A Catalog Is a Contract with the Machine
A catalog was built for browsing. An agent approaches it as a set of claims that must be compared, checked, and acted upon. That change gives the ordinary product page a second responsibility: it must explain an offer to a machine that may search, negotiate, purchase, schedule, or hand the decision back to a person.
The public signals are now concrete enough to study. The W3C announced a September 8–9 workshop with GS1 on e-commerce for humans and AI agents, explicitly focusing on the experience of creating content with agents in mind. OpenAI’s July 30 RSS item about avatarin describes a GPT-Realtime retail agent giving shoppers multilingual support around the clock. Google’s A2A work and its ADK Go 2.0 runtime describe the coordination layer through which agents can hand work to one another and move through graph-based, human-approved workflows. These sources describe different layers, yet they touch the same commercial object: the information surface that lets a machine understand what can be bought, under which terms, and with what evidence.
The next storefront is a document that can state its own identity, conditions, and limits before an agent recommends it.
The page now has a second reader
Human shoppers forgive ambiguity because they can ask a seller. They can infer that a photograph shows the larger package, notice that a delivery promise excludes their region, or understand that “traditional” describes a method rather than a certification. A machine cannot safely fill every gap with common sense. It needs the merchant’s claims in forms that can be retrieved, compared, cited, and tested.
This makes product data part of the execution path. An agent looking for a replacement component must distinguish compatible dimensions from marketing language. An assistant buying food must resolve unit size, ingredients, stock, delivery area, substitution rules, and payment conditions. An institution procuring equipment must compare warranty, maintenance, tax, delivery, and authorization requirements. The catalog becomes the first policy boundary of the transaction.
The implication is architectural. A commerce system now has at least two surfaces: the visual interface that persuades and assists a person, and the structured interface that lets a machine form a bounded proposition. The second surface cannot be treated as a thin metadata layer added after the page is designed. It carries the fields on which an agent decides whether to continue.
What the machine needs to know before it can sell
A useful agent-readable catalog carries more than a product name and a price. It carries the conditions that allow a recommendation to become an accountable action:
- Identity: who publishes the offer, who fulfills it, which brand or producer is responsible, and how a buyer can distinguish an authorized seller from an imitator.
- Object: what is being sold, in which unit, with which dimensions, ingredients, versions, compatibility limits, and meaningful variations.
- Availability: where the object can be delivered or collected, when stock was last checked, what quantity is available, and what happens when the stated supply changes.
- Terms: price, currency, tax, shipping, returns, warranty, subscription conditions, payment methods, and the point at which a quotation expires.
- Permission: which actions an agent may take automatically, which require confirmation, and which must be handed to a named person or department.
- Evidence: the source of each material claim, the time it was updated, the applicable certification or test, and the route for correcting an error.
These fields describe a contract in the practical sense: a bounded statement of what one party is offering and what another party may rely upon. The contract can remain human-readable while also being expressed in structured data. What matters is that a machine can tell the difference between a fact, an estimate, a promotion, a condition, and an unanswered question.
Commerce agents will expose the cost of missing fields quickly. When a catalog omits delivery territory, the agent either guesses or stops. When a price lacks currency or validity, the agent cannot compare it honestly. When the return policy is hidden in an image, the buyer’s risk has been transferred into an opaque interpretation step. The quality of agent commerce will therefore depend less on fluent conversation than on the integrity of the underlying offer record.
A handoff is not a sale
Google’s A2A and ADK Go 2.0 material clarifies another part of the problem. A commercial agent rarely completes every step alone. One agent may discover products, another may verify eligibility, a payment service may authorize the charge, a logistics system may schedule delivery, and a human may approve a high-value purchase. The path is a graph of responsibilities.
That graph needs a common account of the object moving through it. If the discovery agent passes a product name without its version, the verification agent may inspect the wrong thing. If the payment step loses the currency or expiry time, the buyer may approve a different amount. If the logistics step receives an address without delivery constraints, a successful API call can still produce an impossible promise. Handoff protocols are useful when they preserve the meaning and authority of the transaction, not only when they pass messages.
The merchant’s catalog is therefore part of the protocol. It gives each agent a stable object to carry, a source to cite, and conditions to preserve. The product identifier, offer identifier, seller identity, price, currency, availability timestamp, delivery promise, and approval state should survive the handoff. An agent that cannot explain which version of an offer it used cannot provide a reliable receipt afterward.
This is where the distinction between recommendation and execution becomes important. A recommendation can tolerate an explicit uncertainty if a person evaluates it. An automated purchase requires a permission boundary and a record of the claims that justified the action. The catalog supplies the initial evidence; the transaction system must retain it as the decision moves across services.
Provenance belongs in the merchandise
The provenance question is usually associated with photographs, video, and written media. Commerce brings it into the object being bought. A customer may need to know who made a product, where it was assembled, which materials it contains, whether a certificate applies to this batch, or whether a repair part is an authorized replacement. Those are provenance claims with commercial consequences.
The W3C’s workshop with GS1 is significant at the standards level because it treats e-commerce content as a subject for joint work between human-facing and agent-facing systems. The aim is to make the offer legible across the systems that discover, compare, authorize, and fulfill it, while keeping the storefront readable for people.
For an African merchant, provenance also carries place and language. A cooperative may need to identify a producer group, a harvest period, a processing method, and a settlement route. A publisher may need to distinguish a licensed digital edition from an unauthorized scan. A craft business may need to attach a maker, material, region, and repair instruction in more than one language. If the platform’s default categories cannot express those distinctions, the machine will treat the omission as neutrality. It will not be neutral; it will be a loss of economic and cultural information.
Local ownership of product data is therefore a sovereignty issue. The merchant should be able to export the offer record, correct it, expose it through more than one channel, and decide which agent may act on it. A foreign marketplace can provide distribution without owning the definition of the object. The boundary must be explicit.
African Commerce Has Its Own Market Grammar
A catalog designed for agent commerce in Africa must account for conditions that are often treated as edge cases elsewhere: mobile money, cash-on-delivery, intermittent connectivity, regional delivery zones, multilingual search, informal business names, tax differences, shared phone numbers, and the need to confirm a transaction through a person. These are not awkward exceptions to be cleaned away. They are part of the market’s actual grammar.
An agent that searches only standardized English product names will miss goods whose economic identity is carried by a local term. An agent that assumes one stable address format will misroute a delivery. An agent that treats card payment as the default will exclude a buyer who can pay through a mobile wallet or an assisted merchant. A local catalog layer can encode these conditions while still exposing interoperable product identifiers to global discovery systems.
This is one route toward digital sovereignty that does not require isolation. African builders can own the vocabulary, seller identity, payment rules, delivery logic, and correction process while allowing external agents to discover and request the offer. The objective is a market in which local institutions are legible to machines without having to surrender the terms of legibility to a distant platform.
The work begins with a disciplined inventory:
- Define the product and offer identifiers that remain stable when a marketplace, model, or payment provider changes.
- Record prices, currencies, availability, delivery constraints, returns, warranties, and permissions as explicit fields rather than prose hidden in a page.
- Attach local names, translations, producer identities, provenance claims, and correction routes to the same offer record.
- Test the record through more than one agent and more than one payment or fulfillment path.
- Preserve a receipt that shows which claims and permissions were active when the transaction was authorized.
Where the investable surface is widening
If agents become participants in commerce, the capital-relevant infrastructure sits between the storefront and the model:
- Agent-readable catalog platforms: systems that let merchants publish structured products, offers, conditions, and permissions across storefronts and agent channels.
- Offer identity and provenance: registries that bind a product, seller, batch, certification, and source record to a stable object an agent can carry through a transaction.
- Commerce policy engines: controls that decide when an agent may recommend, reserve, purchase, refund, or escalate, with thresholds tied to price, risk, identity, and jurisdiction.
- Multilingual and local-market data layers: terminology, translation memory, seller resolution, address interpretation, mobile-money support, and offline synchronization for markets that do not fit a single global template.
- Machine-readable receipts: records that preserve the offer version, evidence, permissions, approvals, payment state, and fulfillment commitments behind an agent-mediated purchase.
The underwriting question is specific: does the product reduce the uncertainty between a machine’s recommendation and a transaction that a buyer, seller, and institution can defend? Evidence might include fewer failed handoffs, lower correction rates, better conversion from discovery to authorized purchase, faster reconciliation, clearer returns, higher coverage of local-language searches, and a receipt that a human can audit without reconstructing the entire conversation.
A catalog platform earns infrastructure status when it becomes portable across agents and marketplaces while preserving the merchant’s own definitions. A polished interface may attract attention, but the durable value sits in the record that survives a new model, a new channel, a new payment route, or a dispute.
Write the offer so it can be inherited
The machine-readable catalog is a new publishing discipline. It asks the merchant to state what the product is, who stands behind it, where it can move, what it costs, which evidence supports the claim, and which actions require consent. It asks the platform to carry those statements through discovery, handoff, authorization, and fulfillment without turning ambiguity into false certainty.
For African institutions, this discipline can connect commerce to sovereignty. Product language, seller identity, payment practice, and local provenance become operating assets that can travel across providers without disappearing into a platform’s private vocabulary. The market remains open to global intelligence while the terms of local participation remain owned by the people who produce and exchange value.
The storefront of the agent economy will be built from records that can be read twice: once by a person deciding whether to trust the offer, and once by a machine deciding what it is permitted to do. A catalog becomes a contract when both readers can see the same object, conditions, and evidence.
Sources
- W3C News: “Upcoming: W3C/GS1 Workshop on E-commerce for Humans and AI Agents” (published June 1, 2026; workshop scheduled for September 8–9, 2026; description: sharing experience of creating content with AI agents in mind, with e-commerce as a focus).
- OpenAI News RSS: “How avatarin built a 24/7 retail agent with GPT-Realtime” (July 30, 2026; linked article: https://openai.com/index/avatarin; description: avatarin uses GPT-Realtime to provide multilingual support for Yamada Denki shoppers).
- Google Developers Blog: “How A2A is Building a World of Collaborative Agents” (June 18, 2026; description: secure agent handoffs and scalable collaborative workflows for a shared agent ecosystem).
- Google Developers Blog: “Build reliable multi-agent applications with ADK Go 2.0” (June 30, 2026; description: graph-based workflows, human-in-the-loop orchestration, dynamic routing, and built-in resilience).