Unable to setup dynamic agents for ParallelAgent

[ArnabChaudhury1-upl]

Example:
LLMAgent has subagent ParalleAgent with subagents: researcher_agent1, researcher_agent2, researcher_agent3. Based on decision by LLMAgent want parallel execution of researcher_agent2 and researcher_agent3 or any combination.

Any way to implement this?

[adk-bot]

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

TLDR: While ParallelAgent is designed to run all its sub-agents concurrently, you can achieve dynamic execution by implementing a callback function that conditionally skips agents based on criteria you define. This is a pattern used in ADK samples.

---

Hello! That's an excellent question about creating dynamic, parallel workflows.

By default, the ParallelAgent is a workflow agent that executes all of its sub-agents at the same time. [1, 2] This is different from an LLMAgent, which uses an LLM to decide what to do next.

However, you can achieve the dynamic behavior you're looking for. The recommended approach is to use a before_agent_callback to check if a sub-agent should run or be skipped. The workflow_triage sample is a great example of this, where it checks for "relevance" before executing an agent. [3]

Here’s a conceptual example of how you could apply this to your scenario:

1. LLMAgent Decides: Your primary LLMAgent would first determine which of the researcher agents are needed (e.g., researcher_agent2 and researcher_agent3). It could store this decision in the Conversation state.

2. Callback for a Check: You would define a callback function. Before each agent in the ParallelAgent runs, this function would check the Conversation state to see if that specific agent was selected by the LLMAgent.

3. Skip if Not Selected: If the agent was not selected, the callback would cause the agent to be skipped.

This approach allows you to maintain a single ParallelAgent in your structure while dynamically controlling which parts of it execute based on prior decisions.

Alternatively, for simpler cases, you could programmatically create a new ParallelAgent instance with only the selected sub-agents after your LLMAgent makes its decision. For the most complex scenarios, you could even build a CustomAgent for full control over dynamic agent selection. [4]

I hope this helps you build your dynamic multi-agent system!

---

[1] https://google.github.io/adk-docs/agents/multi-agents/
[2] https://google.github.io/adk-docs/agents/workflow-agents/parallel-agents/
[3] https://github.com/google/adk-python/blob/main/contributing/samples/workflow_triage/README.md
[4] https://google.github.io/adk-docs/agents/custom-agents/

[ArnabChaudhury1-upl]

@adk-bot would it possible to get a small implementation snippet for the solution you suggested using callbacks and/or custom agents?

[adk-bot]

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

TLDR: The standard ParallelAgent does not support dynamic selection of sub-agents to run. To achieve this, you will need to create a CustomAgent.

---

Hello! Based on the information in the ADK documentation, the behavior you're describing requires a CustomAgent. Here's a more detailed explanation:

The ParallelAgent is designed to execute all of its sub-agents concurrently. [2] It does not have a built-in mechanism to dynamically choose a subset of these agents to run at execution time.

For dynamic orchestration logic, such as choosing which sub-agents to run based on a prior decision, the recommended approach is to implement a CustomAgent. [1] A CustomAgent allows you to define your own execution logic.

To solve your use case, you could create a CustomAgent that:

1. Takes the list of all possible researcher agents as input.
2. Receives a list of agent names to execute from the parent LLMAgent.
3. In its execution method, filters the agents based on the list from the LLMAgent.
4. Executes the selected agents in parallel.

This approach gives you the full flexibility to control the parallel execution flow based on dynamic decisions made by other agents.

I hope this helps!

---

[1] Custom Agents
[2] Parallel Agent