Agent protocol stack explained
A technical overview compared three agent protocol approaches — MCP, A2A and AG‑UI — to show how agent systems choose communication and integration patterns with frameworks like AWS Bedrock AgentCore. The write‑up separates layers that govern tool access, agent‑to‑agent communication and human‑facing interfaces. (dev.to)
AI agents now have three different “languages” for three different jobs: Model Context Protocol for tools, Agent2Agent for agents, and Agent User Interaction for screens. (modelcontextprotocol.io) (github.com) (docs.ag-ui.com) Model Context Protocol, or MCP, is the layer that lets a model reach outside its chat window to use a file, database, search engine, or software tool. Anthropic introduced MCP on November 25, 2024, and the specification describes it as a standard way to connect language model applications to external data sources and tools. (anthropic.com) (modelcontextprotocol.io) Agent2Agent, or A2A, handles a different problem: one agent asking another agent to do work. Google said when it announced A2A that the protocol is meant for agents built by different companies, on different frameworks and servers, to discover capabilities, exchange information, and coordinate tasks. (developers.googleblog.com) (github.com) Agent User Interaction, or AG-UI, covers the front end: the stream of events between an agent and the app a person is using. Its documentation describes an open, lightweight, event-based protocol for connecting agent back ends to chat interfaces, copilots, and other user-facing applications. (docs.ag-ui.com 1) (docs.ag-ui.com 2) The split matters because many agent systems now need all three layers at once. A customer service agent might use MCP to fetch account data, A2A to hand billing work to a specialist agent, and AG-UI to stream status updates and buttons back into a web app. (modelcontextprotocol.io) (github.com) (docs.ag-ui.com) That architecture is also showing up in commercial platforms. Amazon Bedrock AgentCore says it works with any framework and model, and Amazon’s prescriptive guidance says AgentCore Gateway can turn application programming interfaces, Lambda functions, and existing services into MCP-compatible tools for agents. (docs.aws.amazon.com 1) (docs.aws.amazon.com 2) The protocols are not interchangeable. MCP is closer to a tool adapter, A2A is a coordination channel between independent agent systems, and AG-UI is a transport for interface events such as messages, state changes, and user actions. (modelcontextprotocol.io) (github.com) (docs.ag-ui.com) The standards are also arriving from different corners of the industry. MCP came from Anthropic and now has a public specification site, A2A was contributed by Google Cloud to a Linux Foundation project, and AG-UI is maintained as an open protocol with public documentation and code on GitHub. (anthropic.com) (developers.googleblog.com) (github.com) For developers, the practical choice is no longer which single protocol wins. The emerging pattern is a stack: MCP for access, A2A for delegation, and AG-UI for presentation. (modelcontextprotocol.io) (github.com) (docs.ag-ui.com)
