I’m working on a use case where an LLM should first analyze a user query, extract intent and parameters, and then drive a workflow of sub-agents based on that reasoning. I’m trying to confirm the recommended ADK orchestration pattern to accomplish this effectively.
Key Questions
-
Architecture Best Practice
What is the recommended pattern for combining LlmAgent-based reasoning with sequential or conditional execution in ADK?
-
Parameter Passing
How should user intent and extracted parameters be passed from a parent reasoning agent to workflow agents?
-
Instruction / Prompt Design
What is the best practice for structuring instructions so the LLM cleanly delegates tasks to workflow agents?
-
Alternatives
Are there other ADK agent types or workflow control patterns better suited for intent-driven execution?
What I Have Already Tried (Without Code)
- Using a LlmAgent as a top-level orchestrator for reasoning followed by delegation
→ Issue: Often answers directly or delegates without consistent parameter extraction
-Using custom instructions to enforce a reasoning-first approach
→ Issue: Parameter passing and control of specific workflow steps remains unreliable
What I’m Looking For
- Recommended ADK architecture for:
- Intent extraction before workflow execution
- Conditional agent selection
- Passing parameters into workflow agents
- Executing only relevant sub-agents
-Guidance or example demonstrating:
- Reasoning agent (intent + parameter extraction)
- Workflow agent (dynamic execution based on extracted insights)
Environment
- Google ADK: Latest release
- Python: 3.12+
- Production context: Healthcare workflow automation requiring dynamic, intent-driven execution
Additional Context
This orchestration pattern appears essential for multi-agent production applications where workflows must be efficient, context-aware, and driven by user input — not static execution sequences.
Documentation or a sample implementation would be extremely valuable for adopters building complex systems on ADK.
I’m working on a use case where an LLM should first analyze a user query, extract intent and parameters, and then drive a workflow of sub-agents based on that reasoning. I’m trying to confirm the recommended ADK orchestration pattern to accomplish this effectively.
Key Questions
Architecture Best Practice
What is the recommended pattern for combining LlmAgent-based reasoning with sequential or conditional execution in ADK?
Parameter Passing
How should user intent and extracted parameters be passed from a parent reasoning agent to workflow agents?
Instruction / Prompt Design
What is the best practice for structuring instructions so the LLM cleanly delegates tasks to workflow agents?
Alternatives
Are there other ADK agent types or workflow control patterns better suited for intent-driven execution?
What I Have Already Tried (Without Code)
→ Issue: Often answers directly or delegates without consistent parameter extraction
-Using custom instructions to enforce a reasoning-first approach
→ Issue: Parameter passing and control of specific workflow steps remains unreliable
What I’m Looking For
-Guidance or example demonstrating:
Environment
Additional Context
This orchestration pattern appears essential for multi-agent production applications where workflows must be efficient, context-aware, and driven by user input — not static execution sequences.
Documentation or a sample implementation would be extremely valuable for adopters building complex systems on ADK.