Meta Built the Ad Machine. OpenAI Built the Agent. Who Controls the Agentic Marketplace?

Meta and OpenAI are competing for the position between a customer’s request and the businesses that can fulfil it. Meta approaches with an advertising system, merchant relationships, messaging and consumer devices. OpenAI approaches with an assistant people already use to explain problems and compare options. Each is building towards more of the other’s territory.

Meta Connect 2026 made that contest easier to see. Its personal agent, Muse, brings a continuing customer task together with connected services, commerce partners and planned access through glasses. The important commercial development is the attempt to coordinate the whole purchase, from understanding the need to remembering the outcome.

If people entrust that work to an agent, the platform around it may influence which merchants get considered, how a purchase happens and which service gets used next. A retailer could gain a new source of customers while losing some control over the interface through which those customers choose.

That is the marketplace question behind Connect. Who holds the relationship with the person, who supplies the useful capabilities, and who can set the terms on which the two meet?

Start with the whole commercial system

An agent-mediated purchase combines several responsibilities that a familiar shopping interface can make look like one. A customer states requirements. An agent interprets them. Merchants supply products and terms. Access controls establish which operations are permitted. Payment and order systems record the transaction. Fulfilment and support determine whether the promise was kept.

Customer goal and constraints
    → agent interprets the task and finds available services
    → merchant systems supply product and delivery evidence
    → customer chooses and grants any required permission
    → an authorised service prepares and executes the purchase
    → order, payment and fulfilment records establish what happened
    → service outcomes inform the next decision, where permitted

Each transition can fail independently. The agent may misunderstand a requirement, the merchant may expose stale stock, or a payment may succeed before the agent receives confirmation. A polished conversation can conceal any of those failures.

Take carry-on luggage as a buying task. A useful comparison needs the relevant airline’s cabin limit, the bag’s external dimensions, a delivered price and a credible arrival date. “Suitable for travel” is too vague. Preparing a shortlist needs evidence; placing an order also needs permission. A recommendation that sounds confident establishes neither.

The business relationships are equally distinct:

Participant	Main responsibility	Commercial interest
Customer and personal agent	Define the task, compare options and control delegation	A suitable outcome with acceptable effort and risk
Merchant and its business agent	Describe the offer, support the customer and fulfil commitments	A valuable sale and continuing relationship
Agent platform	Provide discovery, context, connections and execution surfaces	Adoption, recurring use and opportunities to earn revenue
Commerce and payment providers	Maintain transaction records and supported purchase operations	Participation in completed, supportable transactions

A buyer’s agent and a merchant’s agent can communicate while serving different people or businesses. Agreement between their outputs is not enough to prove that the purchase serves the customer. That distinction becomes particularly important when one platform supplies both sides.

What Meta is assembling

Meta’s Connect announcement links Muse with retailer connectors, Shopify’s catalogue, additional payment partners and access through AI glasses. It describes Muse on glasses as coming in the following months, alongside the announced Muse Charm device. These are staged product developments, rather than evidence that every part already works together for every customer.

The product details below reflect documentation checked on 24 October 2026. Availability belongs to the particular agent, connector, device and market; one component’s launch does not establish another’s availability.

A personal agent with continuing work

Meta introduced Muse on 8 September. It describes a dedicated cloud computer and browser, connected apps, retained context and work that can continue between conversations. Users choose access and review sensitive actions. The launch post says conversations and data on that computer are not shared with Meta’s ad systems.

The original launch was US-focused. The later Muse for Small Business announcement, updated in October, describes Muse as available in the US and Canada. It adds connections to business tools, Shopify and Meta business accounts, with approval before publishing, sending or spending.

That adds a third role to the system. A personal agent works on a buyer’s task. A customer-facing business agent represents a merchant. An agent helping an owner analyse campaigns or run a shop supports the merchant’s internal operations. The same product family can participate in all three roles without making their permissions or incentives identical.

Services the agent can reach

The developer recap connects the consumer story to models, Muse Code, the Meta Model API and the Muse Connector Platform. Those components help developers build and expose services; they do not demonstrate that a particular shopping task is reliable or commercially successful.

There are also different ways for software to reach Meta’s interfaces. The AI glasses developer guide distinguishes mobile-app integration through the Wearables Device Access Toolkit, web apps on Meta Ray-Ban Display, and Meta AI Connectors using an API or MCP server.

MCP, the Model Context Protocol, exposes defined tools to an agent. WebMCP lets a web app declare operations an agent can use within the supported browser environment. Meta’s guide labels both WebMCP for glasses web apps and Meta AI Connectors as developer previews. It also describes future discovery surfaces and an approval process still being developed. A supported API, a preview integration and a published consumer experience are different achievements.

For a luggage merchant, useful operations could include finding a variant, checking stock and preparing a cart. Each should preserve the merchant’s pricing, access and order rules across clients. A tool description helps the agent understand an operation; the server must enforce what it is allowed to do.

Agents on the merchant’s side

Meta Business Agent is the customer-facing part of this picture. Meta describes product recommendations, lead qualification, appointments, sales and human handoff across its business-messaging surfaces. Its platform connects those agents to business systems, with rollout and feature scope varying by product.

That means a personal agent may eventually encounter another agent representing the seller. The buyer wants a suitable bag before a departure date; the seller wants to convert the enquiry. The seller’s assistant should provide verifiable dimensions and delivery terms, while the buyer’s assistant should preserve the person’s constraints. Neither should be able to manufacture the other’s approval.

OpenAI approaches from the other direction

ChatGPT starts close to the moment someone describes a problem. A conversation can contain the budget, trade-offs and exceptions that a short search query leaves out. OpenAI is adding infrastructure around that expressed need.

Its current Agentic Commerce Protocol documentation describes structured merchant catalogues, inventory and contextual product discovery. Alongside that discovery route, OpenAI has added self-serve advertising, bidding and conversion measurement. The ChatGPT Ads playbook follows the buying mechanics and their rollout conditions.

The September announcement also introduced a US test of Sponsored Agents reached through ads. That adds a paid entry to an agent representing a business, distinct from an independent product result or a checkout integration. Together, these developments make OpenAI’s commercial infrastructure more substantial without establishing equivalent scale, performance or advertiser control to Meta’s.

Starting position	Meta’s route	OpenAI’s route
Existing customer encounter	Social discovery, messaging and advertising	Conversational assistance and stated tasks
Commercial infrastructure	Merchant relationships, business messaging and performance ads	Product feeds, commerce integrations and a developing ad system
Expansion	Add a personal agent before merchant selection	Add merchant services and advertising around the conversation
Continuing relationship	Connected accounts, retained context and device access	Connected capabilities, persistent work and shared contexts

This is an interpretation of their strategies, not a claim that either company has a monopoly on a category. Gemini, Amazon’s assistants, Perplexity and other systems also compete for customer tasks. Meta’s combination of advertising, messaging, merchants and hardware makes it a substantial challenger; it does not prove that it will become the default agent.

Our DevDay analysis follows how OpenAI connects capability discovery to use and adoption. Here, the question is what happens when competing platforms assemble those relationships at scale.

Keep advertising and the agent’s private context distinct

OpenAI’s ads guidance says paid placements run separately from answers and advertisers do not receive private conversations. Depending on settings, matching may use the current thread and, for personalised ads, other chats and memory. That is a platform’s use of context, not permission for an advertiser to inspect it.

Meta’s stated separation between Muse data and its ad systems is another specific boundary. The business-account connections in Muse for Small Business do not establish that consumers’ private shopping requests feed Ads Manager. Access an agent has been granted to help a business operate is different from using a buyer’s memory for advertising.

These are provider commitments and product descriptions, not an independent audit of every implementation. The cited Muse and Connect material also does not establish a paid-recommendation product inside Muse. Future sponsored selection would require its own account of disclosure, user control and conflicts of interest.

The tension is structural. A personal agent becomes useful because it understands a person’s circumstances. That understanding can also create commercial influence. Trust depends partly on what the platform refuses to do with it.

How a useful agent could become a marketplace

A marketplace needs more than a checkout button. Its participants must have reasons to join and reasons to stay.

Economists call one important mechanism a cross-side network effect: participation on one side makes the service more valuable to another. More customers can attract merchants and developers; a wider set of dependable services can make the agent more useful to customers. Rochet and Tirole’s work on platform competition explains why getting interdependent groups on board matters to the platform’s business model.

The possible reinforcing loop is:

Useful completed tasks attract customers
    → customer demand attracts merchants and developers
    → more dependable services expand the tasks the agent can complete
    → continued use creates more reasons to participate

Advertising, payments, subscriptions or business software could help fund that system. A platform may subsidise the customer interface because revenue can arise elsewhere. A free agent therefore need not imply a business without a commercial model.

The loop is a hypothesis to test. More connectors can also bring unreliable inventory, inconsistent policies, fraud and additional work for users. A large catalogue improves choice only when the agent can find suitable offers and the merchant can fulfil them. Connection counts and app downloads cannot establish a durable network effect by themselves.

Merchants may use several platforms while customers favour one

A retailer can expose its catalogue through more than one agent platform. This is multi-homing, meaning participation in several competing networks. Shopify’s distribution role makes that a plausible merchant strategy, although maintaining consistent feeds, policies and integrations across channels still costs work.

Customers can also use several assistants. Yet one may become primary for recurring tasks because it holds useful context, connected accounts and a history of what worked. Call that a possible move towards single-homing: alternatives remain available, but the working relationship concentrates in one place.

When one group joins several platforms and the other mainly uses one, each platform can control access to its own customers. Armstrong calls this a competitive bottleneck. Applied here, merchants might need to participate in multiple agents to reach the customers who prefer each one, even when none of the merchants is exclusive.

That is a different source of power from owning unique inventory. The scarce asset would be access to a particular customer’s delegated task. Whether this pattern emerges depends on user behaviour, interoperability, service quality and switching costs. It is not established by announcing an agent.

Memory can make the relationship harder to move

An agent’s accumulated context could include sizes, budgets, preferred brands, delivery needs, loyalty accounts, returns, standing tasks and the reasons past recommendations succeeded or failed. Changing assistants might then require rebuilding more than a list of preferences.

Farrell and Klemperer’s analysis of switching costs and network effects explains how earlier investments can give a provider power over later choices, and why firms may compete aggressively to acquire the relationship first. The application to agents is a proposition: reconstructing trusted context and connections could make leaving costly even without formal exclusivity.

Defaults add another mechanism. Samuelson and Zeckhauser’s research on status quo bias shows that people can favour an existing choice. An assistant reached through habitual messaging or a familiar device may benefit from that tendency, though the research does not measure these products’ adoption.

Portability can reduce some costs, but exporting chat text is only part of it. Useful transfer would preserve the meaning, freshness and uncertainty of retained requirements and lessons, the record discipline developed in Memory & Agency. Account access and purchase approvals need their own secure reconnection; copying credentials or assuming another platform inherits an old permission would create a different problem.

The result need not be one global winner. Several agents could remain useful for different tasks. The question is whether customers can change the primary agent without losing so much context or capability that the choice becomes nominal.

Amazon and Shopify illustrate different incentives

On 20 September, GeekWire reported Amazon’s decision to block Muse, including Amazon’s concerns about unauthorised access, identification and customer security. Axios also reported the block on 21 September. This is a dated access dispute, not confirmation that the same restrictions remain unchanged today.

A customer’s willingness to delegate shopping does not by itself settle a merchant platform’s access rules. Identity, supported operations, responsibility for errors and recovery all need agreement. The security concern has substance even when the platform also has commercial interests to protect.

Amazon combines a destination, marketplace, recommendations, advertising, checkout and fulfilment. If an external agent controls comparison and the next recommendation, Amazon could retain the order while losing influence over how the customer arrived at it. That is my interpretation of the commercial tension, rather than a claim about Amazon’s undisclosed motive.

Shopify occupies a different position. It supplies infrastructure to merchants who benefit from being found across several customer interfaces. An agent can control the conversation while Shopify remains involved in the catalogue, checkout, order and merchant operations. Extending distribution can therefore reinforce Shopify’s role.

Its current Meta channel documentation makes that division concrete. Catalogue access and direct checkout are separate settings. Eligible stores can have direct checkout enabled by default; disabling it can send customers to the online store instead. Removing catalogue access does not guarantee disappearance from other discovery methods.

Shopify lists customer-market eligibility and excludes subscriptions, bundles, customisable and B2B-only products, plus some delivery methods, from Meta direct checkout. It supports certain checkout functions and discounts, so an off-site checkout is neither identical to the merchant’s website nor devoid of merchant rules. The merchant remains responsible for fulfilment, returns and service. Orders receive Meta channel attribution; standard client-side pixels and Google Analytics do not fire in that direct checkout, while supported server-to-server events remain.

These arrangements distinguish permissioned participation, where services agree how agents may interact, from defended destinations, where a platform restricts unsupported access. The choice is more specific than being for or against agents. It determines which agent can reach which service, under what identity, with what commercial terms and evidence.

Marketing operates across the whole decision

Agentic commerce changes where influence can occur without making human preference irrelevant. A campaign can make the customer ask for a particular brand before delegating any work. Product evidence helps the agent decide whether that brand meets the request. Delivery and support establish whether the preference was rewarded.

For the carry-on task, the relevant evidence spans the journey:

Stage	What the business needs to establish	What remains unknown without further evidence
Eligibility	Correct product, dimensions, stock, market and delivery constraints	Whether the platform considered every eligible offer
Comparison and recommendation	Accurate claims and reasons for the proposed choice	The full ranking process or every rejected alternative
Permission and checkout	Current approval, correct variant and total, recorded order state	Whether a click or a fluent answer represents a completed purchase
Fulfilment and service	Delivery, refunds, returns and resolved problems	Whether the customer experienced the promised benefit
Later selection	Permitted evidence of satisfaction and repeat choice	Whether retention reflects preference, convenience or costly switching

Creative, product data and operations perform different jobs. A compelling story can create preference, but it cannot correct a false availability promise. A complete catalogue can support comparison, but it cannot explain why a person should care about a brand before the task begins.

Measurement needs the same distinctions. A merchant may observe orders, returns and service records without seeing the platform’s entire consideration process. A missing landing-page view does not imply no commercial influence, and an attributed order does not explain every earlier cause. Keep unknown host-side events unknown; use an appropriate controlled experiment before claiming incremental sales or retained customers.

This is where Who Owns the Pixel? connects to the marketplace argument. Control over the record of a sale and control over the explanation of that sale are separate forms of influence.

Reliability and permission determine usable supply

For the carry-on purchase, reliability means finding a bag that meets the airline’s dimensions and delivery deadline, while refusing an order that lacks current authority. The Agent Architect’s Playbook defines the shared evaluation, validation and policy responsibilities. Their marketplace value depends on whether connected services actually preserve those requirements.

Meta’s Muse safety architecture describes an isolated execution environment, protected credentials and a separate Sentinel controlling connector actions and network traffic. It presents permissions as scoped grants and acknowledges that the agent can still make mistakes. That is a useful architectural distinction, not a guarantee of safety or proof that another implementation has the same controls.

For a merchant, the corresponding requirements remain concrete. Recheck the account, chosen variant, stock and total at execution time. Bind approval to the actual operation. Make repeated requests safe so a lost response does not become a duplicate order. If payment or order status is uncertain, reconcile the record before trying again.

Test the carry-on task with stale stock, missing dimensions, incompatible delivery dates, expired access, a changed price, a hostile instruction in product text and a failure after payment. A useful result may be a clarified question or a refused purchase. The benchmark should reveal those outcomes rather than reward every apparent completion.

The commerce readiness stack and secure-agent architecture develop the implementation details. The marketplace consequence is that dependable, permissioned supply may matter more than the number of available connections.

Give each team a concrete part of the preparation

The first useful step is a joint audit of one buying task, ending with a record that product, commerce, marketing and engineering can all inspect.

Commerce teams should produce a channel and product record. Compare variants, prices, stock, delivery, returns and disclosures across the website, catalogue and agent output. Record catalogue and checkout settings separately, the supported markets, excluded product types and the path for unsupported purchases. Include post-purchase service and how the order will be identified when a customer needs help.

Advertising and measurement teams should produce an evidence map. Separate brand preference, observed recommendation tests, paid placements, approved purchases and fulfilled orders. Identify which events belong to the merchant and which require platform reporting. Check where pixels disappear and which supported server-side records remain. Judge experiments on incremental business outcomes and customer benefit, with the assumptions visible.

Product and engineering teams should produce a bounded action contract. Expose product and policy reads before purchase actions. Define identity, account scope, required evidence, confirmation, retry behavior and recovery. Test the contract across supported clients rather than burying merchant rules in one platform’s prompt. Monitor failures, retain necessary audit evidence, set a way to disable faulty actions and re-evaluate after meaningful changes.

Agencies and commercial leaders should produce a dependency record. Name who controls discovery, identity, permissions, payment, attribution, service and the next interaction. Audit catalogue quality, recommendation behavior and checkout together. Assign owners for connector changes and defaults that can alter a channel without a website release. Preserve brand knowledge and independent business records so the platform is a route to customers rather than the only place the business understands them.

Preparation does not require predicting the winning agent. It requires knowing which commitments survive when a customer reaches the business through a different interface.

Watch the terms of participation

The decisive signals will concern control as much as capability: whether users can move useful memory and standing tasks; whether merchants can inspect and challenge exclusion; whether future paid recommendations disclose their influence; and whether an access dispute can be resolved through supported, accountable participation.

Watch consumer use alongside merchant participation. If buyers routinely compare several assistants, the bottleneck may remain weak. If one agent becomes the trusted keeper of recurring tasks while merchants must appear everywhere, the balance of power changes. Both possibilities are consistent with today’s announcements.

Meta Connect’s larger commercial ambition is to bring the agent in front of the machinery Meta already operates. OpenAI is adding commercial machinery around the assistant. Their competition may produce a marketplace whose most valuable position is the continuing relationship with the person who delegates the task.

For a brand, the aim is to make its offer verifiable, its execution dependable and its value worth requesting again. Reach through an agent can help build that relationship. The terms on which the agent remembers, recommends and acts will determine how much of it the brand can keep.