Create and Invoke ACP Agents¶

In this lab, you will invoke 2 agents using the ACP SDK (the Python code is provided). One agent is written using the BeeAI Framework and the other agent is written using PydanticAI. We are using 2 different frameworks to illustrate how easy it is to make any agent ACP-compatible.

Agent Overview:

Agent 1 (BeeAI Framework): "Triage Agent" – categorizes the ticket (e.g., billing, tech, complaint) and extracts key facts in a structured way (e.g., using tags, severity, sentiment).
Agent 2 (PydanticAI): "Response Agent" – takes the structured summary and crafts a tone-specific response (e.g., empathetic for complaints, efficient for billing).

Prerequisite

The following steps assume that you have already completed the pre-work section. If not, please return to the previous section before moving forward.

Steps¶

1. Open the Project Directory¶

If you don't already have the intro_acp_beeai folder open in VS Code, navigate there. Your working directory should look something like this: ~/beeai-workshop/intro_acp_beeai.

2. Install Dependencies¶

Open your terminal (either in VS Code or using your preferred terminal) and install the dependencies:

uv sync

3. Run the Ticket Triage Agent¶

In your terminal, run the ticket triage agent (defaults to run on port 8000):

uv run src/ticket_triage_agent.py

Insight

If you take a look at the code you will notice that the agent is wrapped in an @server.agent decorator. The decorator makes it an ACP agent and the metadata provides agent details for discoverability.

4. Test the Agent via Browser Interface¶

Use your browser to invoke the ticket_triage_agent using the FastAPI interface:

In your preferred browser navigate to http://localhost:8000/docs
Pull down GET /agents List Agents
Hit the Try it out button and then click Execute

Expected Results:

Under Responses:

Under Response body, the result shows just one agent named ticket_triage_agent.
Under Curl, you get the curl command that you can run in a terminal (instead of using the UI)

Try the curl command: In a new terminal window, run:

curl -X 'GET' 'http://localhost:8000/agents' -H 'accept: application/json'

Insight

This shows you that your ACP agent is discoverable on port 8000 and if you had more than one agent running on this server you could use the GET /agents command to list them.

Note

To terminate the ACP agent server press Ctrl + C in the terminal where the server is running or close the terminal.

5. Run the Ticket Response Agent¶

In a separate terminal, run the ticket response agent (defaults to run on port 8001):

uv run src/ticket_response_agent.py

Insight

Notice how the ticket_triage_agent and ticket_response_agent are running on separate ports. Agents can either have their own server and be discoverable on multiple ports or be on the same server (as you'll see in lab 2).

6. Invoke the Response Agent¶

Invoke the response agent using a curl command. Remember that the ticket_response_agent expects a formatted output from the ticket_triage_agent as its input. You'll put these together in a sequential workflow in lab 2.

In a separate terminal, run this curl command:

curl -N -X POST http://localhost:8001/runs \
  -H "Content-Type: application/json" \
  -H "Accept: text/event-stream" \
  -d '{"agent_name":"ticket_response_agent","input":[{"parts":[{"content":"{\"category\":[\"Technical\"],\"customer_name\":\"Jane Doe\",\"account_id\":\"872-55\",\"product\":\"ProPlan\",\"issue_summary\":\"ProPlan product is throwing \\\"Error 500\\\" when exporting reports since yesterday, blocking quarter-end close.\",\"severity\":\"high\",\"sentiment\":\"negative\",\"incident_date\":\"2024-06-11\"}","content_encoding":"plain","content_url":null}]}],"mode":"stream"}'

Expected Results:

In the response body, you should see an appropriate human-like ticket agent response.

7. Clean Up¶

If you have not already, remember to stop the ACP agent servers using Ctrl + C or exiting the terminal where they are running.