AI Folks
  • Courses
  • Web Of AI
  • Pricing
  • Enterprise
Log in
PreviousThe Claude Agent SDKNextGuardrails and approval

Vibe Coder · Unit 16: AI Agents · Lesson 8 of 9

Google ADK

Engineers: Google’s Agent Development Kit builds agents around Gemini, with workflow agents for sequences, loops and parallel work, and deployment to Google Cloud.

Read the short notes and pictures.Do everything inside a blue frame.

A kit for agent systems

Google’s Agent Development Kit (ADK) is an open source framework for building, debugging and deploying agents. It is available in Python, TypeScript, Go, Java and Kotlin, which matters if your team is not a Python team. Its distinctive idea is that agents can be composed: some agents do the thinking and others exist only to organise them.

A first agent

Terrazzo, hand-drawn. A code window: import Agent and google_search, then create an agent with a name, a model, an instruction and a list of tools.
Name, model, instruction, tools.

Install

pip install google-adk

agent.py

from google.adk import Agent
from google.adk.tools import google_search

agent = Agent(
    name="researcher",
    model="gemini-flash-latest",
    instruction="You help users research topics thoroughly.",
    tools=[google_search],
)
The same four ingredients as the other kits. Here the search tool comes ready-made.

Model names move

The model string above is the one in ADK’s own example at the time of writing. Model names change often. Copy the current one from the docs rather than from this page.

Workflow agents

Not every step needs a model deciding what happens next. ADK has workflow agents whose behaviour is fixed, and which exist to run other agents in a pattern.

Terrazzo, hand-drawn. Three cards: sequential, one after another; parallel, at the same time; loop, repeat until done.
Workflow agents organise other agents.

Sequential

Runs agents one after another, passing results along. A pipeline: research, then write, then check.

Parallel

Runs agents at the same time, for independent work such as researching three competitors at once.

Loop

Repeats agents until a condition is met, such as a checker approving the draft.

This is the workflow-or-agent choice from the first lesson, inside one framework: fixed structure where you know the steps, model judgement where you do not.

The three kits side by side

What is the same

An agent = model + instructions + tools.
Tools are functions with descriptions.
A runner executes the loop.
Runs can be traced.

What differs

OpenAI Agents SDK: handoffs, guardrails,
  tracing built in.
Claude Agent SDK: built-in file and shell
  tools, permissions.
Google ADK: workflow agents, five
  languages, Google Cloud deployment.
Choose by what your product needs and which model you rely on, not by brand.
  1. 1

    Deploy in a container

    Runs anywhere you can run containers.

  2. 2

    Cloud Run

    A managed service for containers on Google Cloud.

  3. 3

    Agent Runtime or GKE

    Google’s managed runtime for agents, or Kubernetes if you already use it.

ActivityYour turnWrite it yourself first, then compare.

You need to produce a weekly competitor report: gather facts on three competitors, write a summary, and have it checked. Sketch this with ADK’s workflow agents.

ActivityBuild stepAn agent that works for your product

Build the same agent with ADK and compare

Rebuild your agent with Google ADK using the same instructions and one equivalent tool. Run the same test script against all the versions you have built and record the results in a table.

Done when

This comparison is good material for your project write-up.

Keep these

  • ADK is open source and comes in five languages.
  • Sequential, Parallel and Loop agents organise other agents.
  • All three kits share one idea: model, instructions, tools.
Previous lessonThe Claude Agent SDKNext lessonGuardrails and approval