Call the endpoint from a partner client

Publish an agent as an A2A endpoint

By the end of this lesson, you can call the endpoint with the official A2A SDK and with a Google ADK agent, and continue one conversation across two questions.

Lesson 2 of 4~20 minScript: svc_client.py, adk_client.py, 02_context.sql

1. Lesson 1: Publish an agent as an A2A endpoint
2. Lesson 2: Call the endpoint from a partner client
3. Lesson 3: Accept visit requests that the database checks
4. Lesson 4: Prepare the endpoint for partners

Call the desk with the A2A SDK

In lesson 1 you called the desk with curl. A partner system uses a client library. In this lesson, your client acts as Ferrolux and calls the endpoint from lesson 1 with its API key.

The official a2a-sdk reads the agent card, selects an HTTP interface, and sends the message. The script svc_client.py wraps this in about 50 lines. Install the packages and set the endpoint URL and the key as environment variables. Do not write the key into a file:

pip install a2a-sdk httpx

export A2A_BASE=https://api.example.com/ords/demo/a2a/service-desk
export A2A_API_KEY=a2a_…        # the ferrolux key from lesson 1

The core of the script is this call. create_client fetches the card and builds the client. The httpx client sends the key in each request, including the request for the card:

async with httpx.AsyncClient(headers={"X-API-Key": key}, timeout=300) as http:
    client = await create_client(
        base, client_config=ClientConfig(streaming=False, httpx_client=http)
    )
    request = SendMessageRequest(
        message=Message(role=Role.ROLE_USER, message_id=str(uuid.uuid4()),
                        context_id=context_id or "", parts=[Part(text=text)])
    )
    async for event in client.send_message(request):
        if event.HasField("task"):
            task = event.task

Every question after the first one sets context_id to the context_id of the first task. The script takes the questions as arguments. Run it with two:

python -I svc_client.py "Which of our machines are under contract?" \
                        "When was the second one last serviced?"

A development database with a self-signed certificate

A development database often answers with a self-signed TLS certificate. The recorded runs used --insecure, which sets verify=False in httpx. Use verify=False only for a self-signed development certificate. A production endpoint has a certificate that the client can check.

Recorded output of the script, with the host of the endpoint replaced:

--- question 1: Which of our machines are under contract?
state:    TASK_STATE_COMPLETED
context:  aefe581f-3c7c-4e62-8e6b-349e7457acff
seconds:  2.9
tokens:   666 in, 84 out
Both of your machines are currently under contract:

- FX-CONV-07 — Ferrolux belt conveyor 7m
- FX-PRESS-01 — Ferrolux HP-400 hydraulic press

Both are covered under GOLD contract SC-2024-088 through 2027-04-25.

--- question 2: When was the second one last serviced?
state:    TASK_STATE_COMPLETED
context:  aefe581f-3c7c-4e62-8e6b-349e7457acff
seconds:  2.5
tokens:   1108 in, 53 out
FX-PRESS-01 was last serviced on 2026-10-17 for its annual safety inspection.

The two tasks have the same context. A client that sends the contextId of an earlier task continues the conversation. The server uses the contextId as the agent session. It maps the text of the second message to the next user message. Follow-up messages in the reference page lists the rules.

Read the tools that each question called

The answer to question 2 is correct without the history, because the model can list the machines again. To see whether it used the history, read the tools that each task called. Run @02_context.sql. For each task of Ferrolux it prints the first 8 characters of the context, the question and the tools:

select substr(t.context_id, 1, 8) as context
     , json_value(t.request_message, '$.parts[0].text') as asked
     , listagg(m.tool_name, ', ') within group (order by m.seq) as tools
  from uc_ai_a2a_tasks t
  left join uc_ai_agent_messages m
    on m.execution_id = t.execution_id
   and m.role = 'tool_call'
 where t.client_id = 'apikey:ferrolux'
 group by t.task_id, t.context_id, t.created_at, json_value(t.request_message, '$.parts[0].text')
 order by t.created_at;

The script ran after the SDK client, a control run, and the ADK client below. The control run sent question 2 alone, in a new context:

CONTEXT     ASKED                                        TOOLS
___________ ____________________________________________ _____________________________________________
6a453a33    Which of our machines are under contract?    PARTNER_LIST_ASSETS
aefe581f    Which of our machines are under contract?    PARTNER_LIST_ASSETS
aefe581f    When was the second one last serviced?       PARTNER_ASSET_HISTORY
5b29cc95    When was the second one last serviced?       PARTNER_LIST_ASSETS, PARTNER_ASSET_HISTORY
37209785    Which of our machines are under contract?    PARTNER_LIST_ASSETS
37209785    When was the second one last serviced?       PARTNER_ASSET_HISTORY

Context 6a453a33 is the curl call of lesson 1. Context aefe581f is the SDK client, and 37209785 is the ADK client below. In both conversations, question 2 called only PARTNER_ASSET_HISTORY. The model took the second machine of its own earlier answer. Context 5b29cc95 is the control: the same question without a history called PARTNER_LIST_ASSETS first, to find the machines.

Call the desk from Google ADK

The desk is a standard A2A agent. An agent framework can use it as a remote agent without any UC AI code. Google ADK wraps a card URL as RemoteA2aAgent. The script adk_client.py uses it as the root agent of an ADK runner. This is the recorded setup: ADK 2.13 on Python 3.14.

pip install "google-adk[a2a]" httpx

python -I adk_client.py "Which of our machines are under contract?" \
                        "When was the second one last serviced?"

The core of the script:

agent = RemoteA2aAgent(
    name="service_desk",
    description="The partner service desk of a field-service company",
    agent_card=base + "/.well-known/agent-card.json",
    httpx_client=http,            # sends the X-API-Key header
)
runner = InMemoryRunner(agent=agent, app_name="partner")

The local ADK runner needs no model API key. UC AI calls the model from the database with the agent’s credential. ADK keeps one session for both questions and sends the same contextId each time. ADK prints warnings about its experimental A2A support to stderr. The recorded output below shows stdout:

--- question 1: Which of our machines are under contract?
service_desk: These machines are currently under contract:

- FX-CONV-07 — Ferrolux belt conveyor 7m
- FX-PRESS-01 — Ferrolux HP-400 hydraulic press

Both are covered under GOLD contract SC-2024-088 through 2027-04-25.

--- question 2: When was the second one last serviced?
service_desk: FX-PRESS-01 was last serviced on 2026-10-17 for its annual safety inspection.

The framework found the endpoint through the card and authenticated with the header that the card names. The desk answered through the same tools and the same partner_clients lookup as for curl.

Hand a key to a partner

A partner needs a key and a row in partner_clients. You generate the key with create_api_key. The call returns the key only once. Share it through a secure channel, such as your organization’s password manager.

To rotate a key, create a key under a new client name, for example ferrolux2. Add the row apikey:ferrolux2 to partner_clients. Revoke the old key with revoke_api_key after the partner has switched. You choose the length of the overlap. Revoking a key makes it fail with HTTP 401 at once. Lesson 4 shows this.

The tasks of a client belong to its client name. A new client name starts without the tasks of the old one. API keys in the reference page covers keys and expiry.

Next lesson: Accept visit requests that the database checksAdd a booking tool, a booking prompt and four rules, and read the trace of a refused and an accepted request.