A2A

Delegate a task to a Work IQ agent. The app plays the calling agent: it reads the agent card, sends a task with SendMessage and keeps its context for follow-up questions. It sends A2A 1.0 JSON-RPC directly; no multi-agent framework is required.

When to choose it

Choose A2A when your agent should hand a whole task to a Work IQ agent and get an answer back: Microsoft 365 Copilot, or an agent you name. The Work IQ agent does the reasoning, so no model of yours is needed, and the context carries follow-up questions.

Request path

GET /a2a/.well-known/agent-card.json
POST /a2a/  > SendMessage
Follow-up  > contextId

Who does the reasoning?

The remote Work IQ agent owns its reasoning. A calling harness is optional: use one to decide which agents should get which tasks, not just to speak the protocol.

On the wire

Every request goes to https://workiq.svc.cloud.microsoft with a delegated token and the header A2A-Version: 1.0. To address one agent, put its ID in both paths: /a2a/{agentId}/.

Read the agent card

GET /a2a/.well-known/agent-card.json HTTP/1.1
Host: workiq.svc.cloud.microsoft
Authorization: Bearer <delegated token>
A2A-Version: 1.0

POST /a2a/: send a task

{
  "jsonrpc": "2.0",
  "id": "5f0c1c2e-6a4b-4d7e-9b1f-3e8a2d6c9a10",
  "method": "SendMessage",
  "params": {
    "message": {
      "role": "ROLE_USER",
      "messageId": "b2e7d4a1-0c3f-4e58-8a96-7d1b5c2e4f03",
      "parts": [{ "text": "What did Sarah say about the launch date?" }],
      "metadata": { "Location": { "timeZone": "Europe/Berlin", "timeZoneOffset": 120 } }
    }
  }
}

The result is a task. While its status.state is TASK_STATE_SUBMITTED or TASK_STATE_WORKING, poll it with GetTask and { "id": "<task ID>" }; the app waits 750 ms between polls. A task in TASK_STATE_COMPLETED carries the answer in artifacts[].parts[].text. In TASK_STATE_INPUT_REQUIRED the agent needs more from the user: reply with the task’s contextId and taskId. Later questions send the same contextId.

Authentication

A delegated WorkIQAgent.Ask token for your approved Microsoft Entra public-client app. The route doesn’t reuse the local CLI or remote MCP sign-ins. See Authentication.

What it costs

The app calls no model of yours on this route, so it adds no model charge; a harness that adds its own model pays for that model separately. Work IQ is billed in Copilot Credits. Each SendMessage is a Work IQ Chat call, with variable credits: the agent reasons in the background and returns an answer or asks for more input. The app counts no credits for reading the agent card or polling with GetTask. See Copilot Credits.

Keep in mind

A2A-Version: 1.0 is required. Active tasks are polled with GetTask. Stop waiting in the app ends only the local wait; it doesn’t promise that the server cancels the work.

Try it in the app

1. Open the A2A tab and click Connect.
2. Enter the Directory (tenant) ID and Application (client) ID of your app registration. Agent ID is optional; leave it empty for Microsoft 365 Copilot.
3. Click Connect and finish the sign-in in your browser.
4. Ask a question. Flow shows the agent card, SendMessage and each GetTask poll.

The code

src/workiq/a2a.js sends these requests, and src/workiq/http.js adds the token and refuses any other host.

A2A quickstart Register the app, discover agents and send a task.