Question on _find_agent_to_run

[maheshwaritanay]

I am trying to go through and understand the code of adk-python to help in contribution. I am having difficulty in understanding the logic of _find_agent_to_run function - [https://github.com/google/adk-python/blob/0094eea3cadf5fe2e960cc558e467dd2131de1b7/src/google/adk/runners.py#L891](code link)
Specifically the qualified agent doc string, how can the "last user message" be a function response to a function call

[adk-bot]

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

TLDR: The _find_agent_to_run function is part of the ADK's orchestration logic that decides which agent should act next. The "last user message" it evaluates can be a FunctionResponse because the ADK framework treats the output of a tool as the latest input in the conversation flow. This allows an agent's execution to be guided by the results of the tools it calls.

---

Hello! That's a great question that gets to the core of how the ADK orchestrates tasks. Let me break down the logic for you.

The Agent Development Kit (ADK) operates on an event-driven "Reason-Act" loop. [5, 12] In this loop, an agent:

1. Reasons: Decides what to do based on the current conversation history. This often involves an LLM call.
2. Acts: The action can be responding to the user, or calling a tool (a function) to get more information or perform a task.

When a tool is called, the following happens:

1. The agent's LLM generates a FunctionCall event. [8]
2. The ADK framework executes the corresponding Python function (the tool).
3. The return value of that function is wrapped in a FunctionResponse event and added to the conversation history. [4, 14]

This is where your question comes in. The orchestrator, in functions like _find_agent_to_run, looks at the latest event in the history to decide the next step. After a tool is executed, the latest event is the FunctionResponse. The framework treats this response as the new "user message" for the purpose of deciding what to do next.

Why is this useful?

This mechanism allows for powerful, multi-step agent behaviors:

• Chaining: An agent can use a tool (e.g., search_flights) and the FunctionResponse (the flight details) can be used to prompt the same or another agent to take the next step (e.g., book_flight).
• Routing: In a multi-agent system, the output of a tool can determine which specialized agent should handle the next part of the conversation. A router agent might see a FunctionResponse containing financial data and decide to transfer_to_agent to a "FinancialAnalyst" agent. [1, 10]

So, while it seems counterintuitive at first, treating a FunctionResponse as a "user message" is a key design choice that enables the ADK's orchestrator to string together complex sequences of tool use and agent collaboration. The logic is designed to process a sequence of events, where the output of one step becomes the input for the next, regardless of whether that output was generated by a human user or a tool.