Selective context for each Agent

[Sanjaykrishnamurthy]

In most scenarios I don't want each agent to know the entire chat history bloating the context window. It will be great if we can feed selective chat history

[wuliang229]

@Sanjaykrishnamurthy currently we have context compaction support in ADK to compress the previous chat history. Does that solve your problem? If not, could you be more specific about what is the behavior you are looking for?

[Sanjaykrishnamurthy]

Hi @wuliang229 , thanks for the reference. I've used context compaction but still my application takes a long time to respond especially as the conversation grows lengthier.
Consider i am building a multi-agent chat system with a main router and subagents, and each of these sub-agents are specialized for a particular task depending on the user query and the router routes based on the user query. In such a scenario, I don't want the sub-agent B to know what happened in the conversation before with sub-agent A. I'm not very sure but in LangGraph has the option to explicitly pass or not pass the response to the state variable thereby avoiding context bloat. If a similar feature to control the chat history is available it would be great to manage context window. In the current implementation of ADK all the agents in the multi-agent system get to see all of the past chats

[wuliang229]

@Sanjaykrishnamurthy you could use an AgentTool that wraps around one of your sub-agents to limit the context to only the latest request. You can also use before_agent_callback to redact any previous chat history as well. If you just need the latest request sent to your sub-agent, LlmAgent.include_contents might be of interest. Do any of these work for your use case?

Thanks for the feedback and comparison with LangGraph. We will look into whether that behavior is possible in ADK. Btw, we are working on something similar, so stay tuned. But right now, you can try the above-mentioned solutions.

cc @Jacksunwei to add more if needed.

[ashubham]

OpenAI Agents SDK does this for example. https://openai.github.io/openai-agents-python/handoffs/#input-filters