The operation process of the intelligent agent

[Farewell-CK]

For instance, I employed multi-agent technology. Agent A is responsible for task distribution, Agent B is responsible for data preprocessing, Agent C is responsible for other tasks, and so on. Finally, there is an intelligent agent responsible for organizing the work of all the previous intelligent agents. Then I would like to know which specific agent is ultimately responsible for the delivery? Is it A or the last one? If someone could tell me, I would be extremely grateful. I am looking forward to receiving your notification regarding this process.

[Farewell-CK]

第一种情况：

graph TD
    subgraph 多智能体协作流程
        A[A 智能体: 任务分发] --> B[B 智能体: 数据预处理];
        A --> C[C 智能体: 其他工作];
        A --> D[...]
        
        subgraph 工作执行
            B -- 处理完成 --> E{汇总};
            C -- 处理完成 --> E;
            D -- 处理完成 --> E;
        end

        E -- 所有结果 --> A;
        A -- 整合后 --> F[🏆 最终交付];
    end

Loading

第二种情况：

graph TD
    subgraph 多智能体协作流程
        A[A 智能体: 任务分发] --> B[B 智能体: 数据预处理];
        B -- 预处理数据 --> C[C 智能体: 其他工作];
        C -- 处理结果 --> D[D 智能体: 整合与整理];
        D -- 整合所有工作 --> E[🏆 最终交付];
    end

Loading

[Farewell-CK]

Sorry, now I have understood this principle. However, I encountered a rather strange problem. When I used ADK to build a customer service system, a pre-sales agent, a post-sales agent, and a quotation agent. During the final execution, I would receive one or more similar responses - that is, multiple agents replied to me. But I only wanted one answer, yet I received three answers.

[DeanChensj]

Hi @Farewell-CK , for your 1st question, ADK use the handoff pattern so the last agent will return the result.

For your 2nd question, how is your agent tree look like? Do you have a minimal example to reproduce the issue?

[Farewell-CK]

I am glad to receive your reply. The handoff you just mentioned is shown in the green box. When using these agents, does the handoff exist by default? I understand that the handoff will be done by default. You also mentioned mode switching. Can you tell me in detail how to do this mode switching? Or where can I see the relevant content of mode switching in the document?

[DeanChensj]

Hi @Farewell-CK , I'm not sure what do you mean by mode switching.

[Farewell-CK]

Hi @DeanChensj Look at the red box. Thanks!

[DeanChensj]

You might take a look at https://google.github.io/adk-docs/agents/multi-agents/ for multi-agent system