AI agents that shop, negotiate, and buy on your behalf are becoming mainstream in 2026. But the idea is far older than ChatGPT, and the first attempt failed in ways that are instructive for understanding what happens next. This is Part 1 of a three-part series.
There is a common way of telling the agentic commerce story that begins in late 2024 with the arrival of capable AI agents, treats everything before as prehistory, and frames what is happening now as unprecedented. That telling is convenient and wrong. The idea of software agents that shop on your behalf, compare prices across merchants, negotiate deals, and complete purchases is at least thirty years old, was the subject of a serious academic research program in the late 1990s, and failed for reasons that had nothing to do with the vision and everything to do with the technology available to execute it.
Understanding that first attempt matters, because the problems it ran into did not go away. They are the same problems the industry is now racing to solve with protocols, standards, and infrastructure, and the companies and researchers who understand the history are better positioned to see which parts of today’s wave are genuinely new and which are a rerun.
This series covers agentic commerce in three parts. This first part establishes the foundations: where the idea came from, why it failed the first time, why it is working now, and what structurally separates agentic commerce from the e-commerce we have known for twenty-five years. Part 2 covers the supply side: what brands and retailers must do to exist in an agent-mediated market, and how that visibility game can be manipulated. Part 3 covers the money: payment protocols, consumer trust, and the deeper questions of who captures value when machines do the shopping.
The First Attempt: 1994 to 2000
In 1995, a research lab at Andersen Consulting built a system called BargainFinder, widely considered the first comparison-shopping agent deployed on the web. It did something that seems trivial now: given a music CD the user wanted, it queried multiple online merchants for their prices and returned a comparison. The user clicked the cheapest one.
What happened next is the single most instructive event in the history of agentic commerce. A third of the merchants BargainFinder queried blocked its price requests. Their reasoning, documented by MIT Media Lab researchers at the time, was that they inherently did not want to compete on price alone. The value-added services merchants offered on their sites, the brand experience, the recommendations, the relationship, were invisible to an agent that reduced them to a number in a list. Rather than be commoditized, they opted out.
The academic response to this tension came primarily from Pattie Maes’s Software Agents group at the MIT Media Lab. Their 1996 system Kasbah was a genuine agent marketplace: users created buying or selling agents, gave them price parameters and negotiation strategies, and set them loose to transact with each other autonomously. A successor system, Tete-a-Tete, tried to solve the BargainFinder problem directly by letting agents negotiate across multiple terms, warranties, delivery, service, reputation, rather than price alone, precisely so that merchants could differentiate on something other than being cheapest. The University of Michigan built AuctionBot. Excite acquired the shopping agent Jango. A research community formed around what was called agent-mediated electronic commerce, complete with its own conferences and a canonical survey by Guttman, Moukas, and Maes in the Knowledge Engineering Review in 1998, which mapped agent capabilities onto the stages of consumer buying behavior: need identification, product brokering, merchant brokering, negotiation, purchase, and evaluation.
The field even developed the deeper infrastructure thinking that is being redone today. Researchers formalized electronic institutions, rule-governed environments where autonomous agents could transact safely, and agent communication languages like KQML, proposed in 1994, standardized how agents would talk to each other. If that sounds like a 1990s version of today’s agent protocols, that is exactly what it was.
Then the whole thing stalled. The reasons were straightforward. The agents were not intelligent: they followed hand-coded rules and simple price-decay negotiation heuristics, with no ability to understand context, preferences expressed in natural language, or product qualities beyond structured attributes. The web was not machine-readable: there were no standard product feeds, no APIs, and merchants had active incentives to block scraping. And consumers had no reason to trust an automated system with their money when e-commerce itself was still a novelty. The vision required a technology that did not exist. What survived was the degenerate form: comparison-shopping sites like mySimon and PriceScan, and eventually Google Shopping, which kept the price-comparison function and abandoned the autonomy.
The Second Attempt: Why Now
The current wave is best understood as the same vision meeting three preconditions that finally exist, arriving in rapid succession between late 2024 and today.
The first is agent capability. Large language models resolved the core failure of the 1990s agents: they can understand preferences expressed conversationally, reason about trade-offs across unstructured product information, and handle the ambiguity of real shopping tasks. The academic benchmark literature that has sprung up in the last eighteen months, systems like WebMall for evaluating cross-shop agent performance and AgenticShop for personalized product curation, exists because the underlying capability finally warrants serious measurement.
The second is interoperability infrastructure, and here the timeline is remarkably compressed. Anthropic released the Model Context Protocol in November 2024, standardizing how AI models connect to tools and data sources. Google followed with Agent2Agent (A2A) in April 2025, standardizing how agents discover and delegate to each other. Google then launched the Agent Payments Protocol (AP2) in September 2025 with more than 60 partners including Mastercard, PayPal, and Adyen, adding the payments layer. Thirteen days later, on September 29, 2025, OpenAI and Stripe launched the Agentic Commerce Protocol (ACP) alongside ChatGPT’s Instant Checkout feature, with Etsy as the first live marketplace and Shopify brands including Glossier, Vuori, Spanx, and SKIMS following shortly after. In under a year, the industry built the machine-readable, standardized transactional layer whose absence killed the first attempt. Part 3 of this series examines these protocols, and the emerging war between them, in detail.
The third is consumer readiness. A generation of shoppers has spent a decade delegating discovery to recommendation algorithms, trusting one-click checkout flows, and speaking to voice assistants. The behavioral leap from “the algorithm suggests, I click” to “the agent selects, I approve” is real but far smaller than the leap the 1990s asked consumers to make. Morgan Stanley projects that nearly half of online shoppers will use AI shopping agents by 2030, accounting for roughly a quarter of their online spending.
What Actually Changes: Agentic Commerce vs. E-Commerce
It is tempting to describe agentic commerce as e-commerce with a chatbot in front of it. The academic work published over the past year suggests the difference is much more structural, and it helps to be precise about where the structure actually shifts.
The unit of competition changes. Traditional e-commerce competes for human attention: search rankings, ad placements, page design, urgency banners, reviews engineered for social proof. All of it assumes eyes on a screen and a human deciding. When an agent shops, none of that surface exists. The most rigorous study to date, “What Is Your AI Agent Buying?” by Allouah, Besbes, Figueroa, Kanoria, and Kumar, published in the Proceedings of the ACM Web Conference 2026, built a sandbox environment to study how AI agents actually make purchase decisions, and found that agents exhibit their own systematic patterns: sensitivity to how product information is structured, position and framing effects that differ from human biases, and meaningful variation across underlying models. The same shopping task delegated to agents built on different LLMs can produce different purchases. Competition shifts from persuading humans to being legible and attractive to machine evaluators, whose decision processes are partially opaque and model-dependent.
The funnel collapses. The classic e-commerce funnel, awareness, consideration, conversion, retention, is a human psychological journey that brands spend enormous sums shaping. An agent given the task “buy me the best wireless headphones under 200 francs for running” compresses that entire funnel into a single automated evaluation. There is no awareness stage to advertise into, no consideration stage to retarget. Research on LLM consumer behavior theory argues this requires a genuinely new theoretical framework, because the assumptions of consumer psychology, attention scarcity, emotional appeal, and cognitive bias, apply to the delegating human but not to the executing agent, and the two are now separated.
Search costs approach zero, asymmetrically. Economists have understood since the 1990s that comparison agents reduce search costs. What is new is the asymmetry: an agent can evaluate hundreds of options in seconds at effectively zero marginal cost to the consumer, while the merchant’s cost of being evaluated, structured data, feed maintenance, protocol integration, is real and ongoing. The 1990s merchants who blocked BargainFinder were making a rational calculation that visibility to agents was net-negative for them. Today’s merchants face the same calculation with far higher stakes, which is the subject of Part 2.
The transaction itself becomes delegated. This is the deepest change and the one with no real precedent. In traditional e-commerce, the human is present at the moment of payment: reviewing the cart, entering credentials, clicking buy. In agentic commerce, the human expresses intent once and the agent executes, possibly later, possibly repeatedly, possibly negotiating along the way. That delegation gap, between what you meant and what the agent does with your money, creates the trust, verification, and liability questions that an entire branch of the new research literature is now working on, from verify-then-pay escrow architectures to formal models of agent-to-agent negotiation fairness. Those questions are the heart of Part 3.
The Lesson the History Teaches
The BargainFinder episode of 1995 established a pattern worth carrying through the rest of this series. When agents make markets more transparent, someone’s margin is being made transparent. The merchants who blocked the first shopping agent were not being irrational or backward. They correctly saw that a technology framed as consumer empowerment was also a redistribution of market power, away from those who differentiate through experience and relationship, toward whoever controls the comparison layer.
The MIT researchers concluded at the time that merchants with a web presence would eventually be forced into what they called “aggressive interoperability”: if you publish a catalog, agents will access it whether you like it or not, and your only real choice is how to differentiate within an agent-mediated market rather than whether to participate in one. That prediction was thirty years early. It is now the operating reality for every brand and retailer on the planet, and what participation actually requires is where Part 2 picks up.
Key sources:
Chavez, A. and Maes, P. “Kasbah: An Agent Marketplace for Buying and Selling Goods.” 1996.
“LLM Consumer Behavior Theory: Foundations of a Novel Research Field.” arXiv:2606.18005, 2026.
OpenAI. “Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol.” September 29, 2025.
Google Cloud. “Announcing Agent Payments Protocol (AP2).” September 16, 2025.