A2A (Agent2Agent)

A2A covers the case MCP does not. MCP hands an agent a set of tools; A2A hands a task to another agent. Each participant publishes an agent card at a well-known URL describing its skills, endpoints and authentication, and an orchestrator reads those cards to decide where to route work. The task then has a lifecycle of its own — submitted, working, input-required, completed — rather than a single request and response.

It is the least settled of the standards tracked here, and coverage varies sharply between implementations: publishing a card is not the same as accepting a task, and accepting a task is not the same as completing one.

A2A is not a payment rail, not an identity system, and not a guarantee that the agent on the other end is competent. Sato Hub probes for a published agent card and records whether one answered; nothing beyond that is claimed, and Sato Hub deliberately does not publish an agent card of its own, because publishing one is a claim we would break the first time an orchestrator sent a task.