The agent not following the workflow.

[JahangirJadi]

I have created a manager and define a workflow with subagents, now after delegating the task to first agent it stops after 1st step.
below is my instruction for the main agent.

You are a manager of a care home. Your primary task is to help the staff member creating an event by analyzing and classifying the description provided by staff.

Your workflow is as follows:
1. **Validate the event description**: Ensure that the description is clear and contains all necessary information (e.g., name or initials of resident/staff/manager, date, location (optional), etc.). If the description is unclear or missing critical information, provide feedback to the user. To do this use the `event_validator` tool. In the end, return the control the the parent agent.
2. **Extract relevant information**: If the event description is valid, extract key details such as the eventType(can be `resident` or `staff`) and names of residents or staff mentioned. Use the `event_information_extractor` tool for this task. If no resident or staff is found, return an error message to the user.
3. **Classify the event**: Based on the extracted information, classify the event using the `event_classifier` tool.

I am new to Agent development. So any help would be appreciated. Thanks

[adk-bot]

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

TLDR: Your workflow is stopping because the instruction in step 1, "return the control to the parent agent," is telling the agent to end its process. Removing this phrase will allow the workflow to proceed to the next steps as intended.

---

Hello! Based on the information you've provided, I can see why your workflow is not proceeding past the first step. Here’s a breakdown of the issue and how to resolve it.

The problem lies in this part of your instruction for the first step:

"In the end, return the control to the parent agent."

In the context of ADK workflows, this instruction is interpreted by the agent as an explicit command to stop its current sequence of tasks and hand off control. This effectively terminates the workflow, preventing it from moving on to steps 2 and 3.

ADK is designed to execute a series of sub-agents in a predefined order. [1] Once a step is completed successfully, the framework automatically proceeds to the next one in the sequence. You don't need to explicitly tell the agent to continue; you only need to ensure you don't tell it to stop.

To fix this, you should remove the instruction to return control. You can also rephrase the instruction to be more explicit about what to do if the validation is successful, for example, by simply stating that the process should continue.

Here is a revised version of your workflow instructions:

You are a manager of a care home. Your primary task is to help the staff member creating an event by analyzing and classifying the description provided by staff.

Your workflow is as follows:
1. **Validate the event description**: Ensure that the description is clear and contains all necessary information (e.g., name or initials of resident/staff/manager, date, location (optional), etc.). To do this, use the `event_validator` tool. If the description is unclear or missing critical information, provide feedback to the user. Otherwise, proceed to the next step.
2. **Extract relevant information**: If the event description is valid, extract key details such as the eventType(can be `resident` or `staff`) and names of residents or staff mentioned. Use the `event_information_extractor` tool for this task. If no resident or staff is found, return an error message to the user.
3. **Classify the event**: Based on the extracted information, classify the event using the `event_classifier` tool.

By making this small change, the agent will understand that after successful validation in step 1, it should continue to step 2, allowing your workflow to execute as you originally intended.

For more details on how to structure multi-agent systems and control the flow of execution, you can refer to the official ADK documentation on multi-agent systems. [1]

Let me know if you have any other questions!

---

This response was generated by an automated assistant. Please note that the information provided is based on the available documentation and may not cover all edge cases or scenarios.

[1] https://google.github.io/adk-docs/agents/multi-agents/

[JahangirJadi]

I tried the solution but it didn't worked. It's stopping on step 1. For clarity below is my sub agent:

event_validator_agent = Agent(
    name="event_validator",
    model="gemini-2.0-flash",
    description="Event validator agent",
    instruction="""
    You are an event identifier agent that validates events from the event description.
    You should validate the event based on the following criteria:
    1. The event description should be clear.
    2. The event description should not have any missing information like name or name initials of the resident/staff/manager, date, location(optional) and etc.
    3. Occasionally user can use initials of the entities (resident/staff/manager).

    If the description is invalid return the reason to the user.

    
      """
    
)```