Picture a procurement manager asking her assistant, "Find three suppliers who can deliver 200 of these by the 15th, and get quotes." The assistant does not browse your website. It contacts agents that represent suppliers, and your business either has something that answers or it does not.


That picture is partly here and partly still being designed. This post explains, as of November 2026, the two protocols most often mentioned, MCP and A2A, with what the sources say about each, then lists the open questions and what a small business can do now.


Two protocols, two jobs


MCP: connecting an AI application to tools and data


The Model Context Protocol describes itself as "an open-source standard for connecting AI applications to external systems": data sources such as files and databases, tools such as search engines and calculators, and workflows. Its documentation compares it to a USB-C port for AI applications: one standard connector instead of a custom cable for each device.


Anthropic introduced MCP in late 2024. On December 9, 2025, it announced that it was donating MCP to the Agentic AI Foundation, a fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI, with support from Google, Microsoft, AWS, Cloudflare and Bloomberg. In that announcement Anthropic reported more than 10,000 active public MCP servers and support in products including ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code. Those are Anthropic's numbers.


In plain terms: MCP lets someone's assistant use *your* systems the way a person would use an app: look up an order, check stock, read a policy.


A2A: agents talking to agents


The Agent2Agent protocol, announced by Google on April 9, 2025 with more than 50 launch partners, is "an open protocol enabling communication and interoperability between opaque agentic applications," in the words of its documentation. It was later donated to the Linux Foundation, and its site describes it as vendor-neutral and licensed under Apache 2.0; the current specification is version 1.0.


Its main ideas, from Google's announcement and the project documentation:


  • Agent Card: each agent publishes a machine-readable description of what it can do, at a standard web address, so other agents can find and choose it.
  • Tasks: one agent sends a task to another, which can finish at once or take hours or days, with status updates. The output of a task is an "artifact."
  • Opaque agents: agents work together without sharing their internal memory, tools or logic, which protects each side's security and intellectual property.
  • Existing web standards, and authentication built in.

The project's own comparison is simple: MCP connects an agent to tools, APIs and data; A2A connects an agent to other agents. They are meant to be used together.


A buyer's agent meets a seller's agent


Put the two together and a transaction looks like this:



Every step in that chain raises a question that the protocols alone do not answer.


Open questions


Identity: who is this agent, and who does it speak for? An Agent Card says what an agent claims it can do. It does not by itself prove it acts for a real, authorized company or person. Authentication methods are supported, and the practical state of trust between strangers' agents is still developing.


Authority and consent: what is it allowed to ask for or agree to? A buyer's agent may be allowed to request a quote but not to accept terms or pay. Someone has to define and check those limits. On your side, so should you: what may an outside agent see, and what may it commit you to?


Commerce and payment. Separate efforts target this. Google announced the Agent Payments Protocol (AP2) in September 2025, built around signed mandates that record what a user authorized an agent to do, with more than 60 partners. OpenAI and Stripe released the Agentic Commerce Protocol in September 2025, with Shared Payment Tokens that let a user delegate payment to an agent without exposing card data. Coverage describes these as overlapping and still maturing. We would not treat either as settled.


Accountability. If your agent gives a wrong quote to another agent, you are still the party who made the offer. The Air Canada chatbot decision, discussed in our post on reliability, is an early sign that businesses can be held to what their assistants say.


Abuse and noise. Anyone can create an agent, so expect spam, probing and fraudulent requests. Rate limits, keys and review will matter.


Privacy. If a buyer's agent sends personal information to yours, privacy law applies on your side. Ask your privacy lead or lawyer.


What it means for a small business


For SMBs, this has two sides.


Being found. A buyer's agent will favour businesses it can understand and query: clear products, prices, availability, terms and policies, in a structured form. That is the same case we made in make your company queryable. If your answers exist only in a PDF or a person's head, an agent cannot use them.


Answering well. If a customer's agent asks for a quote, you want a fast, accurate, bounded reply, not a blank. A bounded reply means: the right price from your price list, within rules you set, and nothing committed until a person approves it.


What to do now


1. Make your facts machine-readable and current: prices, lead times, service areas, hours, terms. One source, with an owner.

2. Decide what an outside agent may do. Read-only information, a draft quote, a booking request? Never a signed commitment without a person.

3. Keep a person in the approval seat for quotes, prices, payments and anything binding.

4. Keep a log of every agent request and response.

5. Use keys and limits. Authenticated access, rate limits, and the ability to revoke.

6. Watch the standards, do not bet on one. MCP and A2A are open and under the Linux Foundation; payment protocols are younger.

7. Start small. One use, such as answering availability and price questions, before anything transactional.


How this shows up at Humonix


Factually: Humonix exposes Hue through four doors that use the same tools and the same rules: Hue in the app, an MCP server for AI assistants, a REST API for your own code, and A2A for other agents. Hue publishes an Agent Card; another agent that holds a company key or authorizes through OAuth can send it a task and get a result back. Each client's access is separated by company, and agents draft: anything that spends money, changes a price, sends or goes public waits for a named person's approval, recorded on a ledger. Listing in public A2A directories is not done yet; today an agent finds Hue from its Agent Card address.


As with every feature, check the module page for what is Live, Early access or Coming. See MCP and API, the harness, Know and The Agentic Enterprise.


Sources



This post is general information, not technical or legal advice.