The agent world has an acronym problem: MCP, A2A, UCP, AP2, A2UI, and AG-UI, just for starters. In March 2026, Google published a guide that cuts through them by building one agent, a restaurant supply-chain assistant, and adding protocols until it actually works. Here is that map in business terms, plus the one file every website will need and the single line in the guide an owner should not skip.
What Google announced
The guide at developers.googleblog.com/developers-guide-to-ai-agent-protocols is a teaching exercise. It starts from the weakest possible version of an agent and builds up:
We'll start with a bare LLM that hallucinates everything, then add protocols one by one until it can check real inventory, get specialist quotes, place orders, authorize payments, and render interactive, streaming dashboards.
By the end, six protocols each own one job:
MCP connects agents to tools and data. A2A connects agents to other agents. UCP standardizes commerce. AP2 handles payment authorization. A2UI defines what to render. AG-UI defines how to stream it.
The detail that reaches past developers is how agents find each other. It is not a private directory. It is a file at a fixed, public location on a domain:
Each A2A agent publishes an Agent Card at a well-known URL (
/.well-known/agent-card.json) that describes its name, capabilities, and endpoint.
In plain English
Strip the acronyms and the six protocols answer six ordinary questions. How does an agent reach your data and tools? MCP. How does it talk to other agents? A2A. How does it buy something? UCP. Who approved the spending? AP2. What should the screen look like? A2UI. How does that screen stream back in real time? AG-UI. You do not need all six to matter to a business. The two that decide whether an agent can find and understand you are the first pair.
Notice the pattern under the discovery step. The Agent Card lives at /.well-known/agent-card.json, and the commerce protocol uses the same idea with a catalog at /.well-known/ucp. A "well-known" path is just an agreed address every compliant agent already knows to check, the way browsers know to look for a favicon. It turns your domain into a machine-readable front door: an agent arrives, reads the file, and learns what you offer and how to work with you without a human ever introducing the two.
Why this matters for your business
Google closes the guide with advice aimed at developers that reads just as well for owners: adopt standards early. The patterns these protocols establish, discovery through well-known URLs and typed request-and-response schemas, are how your business stays compatible with a fast-growing ecosystem of agents and tools. Publish the well-known files now and every future compliant agent can find you. Skip them and you are absent from the front door agents are already trained to check.
The opening line of the guide is the other half of the warning. A bare LLM "hallucinates everything." A chatbot bolted onto a site with no connection to real data will confidently invent prices, policies, and product details, because guessing is all it has. The fix is not a better personality. It is grounding: wiring the agent to real, verified content so its answers come from your business rather than from its imagination.
Where Trinzik fits
We ship both halves of that front door. The websites we build publish the well-known discovery layer, the Agent Card and the machine-readable files agents look for, so a business is present at the address the standard tells agents to check. And we pair that discovery with grounding, so the agent answering for a business draws only from that business's verified content and returns nothing it cannot support. Our AI agent work is built to fail closed: when it does not have a grounded answer, it says so rather than inventing one.
That combination is the difference between a site an agent can find but not trust and one it can rely on. The technology underneath is standard protocols plus disciplined grounding, not prompt guesswork.
If you want your domain readable at the address agents already check, and answering only from facts, reach out and we will map it with you.
Questions this raises
What do MCP, A2A, UCP, AP2, A2UI, and AG-UI each do?
Per Google's March 2026 developer guide, each of the six agent protocols owns one job: MCP connects agents to tools and data, A2A connects agents to other agents, UCP standardizes commerce, AP2 handles payment authorization, A2UI defines what an agent renders, and AG-UI defines how that rendering streams back to the user in real time.
What is the well-known agent-card.json file?
The agent-card.json file is a JSON document that every A2A agent publishes at the fixed, public address /.well-known/agent-card.json on its domain. It describes the agent's name, capabilities, and endpoint, so any compliant client agent can discover what an agent does and how to reach it without a developer wiring the connection in advance.
Does the Universal Commerce Protocol (UCP) use the same discovery pattern as A2A?
Yes. Just as an A2A agent publishes its Agent Card at /.well-known/agent-card.json, the Universal Commerce Protocol standardizes the shopping lifecycle through a similar well-known discovery pattern, and it remains usable whether the underlying connection runs over REST, MCP, A2A, or browser-based Embedded Protocols.
Sources
- Developer's Guide to AI Agent Protocols · Google for Developers Blog, March 18, 2026
- Model Context Protocol (MCP) documentation
- A2A protocol specification
- Universal Commerce Protocol (UCP) site