Building AI Agents · Part 7 of 7 · MCP, A2A, UCP, AP2, A2UI & AG-UI · Oct 2026
Quick answer: Agent protocols are open standards for how agents talk to tools, to each other and to users, regardless of the framework or language they were built with. The key ones are MCP (agents to tools and data), A2A (agent to agent), UCP and AP2 (commerce and payments), and A2UI and AG-UI (agents to user interfaces).
Why Agents Need Protocols
An agent built with one framework usually can't talk to an agent built with another. Agent protocols fix this. They are open specifications that standardise how agents, tools, resources and users interact over the internet, whatever the framework, language or platform. An agent written in Python can then work with an agent written in TypeScript.
Protocols tackle the challenge of cross-platform, decentralised and secure interaction between agents, which is a major barrier to a true "agentic web". Because protocols are about communication, latency is a key measure of how well they perform.
Protocols also make subagents (Part 5) far more useful: your parent agent can delegate to specialist agents that run anywhere, not just to ones inside your own codebase.
The Protocols You Should Know
| Protocol | Created by | Connects |
|---|---|---|
| MCP (Model Context Protocol) | Anthropic | Agents to tools, data and resources |
| A2A (Agent2Agent) | Google, now under the Linux Foundation | Agents to other agents |
| UCP (Universal Commerce Protocol) | Google and partners | Agents to merchants for shopping and checkout |
| AP2 (Agent Payments Protocol) | Google and partners | Agents to payment systems, with verifiable authorisation |
| A2UI (Agent to UI) | Agents to user interfaces, by describing UI widgets | |
| AG-UI (Agent–User Interaction) | CopilotKit | Agent back-ends to front-end apps, with real-time streaming |
MCP: Model Context Protocol
MCP, created by Anthropic, standardises how agents connect to tools and data. It uses a client–server design: an MCP server exposes tools, resources and prompts, and an MCP client inside your agent connects to it. Write a server once, for example for your accounting system, and any MCP-compatible agent can use it. The Vercel AI SDK supports MCP, so an agent can discover and use tools from MCP servers.
A2A: Agent2Agent
A2A, introduced by Google in April 2025 and donated to the Linux Foundation that June, lets agents discover each other, delegate tasks and collaborate as peers. Each agent publishes an "agent card" describing what it can do, so other agents know when and how to call it. IBM's Agent Communication Protocol (ACP), which offered a similar framework for agents to discover, negotiate with and work alongside each other, merged into A2A in August 2025.
UCP and AP2: Commerce and Payments
The Universal Commerce Protocol (UCP) is an open standard for agentic commerce. It lets agents discover what a merchant offers and complete purchases. It works with the Agent Payments Protocol (AP2), which acts as the trust layer: instead of handing over a card number, the agent presents cryptographically signed proof that the user authorised the payment.
A2UI and AG-UI: Agents and Interfaces
A2UI, from Google, lets an agent return user-interface widgets, such as forms, cards and buttons, as part of its response. The agent describes the UI and the app renders it. AG-UI, from CopilotKit, connects a user-facing app to any agent back-end and streams the agent's events (Part 6) in real time. The two are complementary: A2UI describes what to show, AG-UI handles how it gets to the screen.
AGNTCY
AGNTCY is not a single protocol but a broader framework. Started by Cisco and donated to the Linux Foundation in July 2025, it provides infrastructure for an "Internet of Agents": agent discovery, identity verification, communication and observability.
User-to-Agent Communication
Agents also need to communicate with people, and there are many channels:
- Voice: phone calls or voice chat in a website widget or app.
- Chat: web chat widgets, SMS, RCS, WhatsApp or Facebook Messenger, which support text, images, carousels, links, videos and interactive cards.
- Video: one-way video so the agent can "see what I see", or two-way video with an avatar of the agent.
Each channel shapes what the agent can show and how fast it must respond, so design your agent's output for the channel it lives in.
Advertisement
Choosing Protocols for Your Stack
Not every agent development tool supports every protocol. The Vercel AI SDK supports MCP natively, for example. To go further, such as making your agents callable by other organisations' agents, you can add A2A with its official SDKs or with Google's Agent Development Kit. Start with MCP for tools and data, then add A2A when you need agents to collaborate across systems.
Series Wrap-Up
Across this series we have covered what agents are (Part 1), the tools to build them (Part 2), prompts (Part 3), context, memory and tools (Part 4), the agent loop and subagents (Part 5), workflows, RAG and events (Part 6) and, finally, protocols. Together these are the building blocks of reliable AI agents. The best way to learn them is to build: start with one agent, one clear goal and a few well-described tools, and grow from there.
References
- Framework, Use Cases and Requirements for AI Agent Protocols (IETF draft)
- AI agent development frameworks (arXiv:2512.01939)
- A Survey of AI Agent Protocols (arXiv:2504.16736)
- Model Context Protocol
- Agent2Agent (A2A) Protocol
- Universal Commerce Protocol
- Agent Payments Protocol (AP2)
- A2UI
- AG-UI
- Vercel AI SDK documentation
Building AI Agents: the series
- Part 1: What Are AI Agents?
- Part 2: Agent Development Tools, Vercel AI SDK and Providers
- Part 3: Prompts: Text, Instructions and Messages
- Part 4: Context, Memory and Tools
- Part 5: The Agent Loop, Skills and Subagents
- Part 6: Workflow Patterns, RAG and Events
- Part 7: Agent Protocols: MCP, A2A and Beyond (you are here)
Cover photo: Conny Schneider on Unsplash
Chainbook builds intelligent business systems for African professionals, and AI agents are at the heart of how we build them.