Google Gemini

Google ADK agents, built and deployed

Multi-agent systems on Google’s Agent Development Kit, deployed to Vertex AI Agent Engine or Cloud Run.

Google ADK agent hierarchy: coordinator, refund pipeline, researcher, policy checker, resolver
The agent hierarchy defined in the code below, built against google-adk.

How we use Google ADK

Google’s Agent Development Kit (ADK) is a code-first framework for building agents and multi-agent hierarchies, with built-in workflow agents, tools, session state and evaluation. It fits teams on Google Cloud and Gemini, and it also works with other models.

What we build with it

  • Gemini-powered assistants over Google Workspace and BigQuery data
  • Multi-agent hierarchies with coordinator and specialist agents
  • Agents that call your APIs through function, OpenAPI and MCP tools
  • Agents exposed to other agents over the A2A protocol
See it working

Google ADK, in code

The exact agent definitions, built and checked against google-adk here. Running them needs a Gemini API key, so this shows the structure the code produced, not a live model run.

adk_support_agents.pyCoordinator, sequential pipeline and tool-using agents
"""Google ADK multi-agent hierarchy used on the vithupro.in Google ADK page.
Built (not run) here to verify the definitions against the installed google-adk; running needs a Gemini key."""
from google.adk.agents import LlmAgent, SequentialAgent


def get_order(order_id: str) -> dict:
    """Look up an order by ID (read-only)."""
    return {"id": order_id, "status": "delivered", "damaged": True}


def issue_refund(order_id: str) -> dict:
    """Issue a refund. Guarded by a before_tool_callback that requires human approval."""
    return {"refund": f"{order_id}-ok"}


researcher = LlmAgent(
    name="researcher", model="gemini-2.5-flash",
    instruction="Find the order and summarise its status.", tools=[get_order], output_key="order",
)
policy = LlmAgent(
    name="policy_checker", model="gemini-2.5-flash",
    instruction="Check the refund policy for {order} and cite the clause.", output_key="policy",
)
resolver = LlmAgent(
    name="resolver", model="gemini-2.5-pro",
    instruction="Decide using {order} and {policy}. Only call issue_refund when the policy allows it.",
    tools=[issue_refund],
)
support = SequentialAgent(name="refund_pipeline", sub_agents=[researcher, policy, resolver])
coordinator = LlmAgent(
    name="coordinator", model="gemini-2.5-flash",
    instruction="Route refund requests to refund_pipeline; answer simple questions yourself.",
    sub_agents=[support],
)
root_agent = coordinator  # `adk web` / Vertex AI Agent Engine entry point

if __name__ == "__main__":
    def show(agent, depth=0):
        tools = [t.__name__ for t in getattr(agent, "tools", []) or []]
        print("  " * depth + f"{agent.name} ({type(agent).__name__})" + (f" tools={tools}" if tools else ""))
        for sub in agent.sub_agents:
            show(sub, depth + 1)
    show(root_agent)
Output of this run
coordinator (LlmAgent)
  refund_pipeline (SequentialAgent)
    researcher (LlmAgent) tools=['get_order']
    policy_checker (LlmAgent)
    resolver (LlmAgent) tools=['issue_refund']

Patterns we rely on

LLM agents and workflow agents

LLM agents for reasoning, plus Sequential, Parallel and Loop agents for predictable orchestration.

Agent hierarchies

A coordinator delegates to sub-agents, each with its own instructions and tools.

Sessions, state and memory

Session state for the current task and memory services for what should persist across sessions.

Callbacks and evaluation

Callbacks for guardrails and logging, and ADK evaluation sets to test agent trajectories before release.

Our production checklist

What we put in place before a Google ADK system goes live.

  • Tool definitions with clear schemas
  • Guardrail callbacks before and after model and tool calls
  • Evaluation sets for key user journeys
  • Deployment to Vertex AI Agent Engine or Cloud Run
  • Cloud Logging and tracing
  • Model choice per agent, Gemini or others via LiteLLM

Questions

Do we have to use Gemini with ADK?

No. ADK is optimised for Gemini but supports other models, so we pick per agent based on quality and cost.

ADK or LangGraph?

ADK if you are on Google Cloud and want managed deployment with Agent Engine. LangGraph when you need fine control over state and checkpoints, or run outside Google Cloud.

Often used together with:

Need Google ADK engineers?

Tell us what it needs to do. We reply within one business day with questions, a rough plan and the next step.

Book a call
  • Free 30-minute architecture call
  • NDA before you share details
  • You own all code and IP
  • Reply within one business day