5.3 ADK Callbacks & Human-in-the-Loop (HITL) Controls

1. The 60-Second Anchor

Cloud IAM and OAuth 2.0 only check who is calling a tool—they cannot look inside the tool’s parameters (like amount = 50000). To inspect live payloads and enforce deterministic guardrails inside the Agent Development Kit (ADK), Google Cloud gives you two runtime callback hooks:

Before-Model Callback (before_model_callback): The Thought Guard. Runs before the request reaches the Gemini LLM. Use it to check incoming context/prompts, inject dynamic system instructions, or return a pre-built response (LlmResponse) to skip calling the LLM completely!
Before-Tool Callback (before_tool_callback): The Action Guard. Runs after Gemini decides to call a tool, but before the tool actually executes. Use it to inspect exact tool arguments (args), block dangerous parameter values, or pause execution for Human-in-the-Loop (HITL) approval!

2. Interactive Showdown: Before-Model vs. Before-Tool Callbacks

Toggle between the 3 Exam Callback Scenarios below and step through the 4 stages to see exactly what ADK inspects—and what return values (None, LlmResponse, or dict) control execution on the exam:

3. Exam Cheat Sheet: The ADK Callback Showdown Matrix

Memorize the exact inputs and return values for both ADK callbacks—the exam loves testing what happens when a callback returns None versus an override object!

Feature	`before_model_callback` (Thought Guard)	`before_tool_callback` (Action Guard)
When Does It Run?	Right before the request is sent to the Gemini LLM.	Right before a selected tool function or MCP tool is executed.
What Inputs Does It Inspect?	`callback_context` and `llm_request` (the prompt, conversation history, and system instructions).	`tool`, `args` (the dictionary of tool parameters like `amount=50000`), and `tool_context`.
What Happens If It Returns `None`?	Proceeds normally: Sends the `llm_request` to the Gemini model.	Proceeds normally: Executes the actual tool with the provided `args`.
How Do You Block / Override Execution?	Return an `LlmResponse` object (skips calling Gemini completely and sends your custom response).	Return a `dict` (skips running the tool completely and passes your dictionary back to the agent as the tool output).

4. High-Yield Exam Split: The 2 Types of Human-in-the-Loop (HITL)

On the PAAC exam, “Human-in-the-Loop” can mean two very different architectural patterns depending on whether you are pausing a single high-risk tool call or escalating an entire customer conversation. Here is how to pick the right answer every time:

HITL Pattern	1. ADK Tool-Level HITL (Action Approval)	2. Customer Engagement Suite (CES) Conversation-Level HITL (Live Agent Handoff)
What Gets Paused or Transferred?	Only the specific high-risk tool execution (e.g., wiring $50,000, deleting a production database, or issuing a large refund).	The entire live chat or voice conversation is handed off to a human contact-center rep.
How It Works Under the Hood	Implemented via ADK’s `before_tool_callback` or built-in Tool Confirmation flow. The callback intercepts `args`, pauses execution, requests human sign-off on the parameters, and resumes once approved.	Configured in CX Agent Studio / Customer Engagement Suite (CES) using escalation rules triggered by negative sentiment, repeated fallback/no-match, or explicit user requests (“Talk to a human”).
What the Human Reviewer Sees	A structured Approval Request Payload showing the tool name, exact arguments (`amount: $50,000`), and Approve/Reject buttons.	The Full Conversation Transcript + AI-Generated Context Summary inside the human agent desktop so the customer doesn’t have to repeat themselves.
Exam Keyword Triggers	“High-value transaction,” “sensitive API mutation,” “approve tool arguments,” “deterministic parameter threshold.”	“Frustrated customer,” “contact center escalation,” “live agent warm transfer,” “omnichannel voice/chat support.”