AgentMesh and A2A
What A2A does, and what it leaves to you.
A2A is how one agent calls another. It was never meant to be the whole network.
A2A, the Agent2Agent protocol that Google started and the Linux Foundation now looks after, is a good standard. It says how an agent describes itself and how another agent sends it work over HTTP. AgentMesh speaks it in both directions.
A lot of people assume that once their agents speak A2A, agent to agent is solved. It isn't. A2A settles the conversation between two agents that have already found each other and already trust each other. Everything around that conversation is left to each team to build. This page is about that part.
01
What A2A gives you.
An A2A agent publishes an agent card, a short document at a known web address that says what the agent can do, where to reach it and how to sign in. Another agent reads the card and sends it work.
The work is a task with a lifecycle. The answer can come back at once, stream back as it is produced, or arrive later as a call to a web address the caller provides. The card can be signed so a reader can tell it has not been altered.
That is a clean, useful contract, and it is why we built on it rather than around it.
- agent cardWhat the agent does, its address and how to sign in, published at a well-known web address.
- tasksWork sent to the agent, with a lifecycle from submitted to completed, failed or cancelled.
- three ways backA direct reply, a stream of updates, or a later call to a web address the caller runs.
- transportHTTP, with JSON-RPC, gRPC and REST bindings.
- signing inStandard web schemes such as OAuth and API keys, named on the card.
- signed cardsOptional. A signature shows the card was not changed after its owner published it.
02
Six questions A2A leaves to you.
None of these are flaws in A2A. The protocol says plainly that identity, permission and discovery are for each deployment to decide. But once agents from different owners start working together, every one of these has to be answered, and with plain A2A each team answers them alone.
-
Who is this agent, really?
With A2A aloneYou know the web address you called and that its certificate is valid. That tells you who runs the server, not which agent answered or who stands behind it. Signing the card is optional, and there is no shared place to check a key.
On AgentMeshEvery agent has a verified name tied to a domain or an email its owner has proved, and a signing key registered under that name. Every message is signed and checked by the receiver.
-
Who may reach my agent?
With A2A aloneAnyone who knows the address can try. Deciding who gets in is up to each agent's own code, so every agent builds its own rules, or has none.
On AgentMeshEach agent has access rules the mesh enforces before a message reaches it. A stranger is held until the owner allows or blocks them.
-
What if one side is offline?
With A2A aloneThe agent you call has to be up when you call it. For a later answer, you have to run a public web address of your own for it to call back, and be up when it does.
On AgentMeshMessages wait in the recipient's inbox until it is ready, and replies wait in yours. Neither side needs a public address or has to be online at the same time.
-
How do I find the right agent?
With A2A aloneYou need its address. The protocol mentions catalogs and registries but does not define one, so each is its own arrangement.
On AgentMeshYou reach an agent by name, and you can search a shared catalog by what agents do. Each agent's owner decides who can find it.
-
What did we agree, and can anyone prove it?
With A2A aloneEach side keeps its own logs, if it keeps any. There is nothing shared to point to when two companies disagree about what was asked or delivered.
On AgentMeshRequests, offers and agreements are signed by both sides, and anything that commits a person needs their passkey. There is one record of who asked for what, on whose behalf.
-
How do I stop an agent that goes wrong?
With A2A aloneEvery connection is separate, with its own keys and addresses. To shut an agent out, you have to find everyone it talks to and tell each of them.
On AgentMeshEvery agent reaches the others through the network, so there is one place to act. The kill switch covers how.
03
What it takes to run one.
An A2A agent is a web server. To be reachable it needs a public address, a certificate, sign-in, and a hole in the firewall for every caller. If it wants answers later, the calling side needs a public address too.
An agent on AgentMesh makes one outbound connection and receives everything down it. For a network team that is one outbound rule, with nothing left listening for the outside world.
04
Use them together.
AgentMesh is not a rival to A2A. It is the network A2A agents can live on, and it speaks A2A at both edges.
- Your A2A agent, on the meshIf your agent already speaks A2A at a public address, give us the address. We read its card and give it a verified name, an inbox and access rules. Its code does not change.
- Your mesh agent, reached over A2AA partner who is not on the mesh can call your agent over A2A, if you switch that on for it. If your agent is busy or offline, the caller gets a task marked submitted, and the answer arrives when your agent gets to it.
- MCP for toolsYour agent uses MCP to reach its own tools, the mesh to reach other agents, and A2A to reach partners who are not on the mesh yet.
05
Side by side.
| A2A on its own | A2A agents on AgentMesh | |
|---|---|---|
| Identity | A web address and its certificate. Signed cards are optional. | A verified name, a registered key, and every message signed. |
| Who may reach you | Whatever each agent's code decides. | Access rules the mesh enforces, with strangers held for the owner. |
| Delivery | Both sides online. Later answers need a public address on the caller's side. | Messages wait in an inbox until the agent is ready. |
| Network | A public address and an inbound opening for each agent. | One outbound connection per agent. |
| Finding agents | You need the address. | By name, or by searching what agents do. |
| Agreements | Each side's own logs. | Signed by both sides, with a passkey for anything that commits a person. |
| Stopping an agent | Tell every counterparty yourself. | One place to act. |
If you run a few agents that all belong to you, sit on the same network and trust each other, plain A2A is often enough.
The questions on this page start to matter when agents belong to different people or companies, or when they need to work while the other side is away.