how do agents implicitly use states?

[mohammad-passage]

I have a question and am too lazy to look at the code for that.

Say you have a state like {"first_name": "Alex"} and the query to agent is "What is the first name? You can look at state's first_name key".

q1: Does the agent answer this? Based on my test it does.
q2: How does this happen? There should be somewhere that we put the state values in the prompt/instruction. who does it?

[boyangsvl]

I don't think it would work, as if you don't refer to the state like {first_name} in your instruction, it won't be sent to the model. Could you share with me your code if this actually works?

[mohammad-passage]

For example this prompt worked:

    You are a number incrementer agent. Your task is to:
    1. First, read the current number from state using ctx.state.get("my_number", 123)
    2. Then use the increment_number tool to increment that number
    3. Report the new number to the user

    The initial value of the number should be 123 if not present in state.
    Each time you run, the number should increase by 1.

    Example of how to read from state:
    current_number = ctx.state.get("my_number", 123)

so it somehow knows the state.

[kanlanc]

I defined a tool in which i send the session_id as a parameter, extract the state variable that I want from that session_id and return that information to the agent

So when the agent needs that information, it calls the tool which returns the state information

[kanlanc]

The code in loop agent part of the documentation says it does

# Conceptual Example: Using output_key and reading state
from google.adk.agents import LlmAgent, SequentialAgent

agent_A = LlmAgent(name="AgentA", instruction="Find the capital of France.", output_key="capital_city")
agent_B = LlmAgent(name="AgentB", instruction="Tell me about the city stored in state key 'capital_city'.")

pipeline = SequentialAgent(name="CityInfo", sub_agents=[agent_A, agent_B])
# AgentA runs, saves "Paris" to state['capital_city'].
# AgentB runs, its instruction processor reads state['capital_city'] to get "Paris".

[paulrinaldi]

Aren't we supposed to use {} notation there? e.g. {capital_city}?

[surajksharma07]

You're right @paulrinaldi. {} notation is required. boyangsvl's original skepticism was correct: plain-text mentions of a state key in an instruction don't give the model implicit access, only {variable} templating does. The docs example has since been cleaned up too, the current sequential-agents pipeline example uses proper {generated_code}/{review_comments} templating throughout. See discussion #2489 for the fuller breakdown of how state actually reaches the model. Closing this out since it's answered, feel free to reopen if anything's still unclear.