Automatic Task Decomposition in Google ADK Multi-Agent System

[kolokol427]

I am developing a proof of concept with Google ADK for a multi-agent system composed of several specialized agents:

• orchestrator_agent: coordinates requests
• contact_manager: handles contact information
• internal_system_notification_agent: sends internal notifications
• weather_agent: provides weather data
• time_agent: returns date and time

---

Current Flow and Limitation

When receiving a request such as:

“Find John Smith’s internal code and send him a message saying the meeting tomorrow will be on the 5th floor.”

the framework identifies that it should route the request to contact_manager, retrieves the internal code, and stops the process there and ask the user to send the message. It does not automatically proceed to the notification agent to send the message.

---

Goal

I want the orchestrator_agent to automatically detect and manage composite tasks so that it can:

1. Decompose the request into independent subtasks.
2. Determine dependencies between the subtasks.
3. Assign and execute each subtask with the appropriate specialized agent.
4. Maintain context throughout the entire process.
5. Consolidate the results and return a single response to the user.

The orchestrator_agent should:

• Call contact_manager to fetch the internal id.
• Once the id is returned, invoke internal_system_notification_agent to send the message with the ID.

---

Questions

1. What strategy does Google ADK recommend for implementing automatic task decomposition?
2. How can I model dependencies and flows within the orchestrator?
3. Are there any documented examples, design patterns, or use cases demonstrating this functionality?

Thank you in advance for any guidance or references to examples!

[boyangsvl]

In your case I think one agent with multiple tools may work better. Usually you want fewer LLM agents involved in serving one query, since the hand-off between agents adds latency and cost, and they are not as reliable as tool calls.

[kolokol427]

We ended up opting to use tools directly to avoid the added latency. Thank you very much for your insight!

Quick question: do you have any recommendations for introducing sub‑agents under the main coordinator agent? In some scenarios, for the sake of architectural clarity, it can be better to have specialized agents each with their own set of tools.