Building a marketing agent

Google Cloud Cloud Shell

TODO: Activate Vertex AI API

1. Create and activate virtual environment

 python3 -m venv .gdg_usyd && source .gdg_usyd/bin/activate

2. Install Google ADK

pip install google-adk

3. Initialise ADK project

adk create marketing-agent

Follow the prompts a. select 1. gemini-2.5-flash b. select 2. Vertex AI c. Ensure the correct proejct ID is selected in "Enter Google Cloud project ID" d. Use default us-central1

4. Test the base Agent with native UI

adk web

Ensure you click on this link and not copy paste.

Ensure you select the correct Agent on the top left

5. Prompting In your agent.py file update the instruction to

instruction='''
You are Marketing Assistant, a strategic yet hands-on partner that plans, creates, and optimizes marketing for measurable growth while keeping brand consistency.
If key inputs are missing, briefly ask for goal, audience, and KPI—otherwise proceed.
Deliver concise, ready-to-ship assets (respect platform limits) and clearly mark assumptions vs facts.
Provide 3 creative variations with a one-sentence hypothesis on which might win and why.
Respond using sections: Summary, Deliverables, Variations, Distribution & Measurement, Next Actions, Sources.
'''

Start with this question

Hi, I am hosting a GDG on campus event to build AI Agents. Can you please create a 1 page brief?
Goal: To educate and upskill students.
Audience: Students starting their AI journey.
KPIs: Number of attendees and new sign-ups

Once you are happy with this result, end the server using ctrl+c in your terminal window.

7. Tooling - Google Search Lets provide the Agent access to the internet and allow it to research relevant information from the web

   a. Import ADK's inbuilt google search tool in your agent.py file

   from google.adk.tools import google_search

   b. Update the instruction in your Agent definition to guide the Agent to reserach relevant event's marketing campaign.

       You are Marketing Assistant, a strategic yet hands-on partner that plans, creates, and optimizes marketing for measurable growth while keeping brand consistency. Ensure you research relevant events that have been hosted in the last few months and ensure our campaign is unique but also take inspiration from them.
       If key inputs are missing, briefly ask for goal, audience, and KPI—otherwise proceed.
       Deliver concise, ready-to-ship assets (respect platform limits) and clearly mark assumptions vs facts.
       Provide 3 creative variations with a one-sentence hypothesis on which might win and why.
       Respond using sections: Summary, Deliverables, Variations, Distribution & Measurement, Next Actions, Sources.

   c. Register the tool with your agent

       tools=[google_search]

   d. Reuse the same question and see the difference

       Hi, I am hosting a GDG on campus event to build AI Agents. Can you please create a 1 page brief?
       Goal: To educate and upskill students.
       Audience: Students starting their AI journey.
       KPIs: Number of attendees and new sign-ups

   Once you are happy with this result, end the server using ctrl+c in your terminal window.

8. Tooling - Imagen Lets provide the Agent access to Imagen to generate marketing posters

   a. Import necessary packages in your agent.py file

   from google.adk.agents.llm_agent import Agent
   from google.adk.tools import google_search, ToolContext, load_artifacts
   from google.genai import Client, types
   
   from dotenv import load_dotenv
   import os 

   b. Define a a tool to use Imagen. To avoid excess charges and quickly iterate with some designs, lets use Imagen 3 Fast. Ensure you you paste this function above the Agent definition.

   MODEL = "gemini-2.5-flash"
   IMAGE_MODEL = "imagen-3.0-fast-generate-001"
   # IMAGE_MODEL = "imagen-4.0-generate-001"
   
   load_dotenv()
   
   client = Client(
       vertexai=True,
       project=os.getenv("GOOGLE_CLOUD_PROJECT"),
       location=os.getenv("GOOGLE_CLOUD_LOCATION"),
   )
   
   async def generate_image(img_prompt: str, tool_context: "ToolContext"):
       """Generates an image based on the prompt provided to this tool"""
       response = client.models.generate_images(
           model=IMAGE_MODEL,
           prompt=img_prompt,
           config={"number_of_images": 1},
       )
       if not response.generated_images:
           return {"status": "failed"}
       image_bytes = response.generated_images[0].image.image_bytes
       await tool_context.save_artifact(
           "image.png",
           types.Part.from_bytes(data=image_bytes, mime_type="image/png"),
       )
       return {
           "status": "success",
           "detail": "Image generated successfully and stored in artifacts.",
           "filename": "image.png",
       }

   c. Add these new functions to the Agent

   tools=[generate_image, load_artifacts, google_search]

   c. Run the Agent

       adk web

   Reuse the same question and see the difference

       Hi, I am hosting a GDG on campus event to build AI Agents. Can you please create a 1 page brief?
       Goal: To educate and upskill students.
       Audience: Students starting their AI journey.
       KPIs: Number of attendees and new sign-ups

   Once you are happy with this result, end the server using ctrl+c in your terminal window.

   {
       "error": "400 INVALID_ARGUMENT. 
       {
           'error': {
               'code': 400, 
               'message': 'Multiple tools are supported only when they are all search tools.', 
               'status': 'INVALID_ARGUMENT'}
           }
       "
   }

   d. This is a known caveat for using inbuilt tools with ADK. We need to expose another Agent as tool, import this method

   from google.adk.tools.agent_tool import AgentTool

   e. Create a specialised agent responsible for searching. Create this agent before your root agent so you can use this in your root agent.

   search_agent = Agent(
       model='gemini-2.0-flash',
       name='SearchAgent',
       instruction="""
       You're a specialist in Google Search
       """,
       tools=[google_search],
   )

   f. Update the existing root_agent with this definition

   root_agent = Agent(
       model=MODEL,
       name='root_agent',
       description='A Marketing assistant to help users create marketing campaigns',
       instruction='''You are Marketing Assistant, a strategic yet hands-on partner that plans, creates, and optimizes marketing for measurable growth while keeping brand consistency. 
           Use Google Search tool to research relevant events that have been hosted in the last few months and ensure our campaign is unique but also take inspiration from them.
           If key inputs are missing, briefly ask for goal, audience, and KPI—otherwise proceed.
           Deliver concise, ready-to-ship assets (respect platform limits) and clearly mark assumptions vs facts.
           Provide 3 creative variations with a one-sentence hypothesis on which might win and why.
           Respond using sections: Summary, Deliverables, Variations, Distribution & Measurement, Next Actions, Sources.
       ''',
       tools=[AgentTool(agent=search_agent), generate_image, load_artifacts],
   )

   g. Reuse the same question and see the difference

       Hi, I am hosting a GDG on campus event to build AI Agents. Can you please create a 1 page brief?
       Goal: To educate and upskill students.
       Audience: Students starting their AI journey.
       KPIs: Number of attendees and new sign-ups

   h. Once you are happy with this result. Ask the agent to generate a poster for this event.

   i. With Gemini familiy's multimodal capabilities, you can ask the Agent to take inspiration from an existing image. Try uploading a reference image and direct the LLM to create something similar

Summary

In this workshop we built a marketing-agent to produce a one-page event brief and provided it tools for Google Search and Imagen 3 Fast via a custom tool to generate posters and save artifacts.