Agent-to-agent communication

Why A2A is its own topic

A2A is not a subfield of context engineering, and it is not a subfield of grounding. It is the point where two knowledge layers meet. Each agent arrives with its own picture of the world — its own currentness, ownership, permission, and grounding. The conversation is the place where those pictures have to be negotiated, shared selectively, and verified before either agent acts on what the other said.

This is why A2A fails in the same way single-agent systems fail — stale knowledge, unowned knowledge, weak grounding — but with a second agent in the loop, the failure is harder to see and more expensive when it shows up. Two agents can both be right about their own picture and still produce a wrong joint action, because the joint action is built on an unsafe merge of two partial pictures.

What an agent needs before it can share

An agent that talks to another agent needs more than a message. It needs to know what it is sharing, in what shape, with what basis, and with what permission. The minimal set:

• The knowledge, in a shareable shape. Not the raw internal representation, not a dump. A shareable shape the receiving agent can use — the same discipline as representation, but tuned for exchange rather than internal use.
• The basis. What the sharing agent used to form what it is saying — which knowledge, from where, at what version, with what permission. The receiving agent needs the basis to verify, not just the claim.
• The scope of sharing. What the sharing agent is allowed to share. The boundary between "I know this" and "I am allowed to tell you this" is the boundary that matters most in A2A — more than the message itself.
• The currentness. How current the shared knowledge is, and how the receiving agent can check it. A2A with stale knowledge is two agents confidently wrong together.
• The verification expectation. What the receiving agent is expected to do with the knowledge — trust it, verify it, cross-check it, or treat it as a lead rather than a fact. A2A is not a trust transfer; it is a trust negotiation.

These are the same disciplines as the rest of the knowledge layer — representation, currentness, ownership, permission, grounding — but applied at the boundary between two agents rather than inside one.

What A2A standards cover

A2A standards are not a transport protocol. They are the conventions that sit on top of whatever transport the agents use, and that make the exchange of knowledge dependable. They cover:

The goal is not a universal protocol. The goal is a small set of conventions that make A2A dependable enough to use in real settings — not a research exercise, but a practical layer that sits on top of whatever the agents already do.

A2A and the knowledge layer

A2A is the clearest demonstration that knowledge is a layer and not a module. A single agent with a weak knowledge layer can fail in private. Two agents with weak knowledge layers can fail in public — and the failure is harder to attribute, harder to fix, and more expensive when it touches a joint action.

The knowledge layer is what each agent brings to the conversation. A2A is what happens when those layers meet. The standards for A2A are the disciplines that make the meeting dependable: negotiation of scope, sharing in a usable shape, and verification of the basis before action.

This is also where the agent interface and the discoverability layer meet A2A. If an agent engine can discover another agent's knowledge surface — what it knows, what it is willing to share, how to ask — then A2A is not a custom integration every time. It is a convention applied to known surfaces.

Where to start

If you are building agents that talk to other agents and you are wondering where to start with A2A, the usual sequence is:

• Pick one exchange — one thing one agent tells another, in one setting. Not the whole world of agent-to-agent communication. One exchange.
• Put the basis on what is being shared — what knowledge, from where, at what version, with what permission. Not the message. The basis.
• Negotiate scope before knowledge moves — what the sharing agent is willing and permitted to share, and what the receiving agent is expected to do with it.
• Make the receiving agent verify before it acts — not on every exchange, but on the exchanges where the joint action is expensive or sensitive.
• Expose a discoverable surface — what the agent knows, what it is willing to share, how to ask — so the next A2A exchange is not a custom integration.

The goal is not to standardize all of A2A at once. The goal is to make one exchange dependable, in a way that the disciplines can be carried to the next exchange without starting over.