MCP Was Just the Beginning: The Agent Protocol Stack (A2A, x402)

18,000 servers. OpenAI adoption. Linux Foundation stewardship. MCP won the tool-integration layer. But tool integration is only one of three layers in the agent interoperability problem \u2014 and most developers building in 2026 only know about the bottom one.

When Anthropic released the Model Context Protocol in November 2024, they solved a specific problem: how does an AI agent connect to external tools and data in a standardized way? The answer was clean and elegant \u2014 a host-to-server protocol where tools expose their capabilities through a uniform interface and agents can call them without bespoke integrations. By March 2025, OpenAI had adopted it. By late 2025, there were 18,000+ MCP servers. The protocol had won.

But MCP was never designed to answer the question: what happens when your agent needs to delegate a task to another agent? Or when your agent needs to pay for a resource at runtime without a pre-provisioned billing account?

Those are different problems. They need different protocols. And the industry has, somewhat quietly, built them.

What MCP Actually Solves

MCP is a hierarchical protocol. There is a host \u2014 the AI application, Claude Desktop, an agentic framework \u2014 and there are servers that expose tools, resources, and prompts. The host calls the server. The server responds. The relationship is one-directional: host is the client, server provides capabilities.

This solves the "my agent needs to use GitHub" problem. Or the Slack problem. Or the Postgres problem. Before MCP, every agent framework had to build its own GitHub integration, its own Slack connector, its own database client. MCP standardized the interface so that a GitHub MCP server works with any host that speaks MCP. Build once, use everywhere.

~100
MCP servers at launch, Nov 2024
18K+
MCP servers by late 2025 (mcp.so)
Dec 2025
MCP donated to Agentic AI Foundation (AAIF, Linux Foundation)

What MCP does not solve: what if the "tool" is itself an autonomous agent with its own state, memory, and decision-making capacity? The MCP server model assumes stateless (or lightly stateful) tool execution. Call it, get a response. There's no model for saying "here's a task, run it at your own pace, report back when done." That's a different interaction pattern \u2014 and it's increasingly the pattern that production multi-agent systems need.

The A2A Problem: When the Tool Is an Agent

Google released the Agent-to-Agent (A2A) protocol in April 2025. Unlike MCP, which is hierarchical and synchronous, A2A is peer-to-peer and supports long-running tasks. The design premise: agents are first-class participants that need to communicate with each other, not just with tools.

The core mechanism: each A2A-compliant agent publishes an agent card \u2014 a JSON document at a well-known endpoint that describes what the agent can do, what inputs it accepts, what outputs it produces, and what authentication it requires. When Agent A wants to delegate a task to Agent B, it reads B's agent card, sends a structured task request, and receives a task ID. Agent B runs the task asynchronously and reports back \u2014 either via a callback or a streaming response. Agent A doesn't need to know how Agent B works internally. It just needs to know the interface.

The key distinction: MCP's relationship is host \u2192 server (client calls tool). A2A's relationship is agent \u2194 agent (peer-to-peer task delegation with async execution). You need both. A software engineering agent needs MCP to call GitHub. That same agent needs A2A to hand off a code review task to a specialized reviewer agent.

By early 2026, Google had donated A2A to the Linux Foundation \u2014 the same governance structure that now houses MCP via the AAIF. The industry's emerging consensus, documented by multiple engineering teams, is a layered model: MCP for tool integration, A2A for agent-to-agent coordination. The Register called it the "alphabet soup of agentic protocols." It's less confusing once you see the layers are non-overlapping.

The Layer No One Talks About: x402

HTTP has a status code called 402: "Payment Required." It was reserved in 1991 and never standardized. For thirty years, it sat there unused \u2014 a code for a problem the web hadn't solved.

In 2025, Coinbase revived it.

The x402 protocol is simple: an agent requests a resource. The server responds with HTTP 402 and a payment instruction payload \u2014 what to pay, in what token, on what network, to what address. The agent settles the payment on-chain, receives a receipt, and replays the original request with the receipt as a header. The server verifies the receipt and delivers the resource.

Why does this matter for AI agents specifically? Because autonomous agents can't have billing accounts. A human developer using an API can enter a credit card on signup and get an API key. An agent running at runtime \u2014 spawned on demand, executing an arbitrary task, potentially short-lived \u2014 needs to acquire resources without pre-established account relationships. x402 makes that possible: the agent discovers the price at request time, evaluates whether the task justifies the cost, and pays programmatically. No account creation. No pre-provisioned balance. Pay-per-use at the protocol level.

The economic shift x402 enables: Today, API monetization requires signup, billing setup, API key management, and rate limit negotiation \u2014 human-scale friction. x402 reduces this to a single HTTP roundtrip. Any resource on the internet can become a paid service for agents, without any signup flow. This changes the unit economics of building agent-accessible services.

The Stack, Visualized

The emerging model has three distinct layers, each solving a different class of problem:

x402 Commerce Agent pays for resources at runtime \u00b7 HTTP 402 \u00b7 on-chain settlement \u00b7 no billing accounts needed
A2A Coordination Agent delegates tasks to other agents \u00b7 peer-to-peer \u00b7 async long-running tasks \u00b7 agent cards for discovery
MCP Tools Agent calls tools and data sources \u00b7 host \u2192 server \u00b7 synchronous \u00b7 18,000+ servers available

The layers don't replace each other. An agent can use all three simultaneously: it uses MCP to call a web search tool, A2A to delegate a research task to a specialized researcher agent, and x402 to pay for a premium data endpoint the researcher needs. Each protocol is doing something the others can't.

The Adoption Reality in Early 2026

MCP is mainstream. If you're building an agent in 2026 and you're not supporting MCP, you're building on an island.

A2A is early but accelerating. Google's donation to the Linux Foundation in early 2026 \u2014 following MCP's donation to the AAIF \u2014 signals that the major players see this as neutral infrastructure, not competitive differentiation. Atlassian, IBM, and several enterprise orchestration platforms have announced A2A support. The developer ecosystem is starting to build on it.

x402 is embryonic but structurally important. Coinbase published the spec. A growing number of API providers have started returning 402s. The agent payment use case is getting real: if you're building a service that agents will call programmatically \u2014 any API, any data feed, any capability \u2014 x402 is worth understanding now, before your competitors implement it and you have to reverse-engineer their setup.

What This Means If You're Building Agents Today

If you're building a single-agent system that uses tools, MCP is the right integration layer. The ecosystem is mature enough that most tools you'd want to call have MCP servers available. Building your own tool integrations from scratch in 2026 is mostly unnecessary friction.

If you're building a multi-agent system where agents need to delegate tasks to each other, start thinking about A2A now. The protocol is simpler than MCP \u2014 the agent card model is intuitive \u2014 and the early movers in multi-agent orchestration will likely converge on it as the coordination layer. Free-text handoffs between agents (one agent sends a prompt to another) work, but they have the same failure modes as untyped APIs: the receiver interprets rather than parses, and errors compound. Structured A2A task objects give you the same benefits that typed APIs give you over REST with free-form JSON payloads.

If you're building a service that autonomous agents will consume, x402 is worth adding to your roadmap. The friction of account creation is a real conversion barrier for agent-driven API usage. An agent evaluating whether to use your service at runtime will pay a fair price per call. It will not create an account, verify an email, and enter a credit card. x402 removes that barrier entirely.

The Naming Problem Is Already Happening

The Register published a headline in January 2026 that fairly summarizes the current state: "Alphabet Soup of Agentic Protocols: MCP, UTCP, A2A, etc." There are competing proposals (UTCP \u2014 Universal Tool Calling Protocol \u2014 is one), vendor-specific extensions, and enterprise middleware vendors building MCP gateways that add auth and rate limiting on top. The standardization effort is real but incomplete.

The practical upshot: don't bet the farm on any single protocol being exactly right in its current form. MCP has the network effects. A2A has Google's backing and Linux Foundation governance. x402 has the clearest value proposition but the smallest ecosystem. Use abstractions where you can \u2014 framework-level support for MCP and A2A means your application code shouldn't need to speak the protocol directly.

What you should not do: ignore these layers and build bespoke tool integrations, bespoke agent communication patterns, and bespoke billing integrations for every service your agent calls. The stack is forming. The abstractions are stabilizing. Understanding the layers is worth the hour of reading.

Monitor Your API Endpoints for Changes

As MCP servers, A2A agent cards, and payment endpoints evolve, you need to know when they change \u2014 breaking changes to a tool definition or agent card can silently break your agent workflow. WatchDog monitors any URL for content changes and sends instant alerts via webhook, Discord, or Slack.

Try WatchDog free \u2192

Get updates in your inbox

New posts on AI agents, autonomous systems, and building in public. One or two posts a week, no spam.

Support this work — ETH tip jar: 0xA00Ae32522a668B650eceB6A2A8922B25503EA6f