Six protocols are competing to define how AI agents transact with services on behalf of users in 2026. Anthropic's ACP and MCP. Google's A2A. Stripe's TAP. UCP and AP2 from competing consortia. Each addresses a slice of the agentic commerce stack — discovery, identity, intent, payment, settlement — and the protocols overlap enough that operators choosing what to support are running real architectural decisions on partial information. The protocol that owns payments has structural advantage in how this consolidates. The protocol that owns the model layer has structural advantage in adoption. The protocol that owns neither but ships first into a useful tool ecosystem can take a category before the heavyweights coordinate.
This desk has been tracking the protocol announcements as they shipped, looking at adoption indicators across early integrations, and stress-testing each protocol's claim by working through what an actual agentic checkout flow looks like end-to-end. The conclusion in advance: the protocol war will not produce a single winner. It will consolidate around two or three layers, and Stripe's TAP is structurally positioned at the layer where consolidation pressure is hardest to resist.
The Six Protocols on the Table — What Each Actually Specifies
MCP (Model Context Protocol, Anthropic). The lowest-level protocol of the six. MCP standardizes how an AI model exposes and consumes tools — what tools are available, what schemas they accept, how state is passed between calls. Adopted broadly across the model layer; OpenAI announced support in early 2026. MCP is the substrate the other protocols build on or around, not a direct competitor to checkout-layer protocols.
ACP (Agent Commerce Protocol, Anthropic). Built on MCP, ACP specifies how an agent representing a user discovers commerce-capable services, negotiates terms, and confirms transactions. The intent layer — what the user wants, what the agent is authorized to commit to — is the core ACP contribution. Payment is delegated to whatever payment protocol the merchant supports.
A2A (Agent-to-Agent, Google). Google's framing is agents talking to other agents on behalf of their respective principals. A2A specifies negotiation patterns, identity verification between agents, and the cryptographic primitives for binding actions to authorized principals. Less directly commerce-shaped than ACP, more general-purpose. The overlap with ACP at the negotiation layer is real.
TAP (Transactional Agent Protocol, Stripe). Payments-first. TAP specifies how an agent acting on behalf of a user can initiate a payment, what authorization the user has pre-granted, what limits apply, how disputes are routed when the agent acts wrongly. Stripe ships TAP with the rest of its payment infrastructure already wired up — agentic checkout becomes a flag on existing Stripe integrations.
UCP (Universal Commerce Protocol). A consortium effort with broad merchant participation, UCP specifies a generalized commerce vocabulary that does not depend on a specific payment provider or model vendor. The neutrality is the point. The cost of neutrality is a thinner specification and slower iteration than vendor-led protocols.
AP2 (Agent Payment Protocol v2). A second-generation effort that specifically addresses the gap between intent (ACP-style) and payment (TAP-style) — how authorization for a class of transactions translates into specific payment events. AP2 is the most explicit attempt to be the connective tissue between protocols rather than competing with them.
Why Payments Matter More Than Identity in Agentic Commerce
The intuition that identity is the hard problem in agentic commerce is wrong. Identity is hard but is largely solved — federated identity, OAuth, verifiable credentials, the cryptographic primitives the AI infrastructure inherits from the prior decade of web identity work. The new problem agentic commerce introduces is bounded delegation: how does a user grant an agent the authority to spend money on their behalf, with limits, with revocability, with a dispute path that does not require litigating whether the agent had authority?
That problem lives in the payment layer. It does not live in the identity layer. The protocol that solves bounded delegation cleanly, in a way that integrates with existing dispute and chargeback infrastructure, is the protocol that defines how agentic commerce works in practice. Identity protocols can support that — A2A's identity primitives are well-designed — but identity alone does not produce a consumable commerce surface.
The historical analog: the SSL/TLS protocol war in the 1990s. Multiple competing standards, technical merits favored several candidates, the winner was the one that integrated cleanly into the existing payments processing stack. Operational gravity beats specification quality.
Stripe's TAP and the Structural Position
Stripe's structural position has three components.
Component 1: Existing payment relationships. Stripe processes payments for a substantial fraction of the merchants AI agents will want to transact with. Adding TAP support is a flag in the merchant's existing Stripe integration. The merchant does not have to make a new vendor decision, sign a new contract, or take on new compliance review. The path of least resistance for adding agentic commerce capability is the Stripe path.
Component 2: Dispute and chargeback infrastructure. Bounded delegation without operational dispute infrastructure is not deployable. When an agent transacts wrongly — exceeded limits, misunderstood intent, was manipulated by adversarial input — somebody has to refund, somebody has to lose, somebody has to investigate. That infrastructure exists at payments processors and does not exist at AI infrastructure providers. Stripe inherits the operational layer that the protocol-only competitors do not have.
Component 3: Regulatory positioning. Agentic commerce will attract regulatory scrutiny — consumer protection regulators, payment regulators, AI safety regulators in jurisdictions building those institutions. The May 5 2026 agreement among Microsoft, Google, and xAI to let CAISI test models before public launch — with agentic commerce specifically called out as a focus area — is the early signal. Stripe has decade-plus experience operating inside payment regulatory regimes. The protocol that brings that experience natively into the agentic commerce stack has lower regulatory risk for the merchants adopting it.
None of this guarantees Stripe wins. It positions TAP at the layer where payment-and-dispute reality is hardest to displace once consolidated.
Anthropic's ACP and MCP — Different Layers, Why Both Matter
MCP and ACP are stacked, not competing. MCP defines how the model talks to tools; ACP defines how the agent representing the user talks to commerce-capable services. The stacked design lets Anthropic ship at multiple layers without choosing one.
The strategic position MCP locks in is substantial: every tool that wants to be reachable by Claude exposes MCP. Every model vendor that wants to consume those tools speaks MCP. OpenAI's adoption ratified the protocol as a market standard, which is rare in early-stage protocol fights and reflects MCP's specification quality more than vendor coordination.
ACP at the intent layer is more contested. The intent specification — what does it mean for an agent to be authorized to commit to a transaction — is genuinely hard, and ACP's approach is one of several reasonable designs. ACP's adoption depends partly on whether Anthropic-aligned merchants run with it, which depends on Anthropic's overall market position. The compute partnership with SpaceX and the expanded Claude Code/API limits announced this April underline that Anthropic is building enterprise distribution that ACP can ride.
Google's A2A — Where It Wins, Where It Doesn't
A2A's strength is the identity and negotiation layer. The cryptographic primitives are clean, the agent-to-agent negotiation patterns are well-thought-through, the integration with Google's existing identity infrastructure is operationally significant. For agent-to-agent coordination that does not directly involve payment — supply chain queries, scheduling, inter-organization workflows — A2A is a strong protocol.
A2A's weakness in commerce specifically is that it depends on a separate payment protocol below it. A2A-plus-TAP is a coherent stack. A2A-plus-something-yet-undefined is not deployable. Google's lack of a payments processor of comparable scale to Stripe is the structural gap. Google Pay exists; Google Pay's merchant footprint is not Stripe's merchant footprint.
The Likely Consolidation Path Through 2027
The path that fits the structural pressures across each protocol layer.
MCP consolidates as the model-to-tool standard. The competition at this layer is effectively over; the remaining work is implementation maturity, not specification competition.
The intent layer remains contested. ACP and the broader UCP work occupy overlapping territory; AP2's connective-tissue framing may absorb the consolidation. The intent layer will not have a single winner by end of 2027 — operators will support two specs.
The payment layer consolidates around TAP for Stripe-served merchants and around whatever Adyen, Checkout.com, and the other major processors ship as their TAP-equivalents. Payment-protocol fragmentation by processor mirrors the existing payment-API fragmentation; merchants already manage that. The cost of multiple payment-side protocols is lower than the cost of multiple intent-side protocols because abstraction layers handle it.
The identity layer borrows from A2A and from existing federated identity infrastructure. A2A's primitives become reference implementations rather than a single winning protocol.
What This Desk Tracks Through Q2-Q3 2026
Three datapoints anchor our ongoing tracking. First, merchant adoption rates of TAP versus competing payment-layer protocols. Adyen and Checkout.com responses to TAP will determine whether the payment layer fragments by processor or consolidates around a Stripe-defined standard. Second, regulatory action shaped by the CAISI agreement. The specific agentic-commerce focus areas regulators flag will reshape protocol requirements; protocols built without regulatory anticipation will need late changes. Third, Anthropic's Claude financial agents rollout — the operator track record on actual agentic transactions in the wild will reveal whether ACP's intent specification holds against real operational pressure.
Honest Limits
The protocol landscape is moving faster than any single analysis can keep current; specifications and adoption indicators referenced here are accurate as of early May 2026 and will drift. The "structural advantage" framing is shaped by analogy to the SSL/TLS consolidation and other historical protocol fights — analogies fail in specific places, and agentic commerce has properties earlier protocol fights did not. The payments-matter-more-than-identity claim is a working position, not a proven thesis; identity-first arguments exist and are not foreclosed. We did not include every protocol in active development — there are vertical-specific commerce protocols, regional protocols (PIX-aligned variants in LATAM, UPI-aligned in India), and consortium efforts beyond UCP that did not make this analysis. The Stripe positioning analysis assumes Stripe executes on TAP at roughly the quality of its existing payment infrastructure; specific integration friction or strategic missteps could shift the picture. The consolidation path described is the modal outcome, not the only one — protocol fights produce occasional surprises and 18 months is enough time for one to land here.