Using a real model

Everything so far used the stand-in you wrote in lesson 4. Swapping in a real model is one line, and nothing else in any of these programs changes.

agent = LlmAgent(
    name="support",
    model="gemini-flash-latest",
    instruction="Help customers with their orders.",
    tools=[lookup_order],
)

A string instead of an object. ADK looks it up, builds the client, and reads the key from your environment. Everything downstream, the runner, sessions, tools, callbacks, workflow agents, was only ever talking to something with a generate_content_async method.

Setting up the key

Get a Gemini API key from Google AI Studio, then put it in a file called .env in the agent's folder, next to agent.py. That is where ADK looks, and it is why adk create generates the file for you in lesson 1.

# my_agent/.env
GOOGLE_API_KEY="PASTE_YOUR_GEMINI_API_KEY_HERE"

That single line is the whole setup for the Gemini API. Nothing in your code reads it: ADK loads the file and the client picks the key up from the environment, which is why the agent above only names a model.

Or through Google Cloud

If you are going through the Gemini Enterprise Agent Platform rather than the Gemini API, the same file holds a project and a location instead of a key:

# .env configuration file
GOOGLE_CLOUD_PROJECT=your-project-id
GOOGLE_CLOUD_LOCATION=us-central1
GOOGLE_GENAI_USE_ENTERPRISE=True

There is no output on this page, because a real model needs a key and this site only shows output that came out of a program that ran here.

Three ways to reach a model

LiteLLM is the one worth knowing about, because it is how ADK reaches a hundred other providers with one wrapper. Ollama and vLLM are the same idea for models you host yourself.

What changes when the model is real

• It decides for itself. The same question can produce a different tool call twice, which is why lesson 6 taught you to read events.
• Bad descriptions start costing you. The stand-in never misread a tool. A real model does, and lesson 9 is the fix.
• Limits matter. The max_llm_calls from lesson 15 stops being theoretical.
• It costs money per turn. Long sessions send the whole history, so lesson 21's advice about small state and small tool results turns into a bill.