Google ADK "supervisor" architecture · AI Developer Accelerator

Google ADK "supervisor" architecture

Hi everyone, I’m currently testing Google’s ADK and building my first multi-agent system (a customer service agent).

My idea is to have a root agent that communicates directly with the customer, while the sub-agents only interact with or assist the root agent (not the customer directly).

The problem I’m facing is that all agents are responding to the customer, which breaks some intended features. For example:

The root agent can open a ticket using a tool when it cannot solve the customer’s problem.
If the customer’s prompt is about an order and the root agent delegates it to the “orders” sub-agent, the sub-agent should send its answer back to the root agent.
However, what happens now is that the sub-agent replies directly to the customer saying it can’t do anything, instead of sending this information to the root agent, which should then decide whether to answer or propose opening a ticket.

Has anyone faced this issue before or knows how to make only the root agent respond to the customer?

15 comments