A2A tasks: from submitted to completed, with room for a human
A2A tracks work as tasks with defined states, including states that pause for input or authentication. Why that matters for approvals.
In A2A, a client sends messages to an agent, and longer work becomes a task with an identifier and a state. The version 1.0 specification lists states including submitted, working, completed, failed, canceled and rejected, plus two interrupted states: input required and auth required. Tasks produce artifacts, the outputs of the work.
The specification describes A2A as async first, designed for potentially very long-running tasks and human-in-the-loop interactions. By default a send operation waits until the task reaches a terminal or interrupted state; a client can ask for an immediate return and then poll, stream or receive push notifications.
In a finance flow, the interrupted states are the natural place for control. A supplier's agent preparing an invoice dispute can stop at input required and ask for a missing document. A payment agent can stop at auth required until the customer completes strong authentication with their bank. The task history becomes part of the record of what happened.
Limits. A state name does not define who must approve, how, or what evidence is kept. "Auth required" tells the client that authentication is needed; the bank's own authentication and the mandate checks still decide whether the action is allowed. Long-running tasks also need clear expiry and cancellation, which the client and agent have to agree on.