How to build a deep research agent for lead generation using Google’s ADK

In this architecture blueprint published on the Google Cloud Blog, I detail how to move beyond static scripts and brittle scrapers by building a sophisticated, multi-agent lead generation system utilizing Google’s Agent Development Kit (ADK).

Rather than relying on a single monolithic prompt, this system decomposes the problem into discrete phases, managing state transitions across user interactions and orchestrating parallel execution workflows.

Architectural Breakdown:

• Step 1: The Primary Orchestrator: At the core of the architecture sits a root agent named InteractiveLeadGenerator. Running on gemini-2.5-pro, its job is to manage the high-level workflow, delegate to specialized sub-agents via localized tools, and interact directly with the user.
• Step 2: Intent Extraction: Before running complex workflows, the root agent utilizes a specialized intent_extractor_agent (running on gemini-2.5-flash) driven by an explicit INTENT_EXTRACTOR_PROMPT to parse the user’s initial request into structured data (IntentExtractionResult).
• Step 3: Dual-Squad Workflows: The system splits execution into two clear pipelines:
  • The Research Squad (Learning from the Past): A cohort of cooperative agents tasked with executing pattern discovery workflows to discover the signals of success across historical data.
  • The Hunter Squad (Predicting the Future): A cohort that uses those confirmed signals to execute the actual lead generation workflow and target matching prospects.
• Step 4: Control Mechanisms: To maintain deterministic boundaries, the system uses precise control mechanisms like before_agent_callback and after_tool_callback hooks to safely track state updates and maintain absolute conversation context.