Google ADK — LlmAgent Reference
Import
from google.adk.agents import Agent # Alias for LlmAgent
from google.adk.agents import LlmAgent # Full name
Basic Agent
root_agent = Agent(
name="my_agent",
model="gemini-2.5-flash",
description="What this agent does (used by parent for delegation).",
instruction="You are a helpful assistant that answers questions.",
tools=[my_tool_function],
)
All LlmAgent Parameters
| Parameter |
Type |
Description |
name |
str |
Unique agent identifier (required) |
model |
str |
Model ID like "gemini-2.5-flash" (required) |
instruction |
str | Callable |
System prompt or dynamic instruction provider |
description |
str |
Used by parent agent for delegation decisions |
tools |
list |
List of tool functions, BaseTool instances, or toolsets |
sub_agents |
list |
Child agents for multi-agent systems |
output_key |
str |
State key to store agent's final output |
output_schema |
type |
Pydantic model or dict for structured output |
include_contents |
str |
"default" or "none" — controls conversation history visibility |
before_agent_callback |
Callable |
Called before agent runs |
after_agent_callback |
Callable |
Called after agent completes |
before_tool_callback |
Callable |
(tool, args, tool_context) -> Optional[dict] — called before tool |
after_tool_callback |
Callable |
(tool, args, tool_context, response) -> Optional[dict] — called after tool |
generate_content_config |
GenerateContentConfig |
Advanced model configuration |
code_executor |
CodeExecutor |
Built-in code execution capability |
planner |
BasePlanner |
Enable planning/thinking (e.g. BuiltInPlanner) |
context_cache_config |
ContextCacheConfig |
Cache context for efficiency |
Dynamic Instruction Provider
from google.adk.agents.readonly_context import ReadonlyContext
def my_instruction(readonly_context: ReadonlyContext) -> str:
user_name = readonly_context.state.get("user_name", "user")
return f"You are a helpful assistant for {user_name}. Be concise."
root_agent = Agent(
name="personalized_agent",
model="gemini-2.5-flash",
instruction=my_instruction,
tools=[],
)
Structured Output (output_schema)
from pydantic import BaseModel
class AnalysisResult(BaseModel):
summary: str
sentiment: str
confidence: float
agent = Agent(
name="analyzer",
model="gemini-2.5-flash",
instruction="Analyze the given text and return structured results.",
output_schema=AnalysisResult,
)
include_contents Options
| Value |
Behavior |
"default" |
Agent sees relevant conversation history (default) |
"none" |
Agent sees NO prior conversation — only instruction + current input |
summary_agent = Agent(
name="summarizer",
model="gemini-2.5-flash",
instruction="Summarize the provided text.",
include_contents="none", # Fresh context each time
)
output_key — Storing Results in State
research_agent = Agent(
name="researcher",
model="gemini-2.5-flash",
instruction="Research the given topic.",
output_key="research_output", # Stored in state["research_output"]
)
Generate Content Config
from google.genai import types
agent = Agent(
name="creative_agent",
model="gemini-2.5-flash",
instruction="Write creative stories.",
generate_content_config=types.GenerateContentConfig(
temperature=0.9,
top_p=0.95,
max_output_tokens=2048,
),
)
Code Executor
from google.adk.code_executors import BuiltInCodeExecutor
agent = Agent(
name="coder",
model="gemini-2.5-flash",
instruction="Write and execute Python code to solve problems.",
code_executor=BuiltInCodeExecutor(),
)
Key Conventions
name must be a valid Python identifier (letters, digits, underscores)description is critical for multi-agent delegation — be specificinstruction can be a string or a callable that returns a string- Tools are plain Python functions with docstrings (auto-converted to FunctionTool)
- The module-level variable MUST be named
root_agent for CLI discovery
Related Skills
google-adk-function-tool — Creating custom tools for agentsgoogle-adk-callbacks — Lifecycle hooks (before/after agent and tool)google-adk-multi-agent — LLM-routed multi-agent systemsgoogle-adk-workflow-agents — Deterministic orchestration (Sequential, Parallel, Loop)
eagleisbatman