AI Agent Protocols: The Missing Link for Interoperable Agents

Updated October 2026

AI agent protocols are open specifications that set how agents talk to tools, to other agents, to people, and to payment systems. In 2026 three protocols carry most traffic. MCP connects agents to tools and data, A2A connects agents to agents, and AG-UI connects agents to user interfaces.

What are AI agent protocols?

AI agent protocols are contracts for machine to machine conversation. Each one fixes the message format, the transport, how a caller finds out what the other side can do, and how both sides prove identity. Without them, every pairing of agent and tool needs its own custom connector.

A 2025 survey of the field sorted protocols along two dimensions. The first asks whether a protocol connects an agent to context such as tools and data, or connects agents to each other. The second asks whether it is general purpose or built for one domain. That split still explains the 2026 landscape well.

Protocols are not orchestrators. They carry requests and results, but they do not decide which agent does what or in what order. That planning lives in the agent framework or the orchestration layer above the protocol.

This article is the protocol reference. For coordination patterns, failure modes, and design rules, read our companion guide to AI agent communication.

How do MCP, A2A, and AG-UI compare?

MCP, A2A, and AG-UI solve different hops, so the useful comparison is by layer rather than by winner. MCP treats the other side as a capability to call. A2A treats it as a peer with its own plan. AG-UI streams what the agent is doing to the person watching.

Message format and transport

MCP sends JSON-RPC 2.0 messages between hosts, clients, and servers. A2A version 1.0 defines three standard bindings, which are JSON-RPC, gRPC, and HTTP plus JSON, and it allows custom bindings. AG-UI defines about 16 standard event types and runs over any event transport, including server sent events and WebSockets.

Discovery

In A2A each agent publishes an Agent Card listing its identity, skills, endpoints, and security schemes. Version 1.0 adds signed Agent Cards so a client can verify who published a card. MCP clients can call server/discover to get a server's versions, capabilities, and identity in one request.

State and long running work

A2A models work as a Task that moves through states such as working, input required, auth required, completed, failed, canceled, and rejected. Clients can stream updates or register webhooks for push notifications. MCP handles long jobs through an optional Tasks extension with polling and durable handles.

What is the Model Context Protocol (MCP)?

MCP is the open standard for connecting an AI application to external tools, data, and prompts. A server exposes resources, prompts, and tools. A client inside the host application calls them. Clients can also offer elicitation, which lets a server ask the user for more information.

Adoption is broad. When Anthropic donated MCP to the Agentic AI Foundation in December 2025, it counted more than 10,000 active public servers and over 97 million monthly SDK downloads. The same announcement listed ChatGPT, Cursor, Gemini, Microsoft Copilot, and Visual Studio Code as supporting platforms.

What changed in the 2026-07-28 revision

The current MCP revision is 2026-07-28. Every request now declares its protocol version, and the server accepts or rejects each request on its own. The foundation's migration guide says this lets any server instance handle any request, so plain round robin load balancing works.

The revision also deprecates four features. Roots, Sampling, Logging, and Dynamic Client Registration stay in the spec but are listed in the deprecated features registry. None can be removed before revisions released on or after July 28, 2027.

Extensions worth knowing

MCP Apps became the first official MCP extension in January 2026. It lets a tool return an interactive interface that the host renders in a sandboxed iframe. Claude, goose, Visual Studio Code Insiders, and ChatGPT supported it at launch. The spec also lists Tasks and Skills over MCP as optional extensions.

What is the Agent2Agent (A2A) protocol?

A2A is the open standard for one agent to delegate work to another agent as a peer. Neither side has to expose its prompts, tools, or memory. The client reads the remote agent's Agent Card, sends a message, and tracks the resulting Task until it ends.

Google launched A2A in April 2025 and donated it to the Linux Foundation. Version 1.0, the first stable specification, shipped in March 2026. On August 17, 2026, A2A joined the Agentic AI Foundation as a hosted project alongside MCP, goose, AGENTS.md, and agentgateway.

The Linux Foundation reported in April 2026 that more than 150 organizations back A2A. It named integrations in Azure AI Foundry, Copilot Studio, and Amazon Bedrock AgentCore Runtime. The same release cited production ready SDKs in Python, JavaScript, Java, Go, and .NET.

On security, the A2A spec lists API keys, HTTP auth, OAuth 2.0, OpenID Connect, and mutual TLS as supported schemes. Servers must reject missing or invalid credentials. The extended Agent Card, which can reveal more skills, requires authentication.

What happened to ACP, ANP, LMOS, and Agora?

The crowded protocol field of early 2025 has narrowed. IBM's ACP merged into A2A, and the other early contenders remain niche or research stage. They are worth knowing because older comparisons still list them as equals of MCP and A2A.

Agent Communication Protocol (ACP)

IBM Research built ACP for its BeeAI platform. IBM's project page now states that ACP is part of A2A under the Linux Foundation and links to a migration guide. New agent to agent work should target A2A.

Agent Network Protocol (ANP)

ANP aims at a decentralized web of agents built on decentralized identifiers. Its white paper lives in the W3C AI Agent Protocol Community Group, which launched in May 2025. Community groups publish drafts but are not the W3C standards track, so treat ANP as an early proposal.

Eclipse LMOS

The LMOS protocol builds on the W3C Web of Things model and uses JSON-LD descriptions signed with decentralized identifiers. It supports local discovery and registries. The project page calls the spec a work in progress and says it is not on the W3C standards track yet.

Agora

Agora comes from an October 2024 research paper on networks of language model agents. Agents use standard routines for frequent exchanges, natural language for rare ones, and model written routines in between. It is a useful idea for research teams rather than a production standard.

Which protocols handle discovery, identity, and payments?

A second layer of protocols sits on top of MCP and A2A. AGNTCY handles discovery and identity across vendors. A group of commerce protocols handles checkout and payment when an agent spends money for a person.

AGNTCY

Cisco open sourced AGNTCY in March 2025, and it joined the Linux Foundation on July 29, 2025. It provides agent discovery through the Open Agent Schema Framework, verifiable identity, the SLIM messaging protocol, and observability. Its directories can list both A2A agents and MCP servers, so it complements them.

Commerce and payment protocols

Google announced the Agent Payments Protocol in September 2025. It uses cryptographically signed mandates to record what the user asked for and which cart they approved. In April 2026 Google donated AP2 to the FIDO Alliance and released version 0.2 with support for purchases made while the user is away.

Stripe and OpenAI publish the Agentic Commerce Protocol under Apache 2.0, with ChatGPT as the first implementing platform. Google's Universal Commerce Protocol, launched in January 2026 with Shopify, Etsy, Wayfair, Target, and Walmart, can run over REST, A2A, or MCP. The Linux Foundation launched the x402 Foundation in April 2026 after Coinbase contributed the protocol.

How do you choose an AI agent protocol?

Choose by the hop you need to standardize, then check maturity and security for that hop. Most teams need MCP first, A2A second, and the rest only for specific cases. The table maps common needs to a starting point.

Check governance and versioning

Prefer protocols with a neutral home and a written change policy. MCP now requires deprecated features to stay supported for at least twelve months before removal. A2A clients send a version header with each request, and the spec treats an empty header as version 0.3, which helps older clients keep working.

Check security at every hop

The MCP spec requires hosts to get user consent before invoking any tool. It also says tool annotations should be treated as untrusted unless they come from a trusted server. Apply the same rule to A2A peers. Verify signed Agent Cards, pass the end user's identity through each call, and log every exchange. Our guide to data governance for AI agents covers access control and audit in more depth.

Check what the protocol does not cover

Protocols move messages faithfully, but they do not make two agents agree on what "active customer" or "net revenue" means. That agreement has to come from shared definitions in a semantic layer, ontology, or knowledge graph. Our comparison of RAG, knowledge graphs, and semantic layers explains the options. The guide to context engineering for AI agents shows how to feed those definitions to agents.

For products that package these protocols, see the best enterprise AI automation platforms.

Frequently asked questions

What are AI agent protocols?

AI agent protocols are open specifications that define how agents exchange messages with tools, other agents, user interfaces, and payment systems. They fix the message format, the transport, how capabilities are discovered, and how each side proves its identity. In 2026 the core set is MCP for tools and data, A2A for agent to agent work, and AG-UI for user interfaces.

What is the difference between MCP and A2A?

MCP connects an AI application to servers that expose tools, resources, and prompts. The server is a capability the model calls. A2A connects one agent to another agent that plans and reasons on its own, often over a long task. A2A uses Agent Cards for discovery and a task lifecycle with states such as working, input required, and completed.

Which AI agent protocol should I start with?

Start with MCP if your agent needs to reach internal tools, databases, or documents. It has the widest client support, including ChatGPT, Gemini, Microsoft Copilot, Cursor, and Visual Studio Code. Add A2A only when you need to delegate work to agents that another team or vendor owns. Add AG-UI when you build your own front end for an agent.

Is the Agent Communication Protocol (ACP) still active?

IBM's Agent Communication Protocol is no longer a separate effort. IBM Research states that ACP is now part of A2A under the Linux Foundation, and the project links to a migration guide. A different protocol, the Agentic Commerce Protocol from Stripe and OpenAI, uses the same ACP acronym. Check which one a vendor means before comparing claims.

Who governs MCP and A2A?

Both sit in the Agentic AI Foundation, a directed fund of the Linux Foundation. Anthropic donated MCP when the foundation launched in December 2025, alongside goose from Block and AGENTS.md from OpenAI. A2A joined as a hosted project on August 17, 2026. Each project keeps its own maintainers and specification process inside the foundation.

What changed in the MCP 2026-07-28 specification?

The 2026-07-28 revision makes MCP requests stateless and self-contained. Each request declares its protocol version in metadata, and a new server/discover call returns capabilities on demand instead of a startup handshake. The revision also deprecates Roots, Sampling, Logging, and Dynamic Client Registration. None of these can be removed before revisions released on or after July 28, 2027.

What is AG-UI and how does it relate to MCP Apps?

AG-UI is an MIT licensed, event based protocol that streams about 16 standard event types between an agent backend and a user facing app. MCP Apps is an official MCP extension that lets a tool return an interactive interface rendered in a sandboxed iframe. AG-UI covers the whole agent session, while MCP Apps covers interfaces attached to single tool results.

Are there protocols for AI agents that make payments?

Yes, and they sit above A2A and MCP. Google's Agent Payments Protocol uses signed mandates to prove what a user approved, and Google donated it to the FIDO Alliance in April 2026. Stripe and OpenAI publish the Agentic Commerce Protocol for checkout. Google's Universal Commerce Protocol covers discovery and checkout, and x402 moved to a Linux Foundation project.

What happened to ANP, LMOS, and Agora?

They remain active ideas but have far less adoption than MCP and A2A. ANP is developed in a W3C community group, which is not the W3C standards track. The Eclipse LMOS protocol describes itself as a work in progress. Agora began as an October 2024 research paper on a meta protocol for networks of language model agents.

Do agent protocols make agents agree on what data means?

No. A protocol guarantees that a message arrives in a known format from a verified sender. It does not guarantee that both agents use the same definition of revenue, customer, or churn. Shared definitions have to come from a semantic layer, ontology, or governed glossary that every agent reads. Without that, protocol compliant agents can still return conflicting answers.