Agent Communication Protocols

This unlocks the capacity to make different AI agents—built by different teams, or even different companies—work together as one system. Once you understand what a communication protocol is, you'll be able to see why modern AI systems (like the agentic workflows in your library) are designed the way they are: they are essentially conversations with rules. You'll also understand a key trade-off in AI security: every new protocol you adopt is a new trust boundary, and knowing this will help you reason about the security of any system you build or use. From here, you can go on to study multi-agent systems, where dozens of agents coordinate to solve problems, and AI security, where protocols are the surface that attackers try to exploit.

Imagine you're in a busy office. People need to share information and ask each other to do things. If everyone just shouted whatever they wanted, in whatever order, with no rules about how to phrase a request or how to confirm it was understood, chaos would result: messages would be lost, requests would be misinterpreted, and nobody would know who was responsible for what. So the office adopts a set of rules: 'When you want something, write it on a form, hand it to the correct person, and they will write back either done or why not.' That set of rules—what a message looks like, how it's addressed, what responses are possible, and in what order—is a communication protocol. It doesn't tell anyone what to say, only the shape of the conversation. In the world of AI agents, the same problem appears. An agent (a program that uses a large language model to plan and act) needs to ask another agent for information, or ask it to perform a task, and then receive the result. Without a shared protocol, each agent would have to be custom-built to talk to each other specific agent—the same as if every person in the office had to learn a separate language for every colleague. A protocol is the shared language and etiquette that lets any agent talk to any other agent, as long as they both follow the rules.

The best analogy is a postal system. When you send a letter, you don't need to know the exact person who will carry it, nor do you need to arrange for them to meet you. You write your message on paper, put it in an envelope, write the address in a standard format (street, city, postal code), and drop it in a mailbox. The postal service guarantees that if the address is right and you've paid the postage, the letter will arrive at the destination. The recipient reads it, and if they choose to reply, they do the same in reverse. The protocol here is the set of rules: the envelope size, the address format, the postage, the delivery guarantee. It doesn't matter what language the letter is written in, or what the content is—the protocol is about the wrapping and the routing. Now, AI agents work similarly. An agent sends a message (like a letter) to another agent (the address). The message is wrapped in a standard format (the protocol's envelope), which includes things like 'who is the sender', 'who is the receiver', 'what type of message this is' (a request, a response, an error), and the payload (the actual content). The receiving agent can unwrap it because it also understands the protocol. The analogy even has a breaking point: in a postal system, the protocol is designed for humans who can adapt to minor irregularities. If a letter is slightly damaged, a human post worker can still read the address. AI agents are much more literal: they adhere to the protocol exactly. If a message is missing a field or has a type mismatch, the agent will often reject it outright, causing a failure. Also, a postal system assumes the address is stable and the sender and receiver are permanent—AI agents can be created and destroyed on the fly, and the protocol must allow for that. This is a 'smart' protocol, not a dumb pipe.

An agent communication protocol is a set of rules that defines how two or more AI agents exchange messages—the structure of a message, the allowed message types, the order in which they can be sent, and the expected responses—so that the agents can cooperate without being custom-built for each other.

You already have notes on AI Agents, and you know that beneath every framework lies a message loop. This concept is about the rules that govern those messages. It connects directly to Agent Cards—those are the directory entries that describe an agent's capabilities, so that another agent knows who to send a message to and what that agent can do. An Agent Executor is the runtime that actually sends and receives messages; a protocol is what the executor uses to format and parse them. The Agent Loop is, in a sense, the whole conversation: each turn in the loop is a message exchange. The idea of a trust boundary is central here: every new protocol you adopt is a new boundary where a malicious agent could try to slip in a harmful message.