What Is This?
In 1974, Vint Cerf and Bob Kahn published the specification for TCP/IP — the protocol that would allow different computers to communicate regardless of who made them. Before that, every computer network was a closed island. After that, any machine could talk to any other machine. The internet as we know it became possible.
AI in 2025 is at the same inflection point. And you're living inside a system that already runs on the protocols about to change everything.
Two open standards have emerged that together define how AI agents talk to tools and to each other:
MCP (Model Context Protocol) — Released by Anthropic in November 2024. Think of it as the USB-C port for AI models. Before MCP, connecting an AI to a database, a file system, an API, or any external tool required bespoke integration work for every combination. MCP standardises the interface: any MCP-compatible tool can plug into any MCP-compatible model. OpenClaw runs on MCP. That's why you can give it access to your files, calendar, GitHub, and Telegram through the same interface.^1
A2A (Agent-to-Agent Protocol) — Released by Google Cloud in April 2025, with 50+ companies contributing. Where MCP handles agent-to-tool communication, A2A handles agent-to-agent communication. An orchestrator agent can discover other agents (via standardised "Agent Cards"), delegate tasks to them, receive results back, and coordinate complex multi-step workflows — without a human in the loop.^2
Together, these two protocols are building the infrastructure layer that turns AI from a collection of isolated chatbots into a functioning networked workforce.
Why Does It Matter?
- This is the infrastructure layer, not the application layer. Most AI coverage focuses on which model is smarter. That's missing the point. Models will keep improving, but without standardised communication protocols, each improvement stays trapped in its silo. MCP and A2A are the reason improvements in any model can be immediately usable by any agent in any system — like HTTP meant any browser could access any website.
- You already live inside this. OpenClaw's entire architecture — Scout scanning X, Researcher writing articles, Claw building products, all coordinating through shared files — is an early implementation of what A2A formalises. The "shared brain" architecture is the primitive version. A2A is the standardised, interoperable version.
- The economic implications are enormous. When agents can autonomously discover, hire, and pay other agents for sub-tasks, labour markets change fundamentally. An agent running a marketing campaign can autonomously hire a copywriting agent, a design agent, a scheduling agent, and a data analysis agent — with no human orchestration required. This isn't science fiction; it's what A2A is designed to enable.
- Whoever controls the protocol controls the ecosystem. Both MCP and A2A have been donated to the Linux Foundation — deliberately designed to be open standards rather than proprietary moats. But the companies that build the first widely-adopted agent directories, registries, and orchestration layers will have enormous leverage. This is the battle being fought right now.
- Understanding this now is like understanding HTTP in 1993. Most people couldn't explain TCP/IP in 1993 and it didn't matter — the web worked anyway. But the people who understood it built every major internet company. The protocol layer only matters if you want to build on it. If you do, now is when to learn it.
Key People & Players
Anthropic — Created MCP. The stated goal: "USB-C for AI." Already supported by Claude, Cursor, Windsurf, and hundreds of tools. Open sourced and donated to Linux Foundation.^3
Google Cloud — Created A2A. Launched with 50+ enterprise partners including SAP, Salesforce, and Deloitte. Explicitly designed to complement MCP rather than replace it: "MCP connects agents to tools; A2A connects agents to agents."^4
OpenAI — Has its own agents framework (Swarm, then the Assistants API, now the Responses API with built-in tool use). Not yet on MCP/A2A, but the competitive pressure from open standards will force convergence or interoperability.
Microsoft — Major backer of A2A. Azure is building A2A support into its AI infrastructure. Given Microsoft's enterprise reach, this will drive A2A adoption in corporate AI deployments faster than anything else.
Linux Foundation — Stewards both protocols. The political decision to put them here (rather than keep them proprietary) signals that the major players want a neutral standards body — similar to how W3C governs web standards.
The Current State
MCP is already production-ready and widely deployed. Thousands of MCP servers exist for everything from GitHub to Slack to local file systems. Claude natively supports it. Cursor and Windsurf use it. If you're using any modern AI coding tool, you're likely already on MCP.
A2A is newer (April 2025) and still being adopted, but enterprise rollout is moving fast. The key concepts:
Agent Cards — Standardised JSON documents that describe what an agent can do, what inputs it accepts, what it returns, and how to authenticate with it. Agents advertise their capabilities; orchestrators discover and select them.
Task lifecycle — A2A defines a clear protocol for task creation, streaming updates, state management, and completion. Orchestrators can monitor long-running tasks, handle failures, and retry intelligently.
Authentication — Built-in OAuth 2.0 support, so agents can delegate credentials securely without sharing raw tokens.
The current gap: discovery and trust. If any agent can advertise any capability via an Agent Card, how does an orchestrator know which agents are trustworthy, well-performing, and accurately described? The agent directory/registry problem is unsolved and will likely be the next major battleground — whoever builds the "App Store for agents" will sit at a pivotal chokepoint.
For builders right now:
- Learn MCP first. It's stable, documented, and immediately useful. Building an MCP server for your own tools is the fastest way to make any AI system interoperable with the ecosystem.
- Watch A2A. The enterprise wave is coming. Understanding multi-agent orchestration before it becomes standard infrastructure is the equivalent of learning REST APIs before SaaS took off.
Best Resources to Learn More
- MCP Official Docs (Anthropic) — The spec, quickstart, and server examples. Start here for implementation.^5
- Google A2A Announcement — The original announcement with architecture overview.^6
- MCP and A2A: The Protocols Building the AI Agent Internet (Medium) — Best plain-English explanation of how the two fit together.^7
- Gravitee: A2A and MCP — Complementary Protocols — Technical comparison for builders.^8
- AI Agent Protocols 2026 Complete Guide — Broader survey including IBM's BeeAgent and other emerging standards.^9