Article · AI governance and project delivery
What Marconi and the Titanic teach us about agentic commerce: open connections matter, but a buyer’s agent must also answer to the buyer.
Agentic commerce needs both interoperability and buyer representation. Connecting an agent to a merchant does not, by itself, establish whose interests the agent serves. The history of Marconi’s wireless network helps explain why those are separate questions.
In March 1902, Prince Heinrich of Prussia was crossing the Atlantic aboard the German liner Deutschland. He had just visited the United States. He asked the ship’s wireless operator to send a message of thanks to President Theodore Roosevelt. The Marconi station at Nantucket refused to accept it. The ship carried German Slaby-Arco equipment, and Marconi stations were prohibited from handling messages from ships that did not lease Marconi equipment.
Marconi’s policy was commercially rational. A shore network, leased apparatus and trained operators made a powerful combination. Whoever controlled access to the connection could influence the customer’s choice of equipment. A rival could be excluded by a refusal to answer.
Germany responded to the Deutschland incident by convening a preliminary wireless conference in Berlin in 1903. Three years later the International Radiotelegraph Convention of 1906 established the principle of exchanging messages regardless of the radio system used. It also gave distress calls absolute priority. Britain signed the Convention but stayed out of the additional agreement on ship-to-ship intercommunication; Italy reserved ratification until its Marconi contracts expired. Interconnection was arriving, but the arrangements were incomplete.
The Titanic and the Californian were both fitted with Marconi radio. At about 11.00 pm, Californian ship’s time, its operator Cyril Evans called the Titanic to report that his ship had stopped in ice. Titanic’s operator was communicating with Cape Race. Evans’s much stronger signal interfered with that traffic, and he was told to keep out. Evans later testified that the traffic was passengers’ private messages. At about 11.30 pm he stopped his detector and went to bed. He had no assistant and no instruction to maintain an overnight watch. The Californian did not receive the Titanic’s subsequent distress calls. This was not a failure of compatibility between rival systems. Compatible equipment still needed the right operating priorities, staffing and authority.
The London Convention signed on 5 July 1912 addressed those operating arrangements. Its service regulations placed the shipboard radio service under the commanding officer’s authority and defined categories of listening duty, including continuous service for the first category. The lesson was not simply that systems must talk to each other. Someone also had to decide whose instructions the operator followed and whether anyone was listening.
The telephone offers another example. The Bell System’s interconnection policies became part of the antitrust challenge that led to the Kingsbury Commitment of 1913. The settlement did not create unrestricted competition, but it illustrates the recurring conflict between control of a network and access to the people who use it.
The principle is now familiar. A customer on one mobile network can ring someone on another. An iPhone can call an Android phone. Apple added RCS messaging support in iOS 18 in September 2024, improving messages to contacts without Apple devices. An email written on a Mac can arrive on a Windows machine. People expect the connection to work across suppliers.
These histories differ in their technology, regulation and commercial outcomes. What they share is a tension: a supplier may own important infrastructure without having an unlimited claim over whom its customers can reach.
Agentic commerce is commerce in which software agents help buyers and sellers discover products, compare offers or carry out transactions. In 2026 the industry has protocols for many parts of that process. Connecting them is one problem. Establishing whom the agent represents is another.
The Agentic Commerce Protocol connects agents with merchant checkout. Google’s Universal Commerce Protocol addresses commerce interactions, while AP2 addresses payment mandates. Visa’s Trusted Agent Protocol helps merchants recognise agents, and x402 supports internet-native payments. In October, Decagon and Instinct published the Personal Agent Consent & Trust Protocol, or PACT, for customer-authorised interactions with business agents. Sierra and Meta also announced the Personal Agent Protocol, an open standard for personal agents to connect with businesses.
These initiatives answer real questions about identity, consent, permissions and transactions. Several are explicitly open standards. PACT keeps the customer’s login with the business and requires the customer to grant permissions; the Personal Agent Protocol combines customer authorisation with parameters set by the business. That is progress towards interconnection. It does not, by itself, establish whether an agent’s recommendations consistently serve the buyer’s interests. A connection can be open while the commercial relationship around it remains concentrated.
Google’s Universal Cart shows the distinction. Its May announcement described a cart spanning Search and Gemini, with YouTube and Gmail to follow. Google said it would use Gemini and Google Wallet, with UCP supporting checkout. It also said shoppers could transfer items to the merchant’s site, and that the brand would remain the merchant of record. The question is therefore larger than where payment happens. If a shopper builds a relationship with an assistant across several services, who controls that accumulated understanding, and how easily can the shopper move it to another representative? An open commerce protocol does not, on its own, answer that question.
Amazon presents a different boundary. In September, SiliconANGLE reported that Amazon had blocked Meta’s Muse shopping agent. Amazon’s stated concerns included agent identification, retailer participation and access to customer accounts. Those concerns deserve to be addressed. They also expose the commercial question: how should a retailer serve a customer who chooses to arrive through a representative? Identification and permission are necessary parts of the answer. They should not become a substitute for considering the customer’s choice.
The analogy with Marconi is about that boundary. It does not mean every restriction is unreasonable, or that radio treaties establish a legal right for shopping agents to access a retailer. My argument is commercial: a business risks losing the customer when it cannot work with the representative the customer has chosen.
The platforms are fighting over the wrong thing, and the reason is a shift in what the agent is.
Sangeet Paul Choudary’s recent Harvard Business Review article examines how AI agents are changing customer relationships. For me, the important distinction is between attracting a customer to a platform and earning authority to act for that customer.
A buyer’s representative can compare offers across marketplaces. In that relationship, a marketplace becomes one source of supply among several. Its catalogue, delivery service and reputation can still matter enormously. What changes is its ability to assume that the customer will begin every decision inside its interface.
Being available inside an agent is therefore useful distribution. It does not establish ownership of the relationship. The agent still has to decide which offers to examine, how to compare them and which constraints to apply. Those decisions need to be accountable to the person whose money is being spent.
That is why protocol proliferation is only part of the story. A protocol can standardise authentication or checkout without resolving conflicts between the buyer’s preferences and a platform’s revenue model. The test is whether the shopper can choose a representative, carry their requirements across suppliers and understand why an offer was recommended.
The risk in disparate approaches is that each platform builds a convenient connection while keeping the decision inside a surface it controls. A buyer’s agent should be able to work across those surfaces, subject to the merchant’s legitimate access rules and the buyer’s authorisation.
VendeeX is built on the other side of this argument. It is a Buyer’s Agent in the strict sense: it works only for the shopper. In the current VendeeX Shopping App, the shopper’s identity, preferences, sizing and cryptographic keys stay on the shopper’s own device. Members do not create an account in the conventional sense, and Vendee never sees their personal data or email address. The agent consults the shopper, interprets the requirement, searches independent merchant sources, applies the shopper’s constraints and produces grounded recommendations. Ranking is separated from commercial metadata, so the order of recommendations is the same whether or not commission information is present. The current Shopping App is distinct from the wider version 2 platform design.
Two design decisions matter for the argument in this article.
The first is how the current Shopping App reaches merchants. It never holds the shopper’s merchant credentials and never acts inside a merchant’s account. It reaches merchants through published feeds and declared interfaces, rather than an automated session dressed up as a human browser. The shopper completes the purchase on the merchant’s own surface, under the merchant’s own authentication. A buyer’s representative should identify itself, respect the merchant’s access rules and remain within the buyer’s authority. Interconnection does not require ignoring a merchant’s refusal or treating access as unlimited.
The second is how the design handles different protocols. VendeeX’s version 2 architecture specifies a Universal Protocol Bridge: a common internal model with adapters for external protocols, including UCP, AP2 and MCP. This is an architectural design, not a claim that every adapter is deployed or every merchant is reachable. Its purpose is to make protocol differences manageable while preserving the buyer’s mandate. From the shopper’s side, the ambition resembles a telephone exchange: the system handles the connection so the person can concentrate on what they want to do.
That is the reversal I believe matters. In 1906, interconnection arrived through an agreement between states. In agentic commerce, buyer-controlled software can also help make interoperability useful in practice. Where a merchant permits access and an adapter is available, the buyer should not need to organise their shopping around the merchant’s choice of protocol. The question that remains is the one I would put first: is this agent working for me?
Marconi’s mistake, as I read it, was to confuse supplying the connection with owning the relationship between the people who needed it. A network can remain valuable after it becomes interoperable. What becomes harder to defend is an exclusive right to determine whom its users may contact.
Amazon’s catalogue, Google’s Shopping Graph and Walmart’s fulfilment network remain valuable. They are the transmitters in this analogy. A buyer’s agent that carries the shopper’s requirements across participating merchants changes how that value is reached. The opportunity is to earn the buyer’s business when their representative arrives.
History does not guarantee the outcome of agentic commerce. It gives us a useful question to ask of each new protocol and each new gate: does this help the customer connect, and does the agent answer to the customer? The merchant that answers both questions well has a reason to welcome the buyer’s representative.
VendeeX. Changing the way the world buys.
Agentic commerce uses software agents to help buyers or sellers discover products, compare offers or carry out transactions. The scope of an agent’s authority depends on the task and the permissions it receives.
It follows the buyer’s requirements and authority, makes conflicts of interest visible and can explain its recommendations. Authentication proves who is connecting; it does not by itself prove that recommendations serve the buyer.
A protocol bridge translates between a common internal model and external interfaces. It can reduce the burden of connecting to different systems, where an adapter is implemented and the merchant permits access. It cannot create permission or guarantee universal coverage.
From the library
A guide for leaders and practitioners to governing AI through principles, ethics and structured assessments: the history and legal setting of AI, eight guiding principles with ways to uphold each, and assessment approaches for organisations.
A workbook that turns AI governance from a compliance burden into a delivery discipline: mapping an organisation against an AI technical standard, an implementation methodology, procurement and contract governance, and oversight across the AI lifecycle.
Cybersecurity explained for boards and executives: how the threat landscape evolved, what compliance requires, how threats differ from vulnerabilities, and what an informed executive response looks like.