Skip to main content

Agent Protocols & Agentic Commerce

Agent protocols are shared rules that let AI agents connect to tools, talk to other agents, sign in to businesses and pay for things, and each protocol does one of those jobs rather than all of them.

Advanced12 min readv1.0Updated Nov 9, 2026
AI-assisted content — reviewed by the author, but verify important details independently

Visual SummaryClick to explore

Learning Objectives

  • Explain why agents need shared protocols instead of one-off integrations.
  • Place MCP, A2A, the Personal Agent Protocol, UCP, ACP and AP2 on one map by the job each does.
  • Trace how an agent finds, checks out and pays for a product on your behalf.
  • Explain how mandates and payment tokens stop an agent from spending more than you allowed.
  • Identify the main security risks of agents that can act and pay.

Why This Matters

In AI-019 you met MCP, the standard way for an agent to use tools. In AI-033 you saw how agents inside one system pass messages. The next step is harder: agents from different companies working together, and agents spending real money. A shopping assistant built by one company has to find products from thousands of shops, check out, and pay, without every shop building a custom connection for every assistant.

That is what this wave of protocols is for. Since late 2025 there has been a rush of them: MCP and A2A for connecting, new standards for agents signing in to businesses, and several competing ones for checkout and payment. If you build products, sell online or lead an AI team, you will be asked which ones matter. This lesson gives you a map so the acronyms stop blurring together.

Everyday Analogy

Think about how international travel works. Your passport proves who you are. A visa says what you are allowed to do in a country. A shared language lets you talk to the people at the counter. A bank card network lets you pay in any shop without the shop knowing your bank.

No single document does all four jobs. Agent protocols work the same way. One protocol handles "how do I use this tool", another "how do I talk to that agent", another "who am I acting for and what may I do", and another "how do I pay without handing over the card number".

The Protocol Map: One Job Each

The easiest way to keep these straight is by the question each one answers.

Question Protocol Started by Status (late 2026)
How does an agent use tools and data? MCP (Model Context Protocol) Anthropic, 2024 Widely used. Hosted by the Agentic AI Foundation
How do agents from different companies talk and hand off work? A2A (Agent2Agent) Google, 2025 v1.0. Hosted by the Agentic AI Foundation
How does my personal agent sign in to a business and what may it do there? Personal Agent Protocol Sierra and Meta, October 2026 Early draft
How does an agent browse, check out and handle orders with a shop? UCP (Universal Commerce Protocol) and ACP (Agentic Commerce Protocol) UCP: Google with large retailers. ACP: OpenAI and Stripe Competing, both in use
How do we prove the person really approved this payment? AP2 (Agent Payments Protocol) Google, 2025 Now under neutral governance

The Agentic AI Foundation, part of the Linux Foundation, now hosts MCP, A2A, AGENTS.md and several other agent projects. That matters because it means no single company controls them, which makes businesses more willing to adopt them.

✓ Tip

When you meet a new agent protocol, ask one question first: which job on this map does it do? Most new announcements fill one row, and many are competing for the same row.

Layer 1: Tools and Agents (MCP and A2A)

MCP connects one agent to tools: a database, a calendar, a shop's product search. The agent is in charge and the tool simply answers. You covered this in AI-019.

A2A connects an agent to another agent. The difference matters. A tool does exactly what it is asked. Another agent has its own goals, its own reasoning and its own owner. A2A gives each agent a public "Agent Card" describing what it can do, and a standard way to send it a task, follow its progress and receive the result.

A simple rule of thumb: MCP is for using things, A2A is for working with someone.

Layer 2: Identity and Permission (Personal Agent Protocol)

When your agent contacts your bank or a shop, the business needs to know three things: is this really an agent acting for this customer, what is it allowed to do, and can the customer take that permission back?

Today agents often fake it, by logging in with your password or by clicking through a website as if they were you. Businesses cannot tell the difference between you and a bot, and the bot has full access to everything.

The Personal Agent Protocol, announced by Sierra and Meta in October 2026 with retail and payment partners, proposes a front door for agents. It is built on OAuth, the same system behind "Sign in with Google". You grant your agent a specific permission, such as "may read my orders" or "may change a delivery address", and the business sees an agent with a limited pass rather than you. The business also chooses how it receives agents: through its website, its APIs, or its own customer service agent.

It is still an early draft. Questions like how quickly a permission can be revoked are still open. Treat it as the direction of travel, not a finished standard.

Layer 3: Shopping and Checkout (UCP and ACP)

Once an agent can talk to a shop, it needs a shared way to search products, build a cart, check out and track the order. Two protocols are competing here:

  • UCP (Universal Commerce Protocol), launched by Google in January 2026 with large retailers, covers the whole journey from product discovery to after-sales. Shops can expose it through their own APIs, through MCP, or through A2A.
  • ACP (Agentic Commerce Protocol), from OpenAI and Stripe, focuses on checkout inside an AI assistant, using a payment token so the agent never sees the card number.

For a shop, the practical question is which AI assistants your customers use, because each assistant favours certain protocols. Many large retailers are supporting more than one while the market settles.

Layer 4: Proving the Payment Was Approved (AP2)

Paying is the riskiest step. If an agent can be tricked, it can be tricked into buying. AP2 tackles the core problem: proving that a real person approved this exact purchase.

It does this with mandates, signed digital records of what you agreed to:

  1. Intent mandate: what you asked for, with limits. "Buy two concert tickets, under 8,000 rupees total, before Friday."
  2. Cart mandate: the exact basket the agent found, which you (or your rules) approve.
  3. Payment mandate: passed to the payment network, so the bank can see an agent was involved and that the purchase matches what was approved.

Because the mandates are signed, a merchant or bank can check them later. If something goes wrong, there is evidence of what the person actually approved, which matters a great deal when disputes and refunds happen.

ℹ Note

Payment card networks are building their own agent programmes too, and some partners belong to more than one camp. Expect some of today's protocols to merge or fade. The four jobs on the map will stay the same even if the names change.

Worked Example: Reordering Printer Ink

You tell your assistant: "Reorder the ink for my office printer, under 2,000 rupees, delivered this week."

  1. Intent recorded. Your assistant creates an intent mandate with the product, price limit and deadline (AP2).
  2. Shop found. It searches retailers that support a commerce protocol and finds three offers (UCP or ACP, often reached through MCP).
  3. Signed in as your agent. At your usual shop it signs in with a limited pass, "may place orders up to 2,000 rupees", rather than your password (Personal Agent Protocol).
  4. Second opinion. It asks the shop's own support agent whether that cartridge fits your printer model (A2A).
  5. Cart approved. The basket comes to 1,780 rupees, inside your limit, so your rules approve it and a cart mandate is signed (AP2).
  6. Paid with a token. Checkout uses a one-time payment token. Neither the agent nor the shop's AI sees your card number (ACP or UCP plus the payment network).
  7. Order tracked. Delivery updates come back through the same commerce protocol.

Notice that no step needed your password, your card number or a custom integration between this assistant and this shop.

Security: New Doors, New Risks

Agents that can act and pay are attractive targets. The main risks:

  • Prompt injection at checkout. A product page hides text telling the agent to buy a more expensive item or to send your address elsewhere (see AI-081). Mandates help, because the purchase must still match what you approved.
  • Over-broad permissions. An agent given "full account access" is a disaster waiting to happen. Grant the smallest permission that does the job, and set spending limits.
  • Fake agents and fake shops. Anyone can publish an Agent Card. Agents need to verify who they are dealing with, just as you check a website's address before paying.
  • Unclear liability. When an agent buys the wrong thing, who pays? Signed mandates are the evidence that settles these arguments.

⚠ Warning

Never give an agent your real password or card number to "make it work". If a service only works that way, it is not ready for agents yet. Wait for proper agent sign-in and payment tokens.

Common Misconceptions

  • "MCP and A2A compete." They do different jobs. MCP connects an agent to tools, A2A connects agents to each other, and many systems use both.
  • "One protocol will do everything." Each protocol covers one layer. A full purchase uses several together.
  • "Agentic commerce means agents can spend freely." The whole point of mandates, tokens and scoped permissions is to keep spending inside limits you set.
  • "These standards are settled." Several are drafts or still competing. Build on the stable ones (MCP, A2A) and keep commerce integrations easy to swap.

Try It Yourself

  1. Pick one online purchase you made recently. Write down which of the four jobs (tools, talking, identity, payment) an agent would need to do it for you.
  2. Write an intent mandate in plain English for a purchase you would actually delegate. What limits would you set on price, timing and substitutes?
  3. If you run a website or shop, list what an agent would need from you: product data, stock levels, a checkout it can use, and a way to know who it is acting for.

Key Takeaways

  • Agent protocols each do one job: tools (MCP), agent-to-agent work (A2A), identity and permission (Personal Agent Protocol), shopping (UCP, ACP) and payment approval (AP2).
  • MCP and A2A are now hosted by the Agentic AI Foundation under the Linux Foundation, so no single company controls them.
  • Agents should sign in with limited, revocable permissions, never your password.
  • Signed mandates prove what a person approved, which keeps agent spending inside limits and settles disputes.
  • The commerce and identity standards are young and competing, so understand the jobs, not just the names.

Glossary

Agent Protocol
A shared set of rules that lets AI agents connect to tools, other agents, businesses or payment systems without a custom integration for every pair.
A2A (Agent2Agent)
An open protocol, started by Google, that lets agents from different organisations discover each other through Agent Cards, send tasks, follow progress and return results. Hosted by the Agentic AI Foundation.
Agent Card
A public description an A2A agent publishes, listing what it can do and how to reach it, so other agents can decide whether to work with it.
Agentic AI Foundation (AAIF)
A Linux Foundation fund, started in December 2025, that hosts open agent projects including MCP, A2A and AGENTS.md so no single company controls them.
Personal Agent Protocol
A proposed OAuth-based standard from Sierra and Meta (October 2026) for how a personal agent signs in to a business with limited, revocable permissions. Still an early draft.
UCP (Universal Commerce Protocol)
A commerce protocol launched by Google with large retailers in January 2026, covering product discovery, checkout and after-sales for agents.
ACP (Agentic Commerce Protocol)
A checkout protocol from OpenAI and Stripe that lets an agent complete a purchase inside an AI assistant using a payment token instead of card details.
AP2 (Agent Payments Protocol)
A protocol started by Google that proves a person approved a purchase, using signed intent, cart and payment mandates.
Mandate
A signed digital record of what a person approved, such as a spending limit or an exact basket, that merchants and banks can verify later.
Payment Token
A stand-in for a card number that works for one purchase or merchant, so agents never handle the real card details.

References

Knowledge Check

8 questions