System prompt - AgentKit 文档

An agent’s system prompt is set through the instruction parameter, which governs the model’s behavior, tone, and how it answers.

Set instruction

The simplest form is a plain string:

from veadk import Agent

agent = Agent(
    name="assistant",
    instruction="You are a professional support assistant. Answer concisely and politely.",
)

Dynamic prompts

When the prompt needs to vary with runtime context, there are two options. Placeholders: use {variable_name} in the string; at runtime it is filled from the matching value in session state.

agent = Agent(
    instruction="You are serving user {user_name}. Answer concisely and politely.",
)

InstructionProvider: pass a function that receives the runtime context and returns the prompt string, letting you read session state to build the prompt yourself. The function can also be an async function.

from google.adk.agents.readonly_context import ReadonlyContext
from veadk import Agent

def build_instruction(ctx: ReadonlyContext) -> str:
    user_name = ctx.state.get("user_name", "there")
    return f"You are serving {user_name}. Answer concisely and politely."

agent = Agent(instruction=build_instruction)

ctx also exposes user_id, user_content, and session; session state is written by the session and tools during a run.

instruction vs. description

instruction: the system prompt that guides how the model reasons and answers.
description: a short summary of the agent’s capabilities, used mainly for routing and selection in multi-agent scenarios — it does not act as the system prompt.

Last modified on September 19, 2026