The Architecture of Agent-to-Agent Commerce: A Field Map

Most of the public conversation about "AI agents" treats agents as a monolithic thing: a single product category to be evaluated as if Phia, Skyfire, OpenAI Operator, and Sierra were all competing in one market. They are not. Agent-to-agent commerce is being assembled across five distinct architectural layers, each with its own buyers, its own incumbents, its own protocols, and its own competitive dynamics. Conflating them produces strategy errors at every level: investors back the wrong thesis, builders spend on the wrong moat, and incumbents miss the layer that will actually disrupt them.

This paper is a field map. It is not a product directory. The aim is to give operators, investors, and policymakers a stable taxonomy for the next decade of customer engagement infrastructure, so that conversations about "agentic commerce" can be grounded in which layer is actually being discussed.

The five layers

The agent-to-agent commerce stack, in our framing, is composed of five interdependent but separable layers. From bottom to top:

Layer 1. Protocols and standards. The wire format and shared vocabulary that lets two agents interoperate at all. Examples already shipping: Anthropic's Model Context Protocol (MCP), Google's Agent2Agent (A2A) protocol, OpenAI's Agentic Commerce Protocol, Klarna's Agentic Product Protocol, Visa's Trusted Agent Protocol (TAP), Mastercard's Agent Pay. This layer is being defined by incumbents and standards bodies, not by venture-backed startups. It is not a competitive market in the conventional sense; it is a coordination problem.

Layer 2. Identity and trust. The infrastructure that lets one agent prove who it represents, what it is allowed to do, and what it has done. KYA (Know Your Agent) is the equivalent of KYC for the agent layer. Funded companies in this layer include Kite (agent identity and settlement), Skyfire (KYA), Nava (financial-agent guardrails), and Mount (agent liability insurance). This is the trust substrate: without it, the higher layers cannot transact.

Layer 3. Payments and rails. The settlement infrastructure that moves money when agents agree on a transaction. Stripe, Visa, Mastercard, Amex, and PayPal are all moving into this layer; pure plays include Nekuda, Basis Theory, Rye, and Firmly.ai. Card-network entry into the agent layer was the unlock event for this layer in 2025; before that, agents could agree on transactions but had no clean way to execute them.

Layer 4. Agents themselves. The two sides of the conversation. On the consumer side: shopping agents (Phia, Daydream, Perplexity Shopping, OpenAI Operator), advocate agents (Teleperson, the next generation of Rocket Money), and household assistants. On the brand side: customer-facing agents (Sierra, Decagon, Ada), specialized sub-agents for retention, dispute, billing, and benefits, and the merchant-side enablement platforms that help retailers expose structured catalogs to incoming agent traffic. Layer 4 is where most public attention sits, but it is downstream of the trust and payments layers; without the lower layers shipping, Layer 4 cannot transact at machine speed.

Layer 5. Marketplaces. The neutral layer where consumer agents and brand agents discover each other and transact. This is the layer that does not yet have a clear winner, and it is the layer Teleperson is building toward. Marketplaces aggregate trust signals from Layer 2, settle through Layer 3, route between agents in Layer 4, and define the negotiation grammar for Layer 1. They are the connective tissue.

Where capital is flowing

A useful test of which layer is structurally most attractive is to look at where capital is actually clearing. In the twenty-four months ending Q1 2026, disclosed venture investment into agent-to-agent commerce companies has been concentrated in the lower layers: identity and payments alone account for the majority of dollars, with merchant-side enablement (a Layer 4 sub-segment) absorbing most of the remainder. Pure consumer-advocate plays in Layer 4, the layer that is most consumer-visible, are a smaller share of capital than the headlines suggest.

Why this distribution? Because the lower layers are picks-and-shovels businesses with clear B2B revenue motions, while consumer agents are platform plays that require both a network effect and a behavioral shift to monetize. Investors prefer the former at the early stage. This will reverse: once Layers 1–3 are stable, capital will rotate to Layer 5 (the marketplace layer) because that is where the network-effect economics live.

What is still missing

A complete map should be honest about gaps. As of mid-2026, the agent commerce stack has working components in all five layers but is not yet end-to-end coherent. Three holes are particularly visible.

A unified identity standard. KYA proposals exist, but no single standard has reached the equivalent of TLS for the web. Brand agents and consumer agents currently authenticate through proprietary handshakes, which limits cross-platform interoperability. This is the most important coordination problem in the stack.

A liability model. When an agent acts on a principal's behalf and causes harm: a cancellation that should have required human approval, a payment to the wrong counterparty, a dispute settled on bad data, there is no settled answer for who pays. The legal architecture lags the technical architecture by roughly five years. Mount and a small handful of insurance-side startups are working on this, but the regulatory side is still developing.

A consumer-side incentive engine. Brand-side agents will get adopted because brands have budgets for them. Consumer-side agents need a different incentive: either subscription pricing (which most consumers resist), success-fee economics (which require trust to earn), or a marketplace where consumer agents are paid by the brands that want to be discoverable to them. The third model is structurally most attractive but also the hardest to bootstrap.

Implications

Three implications follow from the field map.

First, the layer where the category-defining company will emerge is Layer 5, the marketplace. Layers 1–3 are infrastructure that will eventually be commoditized; Layer 4 is a fragmented agent landscape with strong incumbents but no winner-take-all dynamics. The marketplace layer aggregates value from all four lower layers and is the only one with a true two-sided network effect. The companies that win Layer 5 will be the companies remembered as having defined agentic commerce.

Second, buyers should evaluate vendors against the layer they actually occupy, not the layer they claim. Many "agentic commerce platforms" are in fact merchant-side agents (Layer 4) repackaging themselves as marketplaces. Others are payment infrastructure (Layer 3) describing themselves as agentic platforms. The layer determines the buyer, the budget, the integration burden, and the long-term moat. Mismatched evaluation produces buyer's remorse within a quarter.

Third, policymakers regulating "AI agents" generically will produce bad policy. Each layer has different externalities. Layer 1 standards questions are coordination problems best handled through industry bodies, not legislation. Layer 2 identity raises consumer-protection questions that map cleanly to existing financial regulation. Layer 3 payments is already heavily regulated and needs only minor updates. Layer 4 agent behavior is where consumer-harm risk concentrates and where regulation is most needed. Layer 5 marketplace power will eventually become an antitrust question, but only after the layer has matured. Treating these as a single regulatory category will produce rules that are too coarse to be useful in any single layer.

Closing

Agent-to-agent commerce is not a product category. It is an infrastructure transformation that will reshape every interaction between a consumer and a brand involving a recurring service, a dispute, a benefit, or a negotiation. The stack is being built in public, layer by layer, with each layer attracting its own capital, its own talent, and its own competitive logic. The companies that map the stack accurately will build to the right layer; the ones that don't will spend the next three years confusing investors, customers, and themselves about what they actually are.

This map is the version we use internally at Teleperson to orient strategy, partnership, and product decisions. It will need revision as the layers mature. We expect to publish updates quarterly.