How to use dynamic prompt?

[hqta1110]

How to use dynamic prompt in ADK? I tried creating a variable type_error in state that will change depends on user's input, then use a switch case to set the instructions for agents, but it did not work. I think it is because the agents are initialized once at the beginning of the app - when the state and type_error still None

[dexhunter]

Pass a function as instruction, so ADK chooses the prompt from the current session state at runtime. A switch executed while constructing the agent only sees the state at construction time.

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

def choose_instruction(ctx: ReadonlyContext) -> str:
    match ctx.state.get("type_error"):
        case "syntax":
            return "Help the user diagnose and fix syntax errors."
        case "runtime":
            return "Help the user diagnose runtime errors from the traceback."
        case _:
            return "Ask for the error message and relevant code."

root_agent = Agent(
    name="error_helper",
    model="gemini-2.5-flash",
    instruction=choose_instruction,
)

Make sure the tool or callback that determines the error writes type_error into the session state through its context before the next model request. Changing an unrelated Python variable will not update that state.

If you only need to insert a value into a fixed prompt, a string such as instruction="Help with this error category: {type_error?}" also works; ? makes a missing state key optional. Use the callable when you need branching. ADK resolves it through canonical_instruction.