If there is a way to define the context information obtained by the agent?

[207285]

Currently, when defining an agent, the available context options are limited to include_contents: literal ['default', 'none']. This design restricts agents to either accessing the entirety of the session's context or having no context at all.

However, in real-world development, agents often require more granular control over the context they receive. For instance, an agent might only need its own specific context within the broader session. The ability to define and utilize custom context is crucial for implementing advanced features like multi-turn dialogues, where an agent needs to recall specific information from its previous interactions without being overwhelmed by unrelated data from the entire session.

Therefore, optimizing the context handling mechanism to allow for custom context definitions is a critical improvement. This would empower developers to build more sophisticated and efficient multi-turn dialogue agents.

[adk-bot]

Response from ADK Answering Agent (experimental, answer may be inaccurate)

TLDR: While the include_contents parameter is limited to 'default' and 'none', ADK provides a more powerful and granular way to manage context using the session.state object. You can explicitly pass specific pieces of information to an agent by embedding placeholders in its instructions and populating session.state from previous agents or tools.

---

Hello! Thanks for your question. Your observation is correct: the include_contents parameter on an LlmAgent is intentionally limited to either including the conversation history ('default') or not ('none'). [1]

For the granular context control you're looking for, ADK provides a more explicit and powerful mechanism: the session state. Think of session.state as a mutable, in-memory dictionary that serves as the working memory for the entire conversation session. [2, 3] You can use it to store specific pieces of information that subsequent agents can then access.

Here is the recommended pattern for providing specific, custom context to an agent:

1. Store Information in session.state:
An agent can save its output directly to the state by using the output_key parameter. For example, an agent that identifies a topic could save it to session.state['topic'].

2. Inject State into another Agent's Instructions:
A subsequent agent can then be instructed to use that specific piece of information. The ADK framework automatically substitutes placeholders in an agent's instruction string with values from session.state. [4]

Here's a conceptual example of how this works in a multi-agent setup:

from google.adk.agents import LlmAgent, SequentialAgent

# This agent identifies the key topic and saves it to the state.
topic_extractor_agent = LlmAgent(
    name="TopicExtractor",
    instruction="Identify the main topic from the user's message and output only that topic.",
    output_key="topic"  # The result will be saved to session.state['topic']
)

# This agent receives the specific 'topic' as context in its instructions.
deep_dive_agent = LlmAgent(
    name="DeepDive",
    # The {topic} placeholder is filled from session.state['topic'] before being sent to the LLM.
    instruction="Provide a detailed explanation of the following topic: {topic}"
)

# A SequentialAgent orchestrates the flow, ensuring the state is passed along.
pipeline = SequentialAgent(
    name="MyPipeline",
    sub_agents=[topic_extractor_agent, deep_dive_agent]
)

In this example, the DeepDive agent doesn't receive the entire conversation history. Instead, it gets a very specific piece of context ({topic}) injected directly into its instructions, allowing for the kind of precise, multi-turn control you described. [5]

You can also read from and write to session.state from within custom tools by using the ToolContext object, giving you full control over the contextual data available during a session. [6]

This state-passing mechanism is central to building sophisticated multi-turn and multi-agent workflows in ADK. [5, 7]

[1] https://google.github.io/adk-docs/agents/llm-agents/
[2] https://google.github.io/adk-docs/sessions/
[3] https://google.github.io/adk-docs/context/
[4] https://google.github.io/adk-docs/agents/custom-agents/
[5] https://google.github.io/adk-docs/agents/multi-agents/
[6] https://google.github.io/adk-docs/tools-custom/
[7] https://github.com/google/adk-docs/blob/main/tutorials/agent-team.md