Orchestrating Multi Agent Systems

Thu Ya Kyaw

Senior Developer Relations Engineer

LLM in Action

Prompt

LLM (Gemini)

Response

Hallucination

Google I/O Extended 25

Proprietary & Confidential

AI Agent Evolution

LLM + Prompt

LLM + Retrieval

LLM + Retrieval + Actions

+ Many Tools & Reasoning Loop

Multi Agent Systems

LLM

LLM

RAG

LLM

RAG

Tools

LLM

RAG

Tools

Google I/O Extended 25

Proprietary & Confidential

AI Agent in a nutshell

An application that reasons on how to best achieve a goal based on inputs and tools at its disposal.

3 main components:

• Model
• Tools
• Orchestration

​

�

Generative AI Models

(An agent can use multiple models)

APIs

Functions

Databases

Agents

Tools

Profile, goals & instructions

short-term

long-term

Memory

Model based

Reasoning/Planning

(Question Decomp & Reflection)

Orchestration

Google I/O Extended 25

Proprietary & Confidential

Google �Search

Agent

# agent.py

​

from google.adk.agents import Agent

from google.adk.tools import google_search

​

root_agent = Agent(

name="google_search_agent",

model="gemini-3-pro-preview",

instruction=(

"You are a helpful agent who helps user with their queries.",

"Use 'google_search' tool to get the latest and accurate information."

),

tools=[google_search]

)

​

​

pip install google-adk

​

adk create agent

Google �Search

Agent

​

​

Google AI Studio

# __init__.py

​

from . import agent

​

—-------------------------------------------------

​

# .env

​

GOOGLE_GENAI_USE_VERTEXAI=FALSE

GOOGLE_API_KEY="replace-with-your-key"

​

Google �Search

Agent

​

​

Vertex AI

# __init__.py

​

from . import agent

​

—-------------------------------------------------

​

# .env

​

GOOGLE_GENAI_USE_VERTEXAI=TRUE

GOOGLE_CLOUD_PROJECT="your-project-id"

GOOGLE_CLOUD_LOCATION="your-project-location"

goo.gle/adk-foundation

Codelab

goo.gle/adk-using-tools

Codelab

Google I/O Extended 25

Proprietary & Confidential

Google I/O Extended 25

Proprietary & Confidential

Google I/O Extended 25

Proprietary & Confidential

Google I/O Extended 25

Proprietary & Confidential

goo.gle/adk-orch-lab

Notebook

Google Cloud

Google Cloud

goo.gle/adk-mcp-a2a

Codelab

Deploy to Cloud Run

​

​

​

# Deploy to Cloud Run

​

adk deploy cloud_run \

--project=$GOOGLE_CLOUD_PROJECT \

--region=$GOOGLE_CLOUD_LOCATION \

--service_name=$SERVICE_NAME \

--app_name=$APP_NAME \

--with_ui \

$AGENT_PATH

Deploy to Agent Engine

​

​

​

# Deploy to Agent Engine

​

adk deploy agent_engine \

--project=$GOOGLE_CLOUD_PROJECT \

--region=$GOOGLE_CLOUD_LOCATION \

--staging_bucket=$STAGING_BUCKET \

--display_name=$APP_NAME \

$AGENT_PATH

Jensen Huang, CEO of NVIDIA

“Everything that moves will be autonomous someday, whether partially or fully.”

Google I/O Extended 25

Proprietary & Confidential

Google I/O Extended 25

Proprietary & Confidential

Google I/O Extended 25

Proprietary & Confidential