Is there a way to ask the Agent to engage conversation ? Or to relaunch it ?

[charlesdassonville]

Hello !
I am trying to implement an agent that starts the conversation by itself, before the first user's message, when the session is created. In the same way, would it be possible to call the runner to relaunch the conversation if the user has not respond after a while ?

I have tried to implement this:

async_generator = my_runner.run_async(
        new_message=types.Content(parts=[types.Part(text="Start the conversation")]),
        user_id=user_id,
        session_id=session_id,
    )

But later on, the content is considered as a user message which is not what i want.

I also tried to update system instructions to tell the agent to send first message when a session is created but it does not work.

Finally, manually adding a "Hello there" with role = "model" message is not a convenient solution since it's not using agent.

Thanks for your help !

[adk-bot]

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

TLDR: The standard ADK runner is reactive and expects a user message to start the conversation. The recommended approach to have the agent "speak first" is to generate the opening message in a separate step and then manually add it to the conversation history with the role="model" before presenting it to the user.

---

Hello! That's an interesting use case.

The behavior you're seeing is expected because the standard ADK Runner is designed to be reactive. It waits for a user's input, processes it through the agent's reason-act loop, and then generates a response. This is why providing an initial message via run_async results in that message being treated as user input.

While the ADK doesn't have a built-in "agent-initiate" function for general-purpose runners, you can achieve your goal with a slight change in approach.

Proposed Solution: Pre-generating the First Message

Instead of trying to make the runner start the conversation on its own, you can separate the generation of the opening message from the main conversational loop.

1. Generate the Opening Message: When a new session is created, make a one-off call to your agent or the underlying LLM with a prompt like "Generate a friendly opening message to start a conversation."
2. Add to History: Take the generated text and manually add it to the session's history as the first turn, ensuring it has the role="model".
3. Start the Conversation: Now, when the user sees the session, the agent's message is already there. Their first message will be a reply, and from that point on, the Runner can proceed with its normal reactive loop.

This method correctly uses your agent's intelligence to craft the message, avoiding the "inconvenient" hardcoded solution you mentioned.

How to Relaunch a Conversation

You can use the exact same pattern to re-engage an inactive user. An external process (e.g., a scheduled job) could:

1. Check the timestamp of the last message in a session.
2. If the user has been inactive for a certain period, generate a re-engagement message using the agent.
3. Add this new message to the session history with role="model". This would then appear in the user's chat interface, prompting them to reply.

Proactive Features in the Live API

It's worth noting that for specific use cases like live audio streaming, the ADK does have a feature called ProactivityConfig. [1] When enabled, it allows the model to be more proactive, offer suggestions, and anticipate user needs. However, this is currently part of the Gemini Live API and is tailored for native audio models, so it may not apply to your use case.

I hope this workaround helps you implement the desired behavior in your agent!

[1] https://google.github.io/adk-docs/streaming/dev-guide/part5/