0:00Last time we saw that a client reads an agent's card before it sends anything. Now let's follow what it sends, and what comes back.
0:07With a tool, the exchange is a call. Our agent asks, the server answers, and that is the whole exchange.
0:14What comes back is a value, the result, and there is nothing else to keep track of.
0:19With another agent it starts the same way.
0:21The client sends the agent a message. That is the request, in words.
0:26And what comes back is one of two things.
0:29For something simple, the agent can answer with a message of its own, and that is the whole exchange too. For real work, it creates a task.
0:37A direct message ends the exchange there. A task carries on.
0:41A task is a piece of work with a life of its own. The reply is the task itself, and everything about that work is inside it. Let's open one up.
0:48Every task has the same fields, whatever the agent behind it does.
0:52Two ids first. The task id names this piece of work, and the context id names the conversation it belongs to. We come back to the difference in a minute.
1:01Status is where the work stands right now.
1:03A state, and often a message from the agent about what it is doing.
1:07And when that was. In one run of the stock analyst we build in section two, the state went from submitted to working to complete.
1:15History is the messages exchanged so far, ours and the agent's.
1:18Each one says which side sent it, and carries a list of parts.
1:22Artifacts are the deliverables, the work the client asked for. The stock analyst's is a short brief, and there can be more than one.
1:28Each has a name, and its content is again a list of parts.
1:31So parts are the unit of content on both sides.
1:34A part is a piece of text, a file, or structured data. That is the whole object. Now let's follow one task from start to finish.
1:44Our agent sends the request, and the agent creates a task for it.
1:47It starts out submitted. The agent has the work and has not begun.
1:52Then it is working. The agent is fetching prices, deciding what to say, and writing, and that takes a while.
1:58The client sees none of that. It can ask for the task again and read the status, or watch updates arrive as they happen. Both come in section two.
2:07When the brief is written, it goes into the task's artifacts.
2:11And the task comes back to the client with the brief inside it.
2:14Completed is a final state. This task is done.
2:17Not every task runs straight through. If the agent needs something from the client, it stops and asks.
2:23Our stock analyst does this when the request names no period. It asks which period the brief should cover, and the task waits.
2:32The client answers on the same task.
2:34Nothing was thrown away. The question and the answer are both in the task's history.
2:38And the task is working again, then complete, as before.
2:42The same door exists for authorization. If the agent needs permission to go on, the task stops, the client sorts it out, and the work resumes. That one comes up in section three.
2:54A task can also end without a result. The work failed, the client canceled it, or the agent turned it down.
3:00Whichever way it ended, a task that has ended stays ended. A follow-up is a new task, and let's see where that one goes.
3:08The remote agent keeps the tasks it has been given.
3:11The first is the brief we just followed. It has ended, so it takes no more messages.
3:15So when our agent asks a follow-up question, that is a new task, with a new task id.
3:21Both carry the same context id, so the agent can treat them as one conversation.
3:26That is the difference between the two ids. The task id names one piece of work, and the context id names the conversation the work belongs to.
3:33The answer comes back on the new task. Now let's put the three ideas of this section side by side.
3:39Three ideas, and they fit together.
3:42An agent publishes what it can do, and any program can read that before sending it anything.
3:47Whoever reads it and sends the work is the client, and the work goes over as a message.
3:51And everything about that work, from where it stands to the finished deliverable, travels in the task.
3:57Next time we put all of it to use.
3:59A small client of our own, and an agent somebody else wrote.

The Task

Created by Abhyuday DesaiUpdated 4 Oct, v2Runs 4 min 9 sec

0 views

About this video

In Episode 03 of A2A Explained, we look at what travels between two agents. A tool call returns a value; a message to an agent returns either a direct message or a task, and the task is where the protocol earns its keep.

You'll learn how to:

Read a task: its id, a status that moves over time, a history of messages made of parts, and artifacts as the work products
Follow the lifecycle as a path: submitted, working, completed, with input-required and auth-required as side doors and failed, canceled and rejected as exits
Tell the two ids apart: the task id names this piece of work, the context id groups related work into one conversation
See one concrete instance of each noun from a captured run of the analyst agent
Keep the rule that closes the section's theory: the card offers, the client sends, the task carries

This video is for developers who know MCP's vocabulary. The task is lived, not just described, in Episodes 06 and 07.

What changed

v204 Oct 2026abhydesai4 min 9 secd7f62ff

Re-recorded with the closing attribution card, which the first cut left out.

Same chapters and transcript. 5 sec longer.

0 views0 likes on this version

v102 Oct 2026abhydesai4 min 4 sec484dddd

First published version.

0 views0 likes on this version

Watch v1