Step-By-Step Implementation Walk Through:

(1) Create a new directory named "agent-adk" in your C:\ drive by running the command: mkdir agent-adk

(2) Open Visual Studio Code, click File, and select Open Folder.

(3) Select the folder “agent-adk”

(4) Select “New Terminal” from the Menu.

(5) Type the following commands in your terminal:

PS C:>agent-adk>python -m venv .venv and press Enter

PS C:>agent-adk>.venv\Scripts\activate (For Windows) or source .venv\bin\activate (For Mac) and press Enter(.venv) and press Enter

(.venv) PS C:>agent-adk>pip install google-adk pypdf and press Enter

(.venv) PS C:>agent-adk>adk create finance_analyst

(6) You will create two files inside the finance_analyst folder.


Step 1: The Eyes (tools.py)

This script handles the "vision" part of the agent. It reads the PDF and inserts the crucial "Page Markers" that allow the AI to cite sources.

File Location: AGENT-ADK/finance_analyst/tools.py

Step 2: The Brain (agent.py)

This defines the "Employee." We use the root_agent variable (required by ADK) and attach our tool.

File Location: AGENT-ADK/finance_analyst/agent.py

(7) Create a new file called ".env" in Visual Studio Code and replace the placeholder GOOGLE_API_KEY = "YOUR_API_KEY" with your actual API key.

(8) This is what your folder will look like when we are done. We create a parent folder (agent-adk) to hold the project, and a sub-folder (finance_analyst) for the actual agent code.

(9) Type the following commands in your terminal:

(.venv) PS C:\agent-adk> adk web

(10) What just happened? By running a single command, you spun up a "Flight Control" dashboard that provides far more than just a chat interface:

1. The X-Ray Trace: On the left panel, you don't just see the answer; you see the cognitive steps. You can watch the exact moment the agent decides to call parse_earnings_report.
   
   

2. State Inspection: You can view the raw JSON state of the agent's memory, allowing you to debug exactly what context was retained or lost.
   
   

3. Zero Boilerplate: We went from "Concept" to "Interactive UI" without writing a single line of frontend code.
   
   

4. The Result: Navigate to http://localhost:8000. You will see your Finance Analyst ready to work. Download the file “nvidia_10k.pdf” and attach it to the chat, and press Enter
   
   Simply ask it to "Analyze the risks in the Nvidia PDF" and press Enter—watch it analyze the document in real time.
   

Key Takeaways: What We Learned

• Build in Parts, Not One Big Chunk: Instead of writing one giant, messy script, we split the agent into three clear parts: The Eyes (reading files), The Brain (thinking), and The Face (the website). This makes it much easier to fix and upgrade later.
  
  

• Give the AI Strict Rules: The difference between a creative writer and a serious analyst is the Instructions. By telling the AI, "You must cite page numbers," we forced it to be accurate and truthful instead of just making things up.
  
  

• Tools Are Easy to Add: We didn't need complicated code to let the agent read PDFs. We just wrote a simple Python function, and the ADK framework did the rest. It’s like plugging a USB device into a computer—it just works.
  
  

• Skip the Boring Coding: In the past, we had to write hundreds of lines of code just to make a chat box appear on the screen. With the adk web command, we got a full professional dashboard instantly. This lets us focus on making the AI smarter, not building menus and buttons.
  
  

• Good Data = Good Answers: The AI didn't magically know the page numbers. We helped it by adding "Page Markers" when we read the PDF. This proves that preparing your data correctly is still the most important step in building a smart agent.