Skip to main content

Llm Agent

The LlmAgent is the fundamental unit of intelligence in the Google Agent Development Kit (ADK). It wraps a Large Language Model (Gemini) with Instructions, Configuration, and State Management.

When you need an AI to perform a task—whether it's writing code, summarizing text, or analyzing data—you use an LlmAgent.

Visual Guide: Creating an Agent​

1. Drag & Drop​

  1. Locate the Sidebar on the left side of the builder.
  2. Expand the Agents section.
  3. Click and drag the LlmAgent card onto the infinite canvas.
  4. A purple node will appear representing your new agent.

2. Selection & Properties​

Click on the node to open the Properties Panel on the right side of the screen. This panel is divided into three logical sections:

  1. General: Identity and Core Instructions.
  2. Generation Config: Fine-tuning the AI's creativity and output format.
  3. Routing & State: Managing memory and context flow.

1. General Settings​

These are the required fields that define who the agent is.

PropertyDescriptionBest Practice
Agent NameA unique identifier for this agent.Use snake_case (e.g., research_assistant). This name becomes a variable in the code, so avoid spaces.
ModelThe specific Gemini model version to power this agent.gemini-2.0-flash: Best for speed and general tasks.
gemini-3-pro-preview: Best for complex reasoning and coding.
gemini-2.5-flash-image: Specialized for vision tasks.
System InstructionThe "Prompt" that defines the agent's persona and rules.Be specific. Define the role ("You are a coder"), the input format ("You will receive a CSV"), and the output format ("Reply only with Python code").

2. Generation Configuration​

Click the Generation Config dropdown in the properties panel to access advanced controls. These settings determine how the model "thinks" and generates text.

Creativity Controls​

  • Temperature (0.0 - 2.0):

    • Low (0.0 - 0.5): Deterministic and focused. Use this for code generation, data extraction, or factual Q&A.
    • High (1.0 - 2.0): Creative and diverse. Use this for brainstorming, poetry, or creative writing.
    • Default: 1.0

  • Top P (0.0 - 1.0):

    • Also known as "Nucleus Sampling". It limits the pool of potential next words to the top cumulative probability.
    • Lower values make the text more coherent and predictable.

  • Top K (1 - 100):

    • Limits the model to picking from the top K most likely next words.
    • Default: 40.

Output Controls​

  • Max Output Tokens:

    • Sets a hard limit on the response length.
    • Default: 8192. Increase this if you expect the agent to write long essays or large code files.

  • Response MIME Type:

    • text/plain: Standard free-form text response.
    • application/json: Forces the model to output a valid JSON object. This is critical when chaining agents where the next agent expects structured data.

Structured Output (JSON Schema)​

When you select application/json as the MIME Type, a "Configure Output Schema" button appears.

Clicking this opens the Schema Editor, allowing you to define the exact keys and data types the AI must return.

Example Schema for a Recipe Generator:

  • Key: recipe_name | Type: STRING | Desc: The title of the dish.
  • Key: prep_time | Type: INTEGER | Desc: Time in minutes.
  • Key: ingredients | Type: STRING | Desc: A markdown list of items.
Why use Schemas?

Defining a schema ensures reliability. Instead of parsing text with Regex, your code can reliably access response.recipe_name.

3. Routing & State​

This section controls how the agent interacts with the conversation history and how it passes data to other agents.

Context Management​

  • Include History (Toggle):
    • ON (Default): The agent sees the entire conversation history leading up to this point. It knows what previous agents said.
    • OFF: The agent starts with a "Fresh Context". It only sees its own System Instruction and the immediate input. This saves tokens and reduces confusion for specialized tasks.

State Isolation​

  • Block Parent Transfer:

    • If checked, variables created by this agent are not passed back up to the parent orchestrator. Useful for keeping the global state clean.

  • Block Peer Transfer:

    • If checked, variables created by this agent are not visible to sibling agents in the same chain.

Auto-Generated Fields​

  • Output Key:
    • This field is read-only and auto-generated based on your Agent Name.
    • Example: If Agent Name is writer_bot, Output Key is output_writer_bot.
    • This is the variable name used to store the agent's response in the session memory. Subsequent agents can reference this data.

Example Workflow​

Goal: Create a travel planner.

  1. Agent 1: researcher
    • Instruction: "Find 3 hotels in Paris."
    • Output Key: output_researcher
  2. Agent 2: writer
    • Instruction: "Take the hotels listed in {output_researcher} and write a polite email to the user recommending them."
    • Include History: ON (so it can read the output).

By chaining LlmAgents visually, you can build complex, reliable applications without writing boilerplate code.