Agent-to-Agent Protocols: A Comparative Survey of MCP, A2A, ACP, and the Card-Network Entries

A common framing in industry coverage of agent-to-agent commerce is that the protocol layer is fragmented: that eight competing standards from Anthropic, Google, OpenAI, Visa, Mastercard, Klarna, PayPal, and others represent a chaos of incompatible systems that will eventually need to be consolidated. This framing is partly accurate but mostly misleading. The eight protocols are not competing for the same role. They are colonizing different layers of the agent-to-agent stack, and the apparent fragmentation is in fact a layered settlement that is already partially visible if you map the protocols to the layers they actually occupy.

This paper provides a comparative survey of the major agent-to-agent protocols as of late 2025, maps each to its layer in the stack, identifies the genuine areas of competition (where multiple protocols are vying for the same role), and predicts where the settlement is likely to land for each layer.

The layered map

The agent-to-agent protocol stack, from bottom to top:

Layer A. Tool use and capability invocation. How an agent connects to external tools, APIs, and data sources. This is the layer where Anthropic's Model Context Protocol (MCP) operates. MCP was published in late 2024 and has been adopted broadly across the agentic ecosystem as the de facto standard for agent-to-tool integration. The standardization at this layer is essentially complete; competing approaches (proprietary tool-use schemas inside individual agent platforms) persist but are increasingly seen as legacy.

Layer B. Agent-to-agent interoperability. How two agents talk to each other regardless of their underlying platform. This is the layer where Google's Agent2Agent (A2A) protocol operates. Launched in April 2025 with a coalition of partners including PayPal, Salesforce, Atlassian, Intuit, and SAP, A2A defines the wire format and the basic vocabulary for agent-to-agent message exchange. This layer is partially settled, A2A has the most institutional backing, but is still in active competition with proprietary alternatives from major platform vendors that prefer to control the inter-agent protocol on their own surfaces.

Layer C. Commerce-specific agent vocabulary. How an agent expresses commerce-specific concepts (offers, orders, returns, disputes, payments). This is the layer where OpenAI's Agentic Commerce Protocol and Klarna's Agentic Product Protocol operate. Both define vocabularies for commerce transactions specifically, with overlap but also meaningful differences in scope. This layer is the most actively contested in the entire stack: multiple credible standards exist, the technical merits of each are real, and the competitive dynamics among the issuing companies (OpenAI, Klarna, Shopify, others) are still resolving.

Layer D. Payment-network integration. How an agent transaction settles through the financial system. This is the layer where Visa's Trusted Agent Protocol (TAP), Mastercard's Agent Pay, Amex's agent-payment integrations, and PayPal's various integrations operate. Each card network is positioning to be the preferred settlement layer for agent transactions on its rails. The standardization here is least likely, the card networks have decades of competitive coexistence at the human-payments layer and will likely continue to coexist at the agent-payments layer.

Layer E. Identity and trust. Layer A through D handle the syntax of agent transactions; Layer E handles the trust. This layer has the most embryonic standardization. KYA proposals exist from multiple parties (Skyfire, Kite, identity-specific startups), but no single standard has reached the equivalent of OAuth or DIDs for the agent-identity layer. This is the most under-settled layer in the stack and the most consequential for category formation.

Where the genuine competition lives

The eight-protocol framing is misleading because it implies all eight protocols are doing the same thing. In fact, only three layers contain genuine multi-vendor competition for the same role:

Layer C (commerce vocabulary) is the most actively contested. OpenAI's Agentic Commerce Protocol, Klarna's Agentic Product Protocol, Shopify's emerging commerce-agent specifications, and several merchant-side enablement plays are all positioning to define how commerce concepts are expressed in agent transactions. The technical differences are real but bridgeable; the competitive differences are about platform control and ecosystem capture. We expect this layer to settle on two or three coexisting standards (likely OpenAI's, one merchant-side standard, and one financial-services-specific standard) over the next two years, with cross-protocol bridges enabling interoperability.

Layer D (payment-network integration) is contested but the contest is structural rather than zero-sum. Each card network is building its own agent-payment standard because each network has its own merchant base, issuing-bank relationships, and dispute-resolution infrastructure. The card networks have coexisted at the human-payments layer for decades; they will coexist at the agent-payments layer. Standardization here is unlikely and probably unnecessary, the wire format an agent uses to settle a Visa transaction can differ from the wire format it uses to settle a Mastercard transaction without producing meaningful friction for either side.

Layer E (identity) is contested in the sense that no single standard has emerged but uncontested in the sense that everyone agrees one is needed. The under-settlement at this layer is the most consequential gap in the entire stack, because the identity standard is what makes the trust layer possible. We expect a settlement here within eighteen months: likely involving the W3C, the major identity-protocol bodies, and the financial-industry standards organizations, but the path is less predictable than the layers above it.

Where the standards have effectively settled

Three layers are essentially settled, even though the public framing has not caught up.

Layer A (tool use): MCP has won. Every major agent-development framework supports MCP. Every major LLM provider has either adopted MCP or shipped a compatible interface. The remaining proprietary tool-use schemas are legacy code paths that vendors are migrating off of. There is no realistic challenger to MCP at this layer.

Layer B (agent-to-agent interoperability): A2A is the lead, with persistent proprietary fallback. Google's A2A has the most institutional adoption and is the closest thing to a default standard for cross-platform agent communication. The persistent alternative is platform-specific protocols that major vendors prefer for transactions occurring entirely within their own ecosystems. We expect A2A to dominate cross-platform agent traffic and proprietary protocols to handle within-platform traffic, with bridges between the two becoming a standard product category.

Layer C (commerce vocabulary): two-to-three coexisting standards. As noted above, this layer will not settle on a single standard, but the two or three that emerge will be bridged. Vendors building for this layer should plan for multi-standard support from the start.

What this means for builders and buyers

Three implications.

Builders should commit to MCP and A2A early and treat the contested layers as multi-standard. The lower-layer commitments (MCP for tool use, A2A for inter-agent communication) are safe long-term bets. The higher-layer choices (commerce vocabulary, payment integration) should be made with the assumption that multi-standard support will be required indefinitely.

Buyers should evaluate vendors on protocol depth, not protocol selection. A vendor that has shipped good MCP integration, good A2A interoperability, and a clear path to multi-standard support at the higher layers is a more durable bet than a vendor that has bet exclusively on one specific protocol at every layer. The single-protocol bet looks decisive in a deck and looks fragile in a five-year roadmap.

Investors should size the protocol layer as infrastructure, not as a category. The protocol layer is not where the category-defining companies will emerge. The protocols are the substrate on which the category-defining companies (in agent platforms, marketplaces, identity providers, and consumer advocates) will build. Capital deployed at the protocol layer is best understood as infrastructure investing: important, but with different return profiles than the application-layer investments above it.

What the survey leaves out

A note on what this paper does not address. We have not surveyed the regional protocols emerging in China, India, the EU, and other jurisdictions. Several major regional standards are in development, and the cross-jurisdictional interoperability question, whether a Chinese consumer agent can transact with an EU brand agent under whose protocol, is going to be important by 2027. We expect to publish a follow-up specifically on the regional protocols and the cross-jurisdictional interoperability layer in early 2026.

We have also not addressed the open-source agent-protocol ecosystem in detail. Several promising open-source efforts are emerging, and the question of whether open-source protocols can compete with the institution-backed standards above is genuinely open. The open-source side typically lags institutional backing by twelve to eighteen months in coordination problems of this kind, but the gap can close if institutional standards develop in directions that produce broad dissatisfaction.

Closing

The eight-protocol fragmentation framing is misleading. The protocols are colonizing different layers of the stack, three layers are essentially settled, and the genuine contests are concentrated in commerce vocabulary, payment-network integration, and identity. Builders, buyers, and investors who internalize the layered map will make better decisions than those who treat the protocol layer as undifferentiated chaos.

The settlement that is emerging: MCP for tool use, A2A for inter-agent communication, multiple commerce vocabularies bridged by interoperability layers, network-specific payment integrations, and an identity standard yet to be resolved, is the protocol substrate on which the agent-to-agent commerce era will be built. The companies that build on this substrate accurately will compound. The ones that fight the substrate will spend the next three years explaining why their proprietary stack is incompatible with the rest of the ecosystem.