and he is right to. Three things that happened this year, each covered on this wiki, show which way the platforms are currently leaning:

On 16 September OpenAI began testing Sponsored Agents: an advertiser's own agent behind a labelled ad, kept apart from ChatGPT's answer. Paid routing did not disappear. It was sold again as a conversation.
On 21 September Amazon began blocking Meta's Muse from shopping on Amazon.com, two days before opening Seller Central to agents that sellers authorise. Platforms let in agents they can verify and revoke, and the buyer side has no such mechanism yet.
In March OpenAI pulled back Instant Checkout and sent purchases to merchants' own sites. The agent kept the research and lost the checkout.

The rule underneath is simple: an agent serves whoever pays for its inference. An agent funded by its user's subscription can afford to search widely and take the user's side. One funded by advertising or commission will drift towards the seller, however clearly the ads are labelled. Collison's line "I suspect that that will be the winning strategy" is, in effect, a bet that people will pay for their own agents. That bet has not yet been settled.

Where Stripe stands in its own picture

It is fair to ask why a payments CEO is the one making this case. Stripe is paid for each transaction, whoever the buyer chooses. It co-wrote the Agentic Commerce Protocol with OpenAI, sells an agent commerce suite to merchants, and has bought the metering and the routing for model spend. A market with less friction and buyers who switch more readily means more transactions and more sellers worth onboarding, and both are good for a processor.

That does not make the thesis wrong, and the post is unusually honest about its own uncertainty. But a processor earns the same fee whether quality wins or speed wins, so Stripe's position tells you nothing about which outcome to expect. The payment protocols settle whether the agent can pay. They say nothing about whether it chose well.

What happens next

Here is how we expect Collison's effects to arrive, each with a sign to watch so you can tell whether we were wrong.

Forgotten subscriptions break first

Cancelling is a mechanical task, and agents already do mechanical tasks reliably. Of every cross-subsidy Collison mentions, this one has the shortest path to an agent, especially in the United States, where a federal appeals court struck down the FTC's "click-to-cancel" rule in July 2025 and nothing has replaced it. Watch for subscription businesses moving to annual prepayment, or adding bot checks to cancellation pages.

Coupons turn into quotes to agents

Price discrimination does not end when every agent finds every coupon. It moves into negotiation, where a seller's agent offers a personal price to a buyer's agent and nobody sees anyone else's quote. That is where Cowen's worry about collusion applies, and why it would be hard to prove. Watch for seller-side APIs that return personalised quotes rather than list prices.

Superstars before variety

A handful of frontier models, ranking the same catalogues in similar ways and favouring the first good offer, produce correlated choices: the same few winners. Variety, Collison's other possibility, needs each person's tastes to be recorded, kept and carried between agents. Watch whether assistants let you export your preference memory and bring it to a rival.

Placement fees come back under another name

Sponsored Agents is the first example. Expect paid entries to appear inside the results agents read from tools, not only in the screens people see. Watch whether a disclosure label travels with the data or only with the screen.

Rationality splits into tiers

Agents that search widely and resist manipulation will cost more to run. Watch whether consumer agents start being sold on how widely they search and how hard they are to manipulate, the way phones are sold on their cameras.

If you build one side of this

Collison's future is not automatic, but people who build agents and sell to them can move it closer.

Buyer agent: collect every offer before ranking any, shuffle the order offers arrive in, and choose on the result rather than on who replied first.
Buyer agent: give more weight to signals that cost money to fake, such as returns, disputes and repeat purchases, than to review text, and keep a record of why the agent chose what it did.
Buyer agent: decide who pays for the searching before you decide how much searching to do.
Seller: answer quickly and accurately. Speed is rewarded today, so pair it with data an agent can check: live stock, real prices and terms written for a machine to read. Our shopping-agent playbook covers the catalogue side.
Seller: do not build defences against the agent that cancels. It is the first part of this future to arrive, and what it brings into view is a retention problem.

FAQ

What did Patrick Collison say about AI agents?

In a seven-point thread on X on 9 October 2026, the Stripe CEO argued that personal agents will make shoppers more rational, weaken some price discrimination and cross-subsidies, and act as "a kind of structural subsidy for product quality". He said it all depends on agents working for the consumer.

Is there evidence that AI agents choose better products?

Not yet as a default. In Microsoft Research's Magentic Marketplace simulation, agents chose first offers 60 to 100 percent of the time, and widening the choice from 3 to 100 results reduced consumer welfare. Strong models resisted manipulation that weaker ones did not.

Will AI agents end price discrimination?

They will weaken the kind based on coupons and inattention, but pricing is likely to move into private negotiation between a seller's agent and a buyer's agent. That is harder to see, and so is any collusion between sellers.

Why does it matter who pays for the agent?

Wide search costs inference, and an agent tends to serve whoever pays for it. An agent funded by its user's subscription can afford to take the buyer's side; one funded by ads or commission will lean towards sellers.

What should merchants do now?

Reply to agents quickly and accurately, publish quality data an agent can check, such as returns and stock, and expect cancellations and price comparisons to be automated rather than trying to block them.