Gemini 3 Flash

[tdakanalis]

Hi, I have configured gemini-3-flash-preview and set include_thoughts to false. However, in streaming mode, the thoughts are still included in the response. Does anyone know how to resolve this?

root_agent = Agent(
    name="root_agent",
    model="gemini-3-flash-preview",
    static_instruction=types.Content(role="user", parts=[types.Part(text=load_prompt_from_file(Prompt.ASSIST.value))]),
    planner=BuiltInPlanner(
         thinking_config=ThinkingConfig(include_thoughts=False)
    )
)

[tdakanalis]

@adk-bot any ideas?

[adk-bot]

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

TLDR: Your agent configuration to disable thoughts seems correct. The issue might be related to your streaming implementation or a conflicting configuration. Please check your RunConfig and how you initiate the streaming run.

---

Hi there! Thanks for your question.

Your configuration of the Agent and BuiltInPlanner to disable thoughts using ThinkingConfig(include_thoughts=False) looks correct based on the ADK framework. However, since you're seeing this issue in streaming mode, the root cause might lie in how the streaming run is initiated.

Here are a few things you could check:

1. Conflicting RunConfig: The runner.run() or runner.run_live() methods can take a RunConfig object, which might have its own thinking_config that could override the one in your agent's planner. Please ensure you are not passing a different thinking_config when you start the run.

2. Streaming Mode Behavior: In streaming, the model sends back partial responses as they are generated. It's possible that what you are observing are not the model's internal "thoughts" but rather these incremental updates. The final, complete response should not contain the thoughts if include_thoughts is False.

3. Explicitly Set thinking_config to None in generate_content_config: You could try explicitly setting thinking_config=None in the generate_content_config of your agent to be certain no other configuration is interfering. Here is an example of what that might look like:

from google.adk import Agent
from google.adk.planners import BuiltInPlanner
from google.genai import types

root_agent = Agent(
    name="root_agent",
    model="gemini-3-flash-preview",
    static_instruction=types.Content(role="user", parts=[types.Part(text=load_prompt_from_file(Prompt.ASSIST.value))]),
    planner=BuiltInPlanner(
         thinking_config=types.ThinkingConfig(include_thoughts=False)
    ),
    generate_content_config=types.GenerateContentConfig(
        thinking_config=None # Explicitly setting this to None
    )
)

If you've checked these points and the issue persists, it would be helpful if you could share a bit more of your code, specifically how you are calling the runner to start the streaming process. This would help in diagnosing the problem further.

Hope this helps!

[tdakanalis]

@adk-bot What I experience is the streaming of the following parts:

Wait, the prompt says "what are the possible options here?".
Let's write the response.

Wait, the User's Latest Message: "..."
Let's go.

Actually, the "Strategic Consultation" section says: "..."
Wait, I should also check if the user is in "Setup Mode". No, they are asking a question.

I have set thinking_config=None but nothing changed.

[adk-bot]

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

TLDR: The behavior you're observing is likely caused by the instructions within your prompt (Prompt.ASSIST.value), which may be encouraging the model to 'think out loud' as part of its response. This is different from the structured thoughts controlled by ThinkingConfig.

---

Hi again! Thank you for providing the extra details. This is very helpful.

The output you're seeing (Wait, the prompt says..., Let's write the response., etc.) suggests that the model is verbalizing its internal monologue. This is often a result of prompt engineering, where the instructions guide the model to 'think step-by-step' to improve the quality of its final output.

The ThinkingConfig(include_thoughts=False) setting is designed to control the inclusion of the planner's structured thought process in the response metadata, not the narrative content generated by the model. If the prompt itself tells the model to think out loud, the model will generate that text as part of the main response, and include_thoughts will not affect it.

Here are my recommendations:

1. Inspect Your Prompt: Please review the contents of the file loaded by load_prompt_from_file(Prompt.ASSIST.value). It likely contains instructions that encourage this step-by-step thinking.

2. Modify the Prompt: Try modifying the prompt to explicitly instruct the model to be concise and not to show its reasoning. You could add a sentence like:

   "You must provide only the final and direct answer. Do not include your internal thinking process, reasoning steps, or any self-correction narrative in the response."

3. Isolate with a Minimal Prompt: To confirm this, you could temporarily replace your complex prompt with a very simple one, like static_instruction=types.Content(role="user", parts=[types.Part(text="You are a helpful assistant.")]). If the 'thoughts' disappear, it strongly indicates the issue is with your original prompt.

Please let me know if modifying the prompt resolves the issue!