Agent-to-Agent (A2A) Transaction

An agent-to-agent (A2A) transaction is a payment or other exchange of value carried out directly between two software agents, usually AI agents, each acting on behalf of a person or business. One agent might buy data, computing time or a service from another, negotiate a price and settle the payment, with no human approving each step. Such transactions are part of the wider shift towards agentic commerce.

The abbreviation can cause confusion. In payments, "A2A" traditionally means account-to-account payment, a direct bank transfer between two accounts. In the agent context it refers to dealings between autonomous programs, and it also echoes Agent2Agent, an open protocol announced by Google in 2025 for agents to communicate with each other. A2A transactions often rely on an agent wallet or a payment credential delegated by the owner.

Imagine you ask a digital assistant to plan a trip. Your assistant contacts an airline's assistant, compares fares with a hotel's assistant, agrees prices and pays both. You never speak to anyone; the two sides' programs do the talking and the paying, then send you the confirmations.

That is an agent-to-agent transaction. The programs need a shared language to understand each other, a way to prove they are allowed to spend money for their owners, and a payment method fast and cheap enough for many small deals. Those three needs explain most of the current work in this area.

A typical transaction involves several layers:

1. Discovery and communication. Agents find each other and exchange requests using shared protocols that describe capabilities, tasks and prices.
2. Authorisation. The buying agent shows it is acting within its owner's instructions, for example through an agentic mandate or a signed statement of verifiable intent.
3. Payment. The agents settle through a card credential, a bank payment or an on-chain transfer, often in stablecoins.
4. Delivery and confirmation. The selling agent delivers the service or data, and both sides record the outcome.

Some designs use escrow or smart contracts to release payment only when delivery is confirmed, which reduces the need for trust between unknown agents.

Several industry efforts address parts of the problem. Google's Agent2Agent protocol, later placed under the Linux Foundation, focuses on how agents communicate. The Agent Payments Protocol, also led by Google with payment companies, deals with authorisation and payment for agent-led purchases, and the Agentic Commerce Protocol from OpenAI and Stripe targets checkout by agents. Card networks have launched their own programmes, and crypto projects have built on-chain approaches, including the x402 standard that lets a web service request stablecoin payment from an agent within a normal web request. Earlier projects, such as the autonomous economic agent framework developed by Fetch.ai, explored similar ideas before the recent wave of large language models.

• Machine services. Agents pay each other per request for data feeds, translations, model access or computing power, often as micropayments.
• Procurement. Business agents negotiate and place routine orders with suppliers' agents.
• Connected devices. Vehicles, chargers and sensors in the internet of things could pay for services automatically.
• Marketplaces. Platforms where specialist agents sell skills to other agents.

The choice of payment rail shapes which uses are practical. Card payments bring familiar consumer protections and chargeback rights but carry fixed fees that make very small payments uneconomic. Bank transfers are cheap in many countries but can be slow across borders and are not designed for thousands of tiny payments a minute. On-chain stablecoin transfers settle quickly and can be divided into very small amounts, which is why many machine-to-machine experiments use them, but they lack built-in refunds and require the agent's owner to manage keys or rely on a custodian. Some designs open a payment channel or prepaid balance between two agents and settle the net amount later, reducing fees for frequent dealings.

A2A transactions raise unsettled questions. If an agent buys the wrong thing, overpays or is manipulated by another agent, it is unclear who is liable and how dispute resolution should work. Fraudsters can deploy malicious agents that pose as legitimate sellers. Payment providers must still identify the people and businesses behind agents to meet anti-money laundering and sanctions rules. Machine-speed trading between agents could also amplify errors before humans notice. As of November 2026, most A2A payments are pilots or limited deployments, and standards are still converging.