MCP, A2A, AG-UI: The Three Protocols of the Agentic Stack

For two years, agent demos looked the same: a chat box, a handful of hard-coded tools, and a brittle chain of prompts. That era is ending. What replaced it is not a better model — it is a protocol layer. Three open standards now define how agents talk to the world: MCP for tools and data, A2A for other agents, and AG-UI for the humans who still have to trust the outcome.

MCP: how an agent uses tools

The Model Context Protocol, donated to the Linux Foundation's Agentic AI Foundation, is the USB of agent tooling. A host connects to one or more MCP servers; each server exposes tools, prompts, and resources behind a stable schema. The model never executes anything on its own — it proposes a call, your application runs it, and the result comes back as structured context.

That boundary matters. Once tools are schema-defined and host-enforced, you can swap models, add OAuth, log every call, and ship the same capability to Claude, GPT, Gemini, or an open-weight runtime without rewriting the product. In 2026, MCP support is table stakes for any serious agent platform — not because the protocol is fashionable, but because custom tool glue does not survive a second client.

A2A: how agents talk to each other

MCP solves agent-to-tool. It does not solve agent-to-agent. That is A2A: Google's Agent-to-Agent protocol, also under the Linux Foundation, with IBM's ACP folded into it. A2A gives independent agents a way to discover each other, publish capabilities, and negotiate tasks across trust boundaries — your booking agent talking to a partner's hotel agent without either side owning the conversation.

The practical consequence is organisational. Multi-agent work inside one company can still be a private harness. Cross-company work needs a shared language for identity, capability cards, and discourse. If your product will ever be called by someone else's agent — or call theirs — A2A is the protocol that keeps that relationship from becoming a one-off integration.

AG-UI: how agents talk to people

The third leg is the one most teams still skip: Agent-User Interaction. AG-UI standardises how agent state, tool execution, and human decisions stream into a frontend. It is what turns a black-box run into a shared surface — a plan the user can see, an approval they can grant, a hand-off they can take.

Without AG-UI (or an equivalent contract), every product invents its own streaming events, approval widgets, and error states. That is fine for a prototype. It is lethal once you have three surfaces — web, mobile, voice — and a compliance team asking where the human was in the loop. Generative and streaming interfaces are not a visual flourish; they are the UX of autonomy.

One stack, three jobs

Think of it as a division of labour. MCP is how agents act on systems. A2A is how agents collaborate across organisations. AG-UI is how agents remain legible to people. Ship all three and you have an agentic product. Ship only MCP and you have a tool-using chatbot. Ship only A2A and you have a network no one can supervise. Ship only a pretty chat UI and you have a demo.

At Lopezi we treat these protocols as design material, not just plumbing. An MCP tool is a product decision about what the agent may do. An A2A card is a product decision about what partners may ask. An AG-UI stream is a product decision about what humans need to see before they trust the next step. The engineering follows from those choices — and the choices are what separate durable systems from expensive demos.

What to ship first

Start with MCP on the one workflow that already earns its keep. Add AG-UI so that workflow can ask for approval without a custom socket protocol. Introduce A2A when a second organisation — or a second internal agent team — needs to call you without a private webhook. Measure each step with evals and cost gates, not with "it felt smarter."

If you want a concrete starting point, our agent-readiness scan covers the discovery half of this stack — robots.txt, Markdown negotiation, MCP cards, OAuth, Agent Skills — and maps the gaps to a prioritised plan. Book a call at calendly.com/lopezi/ and we will tell you which of the three protocols you are missing before you write another prompt.