System prompts

System prompts define an agent’s task, response style, and behavioral boundaries. Set Agent.instruction to fixed text, a template that reads session state, or a function that builds instructions from the current context Complete installation and model configuration before running these examples. Save each script with the indicated filename and run it with python filename.py

Static prompts

Use a string for a fixed task. Specify the task, information to clarify, and data the agent must not request

main.py

import asyncio
from veadk import Agent, Runner

agent = Agent(
    name="support_assistant",
    instruction=(
        "You help users troubleshoot billing issues. "
        "Ask for missing details before suggesting a solution. "
        "Do not ask for passwords or payment card numbers."
    ),
)

async def main():
    result = await Runner(agent=agent).run(
        messages="My invoice total is different from last month.",
        session_id="prompt-demo",
    )
    print(result)

if __name__ == "__main__":
    asyncio.run(main())

The script prints a response to the billing question. The exact wording depends on the model

Dynamic prompts

Use session state

A {name} placeholder in a string instruction reads the matching key from session state. This example creates the session with an initial name and language before running the agent

state_prompt.py

import asyncio
from veadk import Agent, Runner

agent = Agent(
    name="personal_assistant",
    instruction="Address the user as {user_name}. Answer in {language}.",
)

async def main():
    runner = Runner(agent=agent, app_name="prompt_demo", user_id="demo-user")
    await runner.session_service.create_session(
        app_name="prompt_demo",
        user_id="demo-user",
        session_id="state-demo",
        state={"user_name": "Alex", "language": "English"},
    )
    print(await runner.run(messages="Introduce yourself.", session_id="state-demo"))

if __name__ == "__main__":
    asyncio.run(main())

Syntax	Behavior
`{user_name}`	Reads a required key; a missing key causes the run to fail
`{user_name?}`	Reads an optional key; a missing key becomes an empty string
`{app:language}`, `{user:language}`, `{temp:language}`	Reads a state key in the corresponding namespace

A state value of None also becomes an empty string. Tools and other agents can update state during a run. See session management

Use an InstructionProvider

Pass a function to instruction when you need defaults, conditional logic, or several state values. The function receives a ReadonlyContext and returns a string. Async functions defined with async def are also supported

dynamic_prompt.py

import asyncio
from google.adk.agents.readonly_context import ReadonlyContext
from veadk import Agent, Runner

def build_instruction(context: ReadonlyContext) -> str:
    language = context.state.get("language", "English")
    return f"Answer in {language}. Keep the explanation concise and factual."

agent = Agent(name="assistant", instruction=build_instruction)

async def main():
    print(await Runner(agent=agent).run(
        messages="What is a session?", session_id="provider-demo"
    ))

if __name__ == "__main__":
    asyncio.run(main())

context.state is read-only. Other useful fields include user_id, user_content, session, and agent_name. Placeholders in a function’s return value are not automatically substituted: build the string in the function. To reuse ADK template substitution, call await inject_session_state(template, context) from an async function, importing it from google.adk.utils.instructions_utils

instruction and description

Parameter	Type	Default	Purpose
`instruction`	`str` or `InstructionProvider`	VeADK general task instructions	Guides the current agent’s task and response
`description`	`str`	VeADK general capability description	Helps other agents select a transfer target
`prompt_manager`	`BasePromptManager \| None`	`None`	Provides `instruction` through a manager, overriding explicitly supplied instructions

Give each agent specific instructions and a concise description in a multi-agent application. To maintain prompts in an external service, see prompt management

Overview

Get started

Core components

Observability

Extensions

Frontend

Security

Deployment

Static prompts

Dynamic prompts

Use session state

Use an InstructionProvider

instruction and description

​Static prompts

​Dynamic prompts

​Use session state

​Use an InstructionProvider

​instruction and description

Static prompts

Dynamic prompts

Use session state

Use an InstructionProvider

instruction and description