Q: Does include_contents='none' exclude providing state as input to the agent?

[xuanyang15]

Q: Does include_contents='none' exclude providing state as input to the agent? (When not explicitly embedding state in the agent's instruction.)

From the documentation, it does not seem very clear, it only talks about chat history is excluded.

https://google.github.io/adk-docs/agents/llm-agents/#managing-context-include_contents

[xuanyang15]

it has nothing to do with state

[xuanyang15]

Thanks! After trying things out more with ADK, now I understand where my confusion came from. I think in early starter documentation, I want a clearer explanation of what an agent can see as input.

IIUC, it's just instruction + chat history + tool use results. If I want state visible, it must be templated in instruction.
include_contents='none' excludes chat history, only allow last message + its own tool usages.

Is my understanding right?

[xuanyang15]

Response from ADK Answering Agent

TLDR: Your understanding is mostly correct! include_contents='none' does exclude chat history. However, state is more broadly available than just through the instruction; it's a core part of the session's context accessible by tools and callbacks.

---

Your understanding is very close to the mark. Let me clarify a few points based on the ADK documentation.

• Agent's Input: You are right that an agent's input consists of its instruction, the conversation history (if included), and the results of any tool use. [7, 4]

• include_contents='none': Your understanding is correct here. Setting include_contents='none' prevents the agent from receiving the prior conversation history. It will operate based on its instruction and the input from the current turn. [3, 9]

• State Visibility: This is the main point of clarification. state is not just visible when templated into an instruction. It is a fundamental part of the session context and serves as the primary way to manage data and communicate between different parts of the system within a single conversation. [2, 10] For example:

  • Tools can access and modify the session state via the ToolContext. [4]
  • Callbacks can read and write to the state through the CallbackContext. [8, 12]
  • Workflow Agents like SequentialAgent use state to pass information between sub-agents. [1]

So, while you can use state to dynamically create an instruction [2], its accessibility is much broader. It's the "working memory" for the current session. [7]

[1] https://github.com/google/adk-docs/blob/main/docs/agents/multi-agents.md
[2] https://github.com/google/adk-docs/blob/main/docs/context/index.md
[3] https://github.com/google/adk-docs/blob/main/docs/agents/llm-agents.md
[4] https://github.com/google/adk-docs/blob/main/docs/tools/index.md
[7] https://github.com/google/adk-docs/blob/main/docs/get-started/about.md
[8] https://github.com/google/adk-docs/blob/main/docs/tutorials/agent-team.md
[9] https://github.com/google/adk-python/blob/main/src/google/adk/agents/llm_agent.py
[10] https://github.com/google/adk-docs/blob/main/docs/sessions/index.md
[12] https://github.com/google/adk-docs/blob/main/docs/callbacks/types-of-callbacks.md