# MCP, A2A and micropayments: the infrastructure of autonomous commerce between AI agents

> Anthropic, Google and the big tech companies are driving MCP, A2A and micropayments to make autonomous commerce a reality
> By José Maldonado · AI Agents · 2026-08-01

Anthropic, Google, Mastercard, Visa, Stripe, Microsoft and AWS are leading the development of MCP, A2A and the micropayment protocols that will make autonomous commerce between artificial intelligence agents possible

On July 28, 2026, the Model Context Protocol (MCP) [**published**](https://blog.modelcontextprotocol.io/posts/2026-07-28/) its fifth and most ambitious revision: a stateless core that transforms the protocol from a bidirectional model with persistent sessions into a decoupled request-response model.

The same ecosystem that MCP standardizes (AI agents connecting to tools, data and services) processed more than 120 million autonomous transactions in a single week of February 2026, according to Alipay. But the market research firm [**McKinsey**](https://www.mckinsey.com/), projects that agentic commerce will move between 3 and 5 trillion dollars by 2030. What is happening is not a trend: it is the deliberate construction of a protocol stack that will allow software agents to discover each other, negotiate, execute payments of fractions of a cent and settle agreements without a human touching a button.

## The two protocols that define agent interoperability

The AI agent ecosystem today has two open standards that solve complementary problems, and understanding the difference between them is essential to read what is coming.

MCP, published by Anthropic as an open standard on November 25, 2024, solves communication between an agent and its tools. Before MCP, every integration between a language model and a database, an API or an enterprise system required custom connectors.

MCP standardizes that connection with a client-server model over JSON-RPC 2.0, where servers expose three types of primitives (resources, tools, executable functions and prompts) and any compatible client can consume them. The recurring metaphor in the industry is that MCP is the "USB-C of AI": a universal connector that eliminates the need for proprietary adapters for every combination of agent and service.

**[Claude Mythos breaks the HAWK algorithm and accelerates attacks on AES and LEA](https://observatorioblockchain.com/en/technology/claude-mythos-rompe-hawk-y-acelera-los-ataques-a-aes-y-lea/)**

## The arrival of the A2A protocol

For its part, A2A (Agent-to-Agent), originally developed by Google and donated to the Linux Foundation, solves a different problem: communication between agents. Its version 1.0, published in 2026, defines how agents built on different frameworks (LangGraph, CrewAI, Semantic Kernel, Google's ADK) can discover the capabilities of other agents through Agent Cards, negotiate interaction modalities (text, files, structured data), delegate tasks and collaborate in long-running workflows without exposing their internal state or their proprietary tools. The A2A technical committee includes representatives from AWS, Cisco, Google, IBM Research, Microsoft, Salesforce, SAP and ServiceNow.

![](https://content.observatorioblockchain.com/wp-content/uploads/2026/08/mcp-a2a-scaled.webp)

The relationship between the two protocols is one of complementarity, not competition. MCP operates within the agent: it is how the agent connects to its tools, APIs and data sources. A2A operates between agents: it is how one agent discovers another, delegates a subtask to it and receives the result. A typical multi-agent system will use both: MCP inward, A2A outward.

In any case, MCP adoption has been explosive. By early 2025 (barely months after its launch) OpenAI, Google DeepMind, Microsoft and Meta AI had already adopted it natively. In July 2026, the TypeScript and Python SDKs [**have surpassed 1 billion total downloads each**](https://www.bejob.com/lenguajes-programacion-mas-usados-2026/), with a combined volume close to 500 million monthly downloads. There are official implementations in TypeScript, Python, Go, C# and Rust. The community has built thousands of MCP servers covering everything from Google Drive and GitHub to enterprise databases and design platforms such as Figma.

## The update that changes the deployment rules

But the new MCP specification takes all its possibilities one step further. Until now, MCP required persistent sessions between client and server: **an initial handshake negotiated capabilities, a session identifier was established and all subsequent requests depended on that shared context.**

This worked in local environments, after all, the original inspiration was Microsoft's Language Server Protocol, but it was a burden in production. A remote MCP server could not scale horizontally behind a conventional load balancer without session coordination. Each deployment required draining active connections. Instances were not interchangeable.

The new specification eliminates all of that. Each request travels independently with its own protocol version, client identity and declared capabilities. The handshake and the session header or identifier are officially deprecated. Remote MCP servers become stateless HTTP services that any instance can handle. Round-robin balancing, autoscaling and serverless containers work without modifications. Two new headers (`Mcp-Method` and `Mcp-Name`) allow gateways and load balancers to route MCP traffic at the header level without parsing JSON-RPC bodies.

**[Cybersecurity in the AI Era: How Red, Blue and Green teams protect your company](https://observatorioblockchain.com/en/ai-agents/ciberseguridad-en-la-era-de-la-ia-el-trinomio-red-blue-y-green/)**

## Optimizing MCP capabilities

The update also introduces formal extensions: **Tasks for long-running operations (contributed by AWS) and Enterprise Managed Authorization for corporate authentication aligned with OAuth 2.0 and OpenID Connect.** The proprietary logging channel that MCP used until now is replaced by OpenTelemetry with W3C Trace Context, which enables distributed traces that cross the agent, the MCP client, the gateway and downstream services, rendering as a unified span tree in any compatible collector.

Corporate endorsements are not cosmetic. AWS announced that Amazon Bedrock AgentCore supports the new specification from day zero, with a single call to `UpdateGateway`. Cloudflare integrates MCP `2026-07-28` into Workers and its Agents SDK. Google Cloud called it "a massive leap in enterprise AI scalability." Microsoft Foundry uses it as the basis of its unified toolbox. Netlify, Figma, New Relic, Honeycomb and Intuit published statements of support on the day of the launch. The transition has a twelve-month window with version coexistence, and the protocol maintainers have stated that this is the last revision with breaking changes.

## Fraction-of-a-cent payments: the layer that turns agents into economic actors

An agent that can converse but cannot pay is a mere assistant. An agent that can converse and pay is an economic actor. The infrastructure for that second mode is being built now, and it is not a monolith: it is an ecosystem of protocols converging from different directions.

Mastercard launched Agent Pay for Machines (AP4M) in June 2026, a service that allows AI agents to execute programmatic payments with guaranteed settlement through its global network. The system includes verifiable credentials for each agent, programmatic authorization rules with spending limits and multi-rail settlement: **cards, bank accounts and stablecoins. Cloudflare, Ripple, Solana Foundation and Turnkey participate as launch partners.**

Visa, for its part, published its Trusted Agent Protocol (TAP) and a white paper detailing how card infrastructure can be extended for agentic commerce. Stripe launched its Machine Payments Protocol (MPP), which covers both on-chain payments with stablecoins and fiat payments through shared payment tokens. The x402 protocol (which revives the HTTP 402 – Payment Required status code) allows a server to require a micropayment for a resource and an agent to settle it in milliseconds, with no checkout page, no stored card and no browsing session.

## Banking expansion

The International Monetary Fund, on the other hand, devoted a [**working paper**](https://www.imf.org/-/media/files/publications/imf-notes/2026/english/insea2026002.pdf) in April 2026 to analyzing the impact of agentic AI on payment systems. Its framework proposes three layers: a probabilistic decision layer where the agent interprets the user's intent, a deterministic controls layer where spending limits, identity and restrictions are verified, and a final settlement layer where the transaction is executed without reinterpretation. The separation is crucial: it allows agents to operate with autonomy in the upper layer while the lower layers maintain the legal and financial stability guarantees that payment systems require.

A paper by Stanford and Manchester Metropolitan University, published on arXiv in June 2026, argues that the arrival of native micropayment rails for agents (x402, Google's AP2) changes what is scarce in e-commerce. When the buyer is a tireless agent that can research, compare and verify at scale, product matching becomes cheap; what remains scarce is verifiable and relevant information for the decision.

The authors propose a market of micro-transactions where buying agents spend fractions of a cent to progressively unlock data that a seller or an independent reviewer has chosen to expose, paid à la carte under a freemium model. A product that withstands scrutiny can afford to reveal more; an agent that values its data well earns proportionally more.

**[NVIDIA, Meta and Microsoft back open AI without OpenAI, Google and xAI](https://observatorioblockchain.com/ia/nvidia-meta-y-microsoft-apoyan-la-ia-abierta-sin-openai-y-google/)**

## The agentic economy in numbers

Early adoption data suggest that agentic commerce is not a ten-year projection: it is happening in specific markets at significant scale.

According to NielsenIQ, Alipay AI Pay [**processed**](https://www.stocktitan.net/news/NIQ/120-million-ai-pay-transactions-in-a-single-week-ai-shopping-agents-n2c3f2ern9v8.html) more than 120 million agent-driven transactions during the week of February 5-11, 2026. Alibaba's Qwen assistant reached 100 million monthly active users in two months since its launch. Adobe Analytics reported that traffic from generative AI sources to retail sites in the United States increased 1,200% between July 2024 and February 2025, with visitors showing higher engagement (more pages per visit, lower bounce rate) although conversion still lagged behind traditional channels.

McKinsey projects that the global agentic commerce market will reach between 3 and 5 trillion dollars by 2030. Morgan Stanley estimates that agents will drive between 190 and 385 billion dollars of U.S. e-commerce by that same date, approximately 10% to 20% of the total.

However, full autonomy remains a minority. NielsenIQ data show that around 34% of consumers use AI to research products and 23% to summarize reviews, but only 8% have allowed an AI to complete a purchase. The influence of agents is growing much faster than the delegation of the final decision. The gap between recommendation and execution is where the battle for consumer trust is being fought.

## More "realistic and less reactive" commerce

EY [**published**](https://www.ey.com/en_us/insights/ai/agentic-commerce-strategy) in July 2026 a strategic framework for companies to prepare for agentic commerce. Its diagnosis is direct: agents do not see brands the way humans do. They do not respond to storytelling, images or in-store experiences. They evaluate structured signals: product attributes, prices, availability, eligibility conditions, return policies and expected results.

If a brand's differentiation is not machine-readable, it is invisible to an agent. EY recommends that organizations build a "single commercial truth" (consistent product identifiers, authoritative prices, policies expressed as executable rules) and explicitly define where agents can act with autonomy and where human approval is required.

MIT, through its Initiative on the Digital Economy, published in May 2026 a paper titled [**"The Coasean Singularity"**](https://www.nber.org/papers/w34468) that examines how AI agents will reduce transaction costs to the point of transforming market structures. The central argument: when the costs of searching, comparing, negotiating and verifying collapse (because an agent can do all of that in parallel, at machine speed and for fractions of a cent), markets where human intermediaries exist (insurance, real estate, financial services, hiring) reorganize.

The authors do not present it as a prediction but as a fork: the same technical capability can produce radically different market structures depending on who controls the agents, whether they are provided by the platform (like the shoes rented at a bowling alley) or whether the user brings their own, portable and aligned with their own interests.

**[AI changes free software development forever](https://observatorioblockchain.com/ia/a-ia-cambia-para-siempre-el-desarrollo-del-software-abierto/)**

## How does an agentic transaction work, step by step?

The full flow of an agentic transaction illustrates how the pieces fit together. Suppose a user asks their personal assistant: "I need auto insurance with extended liability coverage, less than 800 euros a year, that also covers roadside assistance in Europe."

The user's agent, running in an MCP client such as Claude Desktop, ChatGPT or an enterprise assistant, receives the instruction. It uses MCP to connect to its tools: a server that queries insurer quotes, another that verifies the user's claims history, another that accesses their calendar to coordinate an inspection if necessary. Each of those connections is a standardized MCP call: the agent does not need to know what proprietary API is behind each server.

To interact with the insurers' agents, the user's agent uses A2A. It discovers the Agent Cards of multiple insurers, each declaring its capabilities, interaction modalities and endpoints, negotiates parameters, sends quote requests and receives structured offers. If an insurer charges for consulting its roadside assistance rate, the user's agent evaluates the price (0.03 cents), authorizes the micropayment through the x402 protocol or Mastercard's AP4M, receives the data and continues. Everything happens in milliseconds, without the user seeing a checkout page or entering a card.

## Culminating the process

When the agent identifies the three best options, it presents them to the user with a breakdown of coverages, exclusions and prices. The user chooses one. The agent executes the contracting (digital signature, premium payment, policy issuance) coordinating with the insurer's agent through A2A and settling the payment through Mastercard or Stripe rails. The complete traceability of the transaction is recorded through OpenTelemetry with W3C Trace Context, propagating through each hop: buying agent, MCP client, insurer's MCP server, payment gateway, settlement system.

This flow is not speculative in all its parts. The MCP and A2A protocols exist in production. The payment layers —Mastercard's AP4M, Stripe's MPP, Visa's TAP— are operational or in advanced deployment. What is still under construction is the full integration of the three layers into multi-agent flows that operate at scale and with consumer trust.

**[Agentic Market: this is how the API and services marketplace for AI agents works](https://observatorioblockchain.com/en/ai-agents/agentic-market-asi-es-el-mercado-de-apis-para-agentes-de-ia/)**

## What's next

Thus, we can see three vectors that will define the next twelve to eighteen months. The first is the consolidation of the protocol stack. MCP and A2A are not the only standards in play, Cisco's Internet of Agents, Google's Agent Payments Protocol, x402, Stripe's MPP, but they are the ones that have reached critical mass of adoption.

The question is not whether they will converge, but what form that convergence will take: **a multi-protocol gateway that translates between them transparently, or a consolidation toward fewer standards?** The historical pattern of the internet suggests the former: TCP/IP did not eliminate other protocols, it became the translation layer.

## Trust and regulation

The second is the trust gap. NielsenIQ data are eloquent: 34% of consumers let AI research, only 8% let it buy. Closing that gap requires not only better technology, but governance frameworks that precisely define what an agent can do, under what conditions and with what traceability. The FIDO Alliance announced in April 2026 two working groups, in collaboration with Google and Mastercard, to develop authentication standards specific to agent-initiated payments.

The third is regulatory. The IMF's three-layer framework (probabilistic decision, deterministic controls, final settlement) offers a solid conceptual model, but its implementation requires coordination between central banks, payment networks and digital identity providers. The BIS already demonstrated in 2025 that generative AI systems can manage treasury functions (liquidity, payment prioritization) without specialized training, which suggests that technical supervision is viable if designed from the ground up.

The agentic economy is not a distant future: it is a protocol stack that is being assembled now, layer by layer, with real enterprise adoption and transaction volumes already measured in millions per week. Companies that structure their data to be machine-readable, define agentic autonomy policies with explicit governance and monitor their presence in agents' consideration sets will be competing in tomorrow's market. Those that do not simply will not appear in the results.
