Ability to test individual local agents apart from the entire workflow

[ansu86d]

At the end of the agent.py file, the root_agent is the one which has the workflow - sequential , parallel etc . But these workflows have a group of agents where each agent might have sub_agents and tools to them. In most ocassions we want to tune or run agents independently without the workflow( of course , ensuring dependencies are resolved like output_key, state , session etc) . I am not ware if such a mode exits because the output of individual agents is very crucial to make the final output deterministic , is there any approach to do it or can we have some tooling around it.
Could not ask the question in the ADK community call - hence posting it here

[HonzaKopecky]

Our approach to this is to have each subagent in a separate folder, so the structure is like this:

- subagents
   - subagent1
      - __init__py
      - agent.py
   - subagent2
      - __init__py
      - agent.py
   - ...
- __init__.py
- agent.py (root agent)

To debug the subagents, you can run adk web from within subagents folder and all the subagents will be available for debugging. The key to make this work is that each subagent (in their agent.py) must define root_agent "alias", so ADK can properly import it. Our agent.py of each agent looks like this:

/subagents/subagent1/agent.py

subagent1 = LlmAgent(...)

root_agent = subagent1

But when we import the subagents into our root agent, we would import it by their names for better readability:

/agent.py

from subagent1.agent import subagent1

root_agent = LlmAgent(
   ....,
   subagents=[subagent1]
)