GDG Cloud Southlake #47: Bala Desikan: Build Autonomous Agentic AI Apps on GCP
JamesAnderson135
15 views
36 slides
Oct 31, 2025
Slide 1 of 36
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
About This Presentation
Agentic AI on Google Cloud: Building Smart, Autonomous Applications
This presentation will demystify how AI is evolving beyond reactive prompts to autonomous, goal-driven systems. In this session, explore the foundational architecture of agentic systems, breaking down key components such as sophist...
Slide Content
Build Autonomous Agentic AI Apps on GCP
From Prototype to Production
From Prototype to Production
Proprietary & ConfidentialGoogle Cloud
Agent
Agent vs. Framework vs. System
Multi Agent Framework
Planner
sub-agent
LLM
Domain-agentDomain-agent
LLM+
+ Plugins
+ Tools
Multi Agent System
+ Agents
+ LLMs
User session User session
User session User session
Session manager
LLM+
Agent
Naive view from user POV
Planner creates / combines sub-agents
for task + selects tools, LLM, etc.
Manages agent combinations per
user & session
Agent
Agent vs. Framework vs. System
Multi Agent Framework
Planner
sub-agent
LLM
Domain-agentDomain-agent
LLM+
+ Plugins
+ Tools
Multi Agent System
+ Agents
+ LLMs
User session User session
User session User session
Session manager
LLM+
Agent
Naive view from user POV
Planner creates / combines sub-agents
for task + selects tools, LLM, etc.
Manages agent combinations per
user & session
Proprietary & ConfidentialGoogle Cloud
Typical Agent Builders Journey
Define a specific and narrow use case that solves a real problem
Define clear success metrics and how you’re going to measure them
Start with the simplest possible architecture and UX to address the core user outcome
PoC
MVP
Implement feedback and metrics collection mechanism
Build evaluation and testing pipelines
Iterate on prompts and architecture
Build at least one entire end-to-end critical user journey and test it with a set of users
Path
to
Prod
Extensive observability and tracing
Load testing and safety evals
CI/CD and continuous monitoring and evaluation (+ incident management)
Systematic collection of data for prompt improvements and/or fine-tuning
Typical Agent Builders Journey
Define a specific and narrow use case that solves a real problem
Define clear success metrics and how you’re going to measure them
Start with the simplest possible architecture and UX to address the core user outcome
PoC
MVP
Implement feedback and metrics collection mechanism
Build evaluation and testing pipelines
Iterate on prompts and architecture
Build at least one entire end-to-end critical user journey and test it with a set of users
Path
to
Prod
Extensive observability and tracing
Load testing and safety evals
CI/CD and continuous monitoring and evaluation (+ incident management)
Systematic collection of data for prompt improvements and/or fine-tuning
Proprietary & ConfidentialGoogle Cloud
Let’s build a customer support refund system
... At Crabby’s Taffy,
a popular candy company.
Right now, the customer refund system
is handled entirely via email, by two
customer service reps.
As a result, sometimes it can take days
for a customer to hear back about their
refund.
You want to transform the refunds system
using an AI agents, to...
1 - Boost customer satisfaction (send the
refund immediately, instead of days later.)
2 - Free up the two customer service reps to
focus on more complex tasks, like handling
requests for wedding favors.
Let’s build a customer support refund system
... At Crabby’s Taffy,
a popular candy company.
Right now, the customer refund system
is handled entirely via email, by two
customer service reps.
As a result, sometimes it can take days
for a customer to hear back about their
refund.
You want to transform the refunds system
using an AI agents, to...
1 - Boost customer satisfaction (send the
refund immediately, instead of days later.)
2 - Free up the two customer service reps to
focus on more complex tasks, like handling
requests for wedding favors.
Proprietary & ConfidentialGoogle Cloud
Prototyping magic w/ Google AI Studio
Google AI Studio is a browser-based integrated development environment
(IDE) for prototyping with generative models - testing input/output.
AI Studio lets you quickly try out models and experiment with different
prompts.
When you've built something you're happy with, you can export it to code in
your preferred programming language and use the Gemini API with it. AI
Studio permits to export what you created into Vertex.
What can you achieve with AI Studio?
•Generate creative content and innovative solutions.
•Gain deeper insights from your data.
•Build cutting-edge AI applications.
Prototyping magic w/ Google AI Studio
Google AI Studio is a browser-based integrated development environment
(IDE) for prototyping with generative models - testing input/output.
AI Studio lets you quickly try out models and experiment with different
prompts.
When you've built something you're happy with, you can export it to code in
your preferred programming language and use the Gemini API with it. AI
Studio permits to export what you created into Vertex.
What can you achieve with AI Studio?
•Generate creative content and innovative solutions.
•Gain deeper insights from your data.
•Build cutting-edge AI applications.
Proprietary & ConfidentialGoogle Cloud
From Prototype to MVP
LLMAgent: Agent relies entirely on the LLM
to reason its way through the task, trusting it
to call the right tools when needed.
SequentialAgent: Agent has a “hardcoded”
workflow that calls sub-agents in a fixed
order.
controlflexibility
Single-agent system: One agent with one
system prompt, set of tools, and model.
Multi-agent system: a system of experts
with their own goals, system prompts,
models, and tools.
simplicity modularity
There are lots of ways to build an agent…
from different control flows… to different numbers of agents
From Prototype to MVP
LLMAgent: Agent relies entirely on the LLM
to reason its way through the task, trusting it
to call the right tools when needed.
SequentialAgent: Agent has a “hardcoded”
workflow that calls sub-agents in a fixed
order.
controlflexibility
Single-agent system: One agent with one
system prompt, set of tools, and model.
Multi-agent system: a system of experts
with their own goals, system prompts,
models, and tools.
simplicity modularity
There are lots of ways to build an agent…
from different control flows… to different numbers of agents
Proprietary & ConfidentialGoogle Cloud
DIY
Low-level LLM framework
(Langchain)
Low-level orchestration framework
(LangGraph)
Agent framework
(Agent Development Kit)
MVP with ADK
client = OpenAI()
response = client.chat.completions.create(
model="gpt-3.5-turbo",
messages=[
{"role": "user",
"content": "What is the capital of France?"}
]
)
prompt = ChatPromptTemplate.from_messages([
("user", "{question}")
])
chain = prompt | llm
question = "What is the capital of France?"
capital = chain.invoke({"question": question})
workflow = StateGraph(GraphState)
workflow.add_node("llm_call", call_llm)
workflow.add_edge(START, "llm_call")
workflow.add_edge("llm_call", END)
Chain
Prompt Model
chain = prompt | llm
START
END
capital_finder_agent = LlmAgent(
name="capital_finder_agent",
model="gemini-1.5-flash-latest",
instruction="You are a helpful assistant
that provides the capital city for a given country."
)
question = "What is the capital of France?"
print(f"Calling ADK Agent with question: {question}")
Tool
Tool
Agent
Tool
Tool
Agent
Tool
Tool
Workflow Agent
Vendor Specific
Restful
DIY
Low-level LLM framework
(Langchain)
Low-level orchestration framework
(LangGraph)
Agent framework
(Agent Development Kit)
MVP with ADK
client = OpenAI()
response = client.chat.completions.create(
model="gpt-3.5-turbo",
messages=[
{"role": "user",
"content": "What is the capital of France?"}
]
)
prompt = ChatPromptTemplate.from_messages([
("user", "{question}")
])
chain = prompt | llm
question = "What is the capital of France?"
capital = chain.invoke({"question": question})
workflow = StateGraph(GraphState)
workflow.add_node("llm_call", call_llm)
workflow.add_edge(START, "llm_call")
workflow.add_edge("llm_call", END)
Chain
Prompt Model
chain = prompt | llm
START
END
capital_finder_agent = LlmAgent(
name="capital_finder_agent",
model="gemini-1.5-flash-latest",
instruction="You are a helpful assistant
that provides the capital city for a given country."
)
question = "What is the capital of France?"
print(f"Calling ADK Agent with question: {question}")
Tool
Tool
Agent
Tool
Tool
Agent
Tool
Tool
Workflow Agent
Vendor Specific
Restful
Proprietary & ConfidentialGoogle Cloud
ADK
Current Version
1.17.0 - Oct 22,2025
ADK Samples
ADK
Current Version
1.17.0 - Oct 22,2025
ADK Samples
Proprietary & ConfidentialGoogle Cloud 9
Architectural Patterns
Single
One LLM for all tasks
LLM
Input Output
Preprocessing Post
Processing
Sequential
Chains multiple specialized LLMs in a pipeline.
LLM
Input
Output
Intermediate
Result…
LLM
Hierarchical
Router model to direct inputs to specialized LLMs
based on content or intent
RouterInput
LLM
LLM
LLM
LLM
Output
Analysis and
Classification
processing
Intermediate
Result…
Parallel
Multiple LLMs concurrently and aggregate outputs
AggregatorInput Output
LLM
LLM
LLM
Architectural Patterns
Single
One LLM for all tasks
LLM
Input Output
Preprocessing Post
Processing
Sequential
Chains multiple specialized LLMs in a pipeline.
LLM
Input
Output
Intermediate
Result…
LLM
Hierarchical
Router model to direct inputs to specialized LLMs
based on content or intent
RouterInput
LLM
LLM
LLM
LLM
Output
Analysis and
Classification
processing
Intermediate
Result…
Parallel
Multiple LLMs concurrently and aggregate outputs
AggregatorInput Output
LLM
LLM
LLM
Proprietary & ConfidentialGoogle Cloud
Cooperative Multi-Agent System
Multiple LLM-based agents collaborating to solve problems
Workspace
Agent
Agent
Agent
Agent
Coordinator
Synthesizer
Problem
Solution
Multi-Agent Interaction
Competitive Multi-Agent System
Independent LLM agents to propose competing solutions,
with evaluator model selecting best or integrating elements
Agent
solution
Agent
solution
Agent
solution
Agent
solution
Problem
Evaluator
Selection
Solution
Cooperative Multi-Agent System
Multiple LLM-based agents collaborating to solve problems
Workspace
Agent
Agent
Agent
Agent
Coordinator
Synthesizer
Problem
Solution
Multi-Agent Interaction
Competitive Multi-Agent System
Independent LLM agents to propose competing solutions,
with evaluator model selecting best or integrating elements
Agent
solution
Agent
solution
Agent
solution
Agent
solution
Problem
Evaluator
Selection
Solution
Proprietary & ConfidentialGoogle Cloud
A Multi-agent, Custom Control Flow for refunds
?????? Customer Refund Agent
(Custom BaseAgent implementation)
session ID 18734830
state[purchases] = {...}
state[is_eligible] = true
✨ Gemini 2.5 Flash ✨
?????? Tool:
get_purchase_history
?????? Tool call:
check_refund_eligibility
?????? Tool call:
process_refund
?????? Final Response Agent
“Output the message.”
?????? Parallel Verifier Agent
(“Hardcoded” Workflow)
?????? Purchase Verifier
Agent
“You verify purchases..”
if
is_eligible
?????? Store Credit
Agent
“Offer an
alternative...”
?????? Process
Credit Decision
“Did they
accept?”
no yes
?????? Refund
Processor
Agent
“You send
refunds..”
?????? Refund Eligibility
Agent
“You check eligibility..”
?????? “Please
refund my
order...”
??????“Refund
successful!”
A Multi-agent, Custom Control Flow for refunds
?????? Customer Refund Agent
(Custom BaseAgent implementation)
session ID 18734830
state[purchases] = {...}
state[is_eligible] = true
✨ Gemini 2.5 Flash ✨
?????? Tool:
get_purchase_history
?????? Tool call:
check_refund_eligibility
?????? Tool call:
process_refund
?????? Final Response Agent
“Output the message.”
?????? Parallel Verifier Agent
(“Hardcoded” Workflow)
?????? Purchase Verifier
Agent
“You verify purchases..”
if
is_eligible
?????? Store Credit
Agent
“Offer an
alternative...”
?????? Process
Credit Decision
“Did they
accept?”
no yes
?????? Refund
Processor
Agent
“You send
refunds..”
?????? Refund Eligibility
Agent
“You check eligibility..”
?????? “Please
refund my
order...”
??????“Refund
successful!”
Key aspects to consider when you build
Dynamic Planning & Orchestration
Given the inherent complexity of agentic
workflows, meticulously plan for dynamic
planning and orchestration during serving
Feedback Loop
Implementing robust feedback loop tools
is vital for collecting performance
samples and continuously improving
agent accuracy.
Tool Calling
Effective tool calling mechanisms are crucial for
agents to both fetch necessary data and execute
actions seamlessly.
#5
#1
Memory & Personalization
Consider how agents will manage memory and
personalize interactions to provide relevant and
continuous experiences
#6
#2
Production Tracing
Tracing dynamic execution paths in
production is critical for troubleshooting
and understanding the real-time behavior of
agentic workflows.
#7
#3
Prompt-first Debugging
Facilitating a simple prompt-first iteration process
is essential for builders to easily inspect thinking
and debug streams during agent development.
Agent Evals
Comprehensive agent evaluations require
tools for evaluating plans, sophisticated
tool calling, precise parameter construction,
and often, automated code generation.
#4
Dynamic Planning & Orchestration
Given the inherent complexity of agentic
workflows, meticulously plan for dynamic
planning and orchestration during serving
Feedback Loop
Implementing robust feedback loop tools
is vital for collecting performance
samples and continuously improving
agent accuracy.
Tool Calling
Effective tool calling mechanisms are crucial for
agents to both fetch necessary data and execute
actions seamlessly.
#5
#1
Memory & Personalization
Consider how agents will manage memory and
personalize interactions to provide relevant and
continuous experiences
#6
#2
Production Tracing
Tracing dynamic execution paths in
production is critical for troubleshooting
and understanding the real-time behavior of
agentic workflows.
#7
#3
Prompt-first Debugging
Facilitating a simple prompt-first iteration process
is essential for builders to easily inspect thinking
and debug streams during agent development.
Agent Evals
Comprehensive agent evaluations require
tools for evaluating plans, sophisticated
tool calling, precise parameter construction,
and often, automated code generation.
#4
Proprietary & ConfidentialGoogle Cloud
Multi-agent: Custom Control Flow
Multi-agent: Custom Control Flow
Proprietary & ConfidentialGoogle Cloud
Multi-agent: Custom Control Flow
Multi-agent: Custom Control Flow
Proprietary & ConfidentialGoogle Cloud
ADK Agent - Base
Base Agent
Model
Name
Instruction
Tools
Sub Agents
Input Output
Video (Stream)
Database
Files
Signals
Audio (Stream)
Cache
Vector
API
gRPC
HTTP
AMQP
MQTT
TCP/IP
ROS
DDS
Out of Box
Vertex AI Search
(RAG)
Google Search
Custom
^&@!#%
Agents
Models
ADK Agent - Base
Base Agent
Model
Name
Instruction
Tools
Sub Agents
Input Output
Video (Stream)
Database
Files
Signals
Audio (Stream)
Cache
Vector
API
gRPC
HTTP
AMQP
MQTT
TCP/IP
ROS
DDS
Out of Box
Vertex AI Search
(RAG)
Google Search
Custom
^&@!#%
Agents
Models
Proprietary & ConfidentialGoogle Cloud
ADK Agent - Base
from google.adk.agents import Agent
from tools.tools import get_purchase_history, check_refund_eligibility, process_refund
from tools.prompts import top_level_prompt
logger = logging.getLogger(__name__)
GEMINI_MODEL = "gemini-2.5-flash-preview-05-20"
root_agent = Agent(
model=GEMINI_MODEL,
name="RefundSingleAgent",
description="Customer refund single-agent for Crabby's Taffy company",
instruction=top_level_prompt,
tools=[get_purchase_history, check_refund_eligibility, process_refund],
)
logger.info(f"Initialized {root_agent.name}")
ADK Agent - Base
from google.adk.agents import Agent
from tools.tools import get_purchase_history, check_refund_eligibility, process_refund
from tools.prompts import top_level_prompt
logger = logging.getLogger(__name__)
GEMINI_MODEL = "gemini-2.5-flash-preview-05-20"
root_agent = Agent(
model=GEMINI_MODEL,
name="RefundSingleAgent",
description="Customer refund single-agent for Crabby's Taffy company",
instruction=top_level_prompt,
tools=[get_purchase_history, check_refund_eligibility, process_refund],
)
logger.info(f"Initialized {root_agent.name}")
Proprietary & ConfidentialGoogle Cloud
ADK - Runner
Base Agent
runner = Runner(
app_name='Agent Refund System',
agent=root_agent,
session_service=VertexAI
)
runner.run_async(
session_id=session.id,
user_id=session.user_id,
new_message=content
)
Sub Agent
Sub Agent
Session config
Runner
ADK - Runner
Base Agent
runner = Runner(
app_name='Agent Refund System',
agent=root_agent,
session_service=VertexAI
)
runner.run_async(
session_id=session.id,
user_id=session.user_id,
new_message=content
)
Sub Agent
Sub Agent
Session config
Runner
Proprietary & ConfidentialGoogle Cloud
Short Term
State
Events
Session
ADK - Memories
^&@!#%
756798
)($$%&*)
^&@!#%
user messages
756798
agent response
&^)**%%
&^)**%%
user messages
HGFKUY
yiueuecall tool
yiueue
tool response
yiueue abc
key value
yiueuestate change
384723
384723
agent response
)($$%&*)
user messages
Long/Short Term (Your Call) - Memory
In memory
Vertex AI
Session
Session
Session
And artifacts…
Short Term
State
Events
Session
ADK - Memories
^&@!#%
756798
)($$%&*)
^&@!#%
user messages
756798
agent response
&^)**%%
&^)**%%
user messages
HGFKUY
yiueuecall tool
yiueue
tool response
yiueue abc
key value
yiueuestate change
384723
384723
agent response
)($$%&*)
user messages
Long/Short Term (Your Call) - Memory
In memory
Vertex AI
Session
Session
Session
And artifacts…
Proprietary & ConfidentialGoogle Cloud
ADK - Runner
Base Agent
Runner
Initiate/Get Session
Invoke agent
Memory
Update Session
Invoke Tool
Event
Event
ADK - Runner
Base Agent
Runner
Initiate/Get Session
Invoke agent
Memory
Update Session
Invoke Tool
Event
Event
Proprietary & ConfidentialGoogle Cloud
Sequential Agent
Parallel Agent
Loop Agent
LLM Agent
Sequential
Chains multiple specialized LLMs in a pipeline.
LLM
Input
Output
Intermediate
Result…
LLM
Parallel
Multiple LLMs concurrently and aggregate outputs
AggregatorInput Output
LLM
LLM
LLM
Hierarchical Routing
Router model to direct inputs to specialized LLMs
based on content or intent
RouterInput
LLM
LLM
LLM
LLM
Output
Analysis and
Classification
processing
ADK Agent - Workflow
Sequential Agent
Parallel Agent
Loop Agent
LLM Agent
Sequential
Chains multiple specialized LLMs in a pipeline.
LLM
Input
Output
Intermediate
Result…
LLM
Parallel
Multiple LLMs concurrently and aggregate outputs
AggregatorInput Output
LLM
LLM
LLM
Hierarchical Routing
Router model to direct inputs to specialized LLMs
based on content or intent
RouterInput
LLM
LLM
LLM
LLM
Output
Analysis and
Classification
processing
ADK Agent - Workflow
Proprietary & ConfidentialGoogle Cloud
ADK Agent - workflow
parallel_agent = ParallelAgent(
name="RefundChecks",
sub_agents=[refund_eligibility_checker, get_purchase_history],
)
sub_agents_list = [
parallel_agent,
offer_store_credit,
process_store_credit_response,
]
ADK Agent - workflow
parallel_agent = ParallelAgent(
name="RefundChecks",
sub_agents=[refund_eligibility_checker, get_purchase_history],
)
sub_agents_list = [
parallel_agent,
offer_store_credit,
process_store_credit_response,
]
Proprietary & ConfidentialGoogle Cloud
Call backs
Base Agent
Model
Tools
before agent call back
a\er agent call back
Before model callback
a\er model callback
Before tools callback
A\er tools callback
Observe & Debug:
Log detailed information at
critical steps for monitoring
and troubleshooting
Control:
Modify data flowing
through the agent or
bypass certain steps.
Implement Guardrails
Enforce safety rules,
validate inputs/outputs,
or prevent disallowed
operations
Manage State:
Read or dynamically
update the agent's
session state
Caching:
Temporary data store
Notification:
Tracing, or notify
human
Call backs
Base Agent
Model
Tools
before agent call back
a\er agent call back
Before model callback
a\er model callback
Before tools callback
A\er tools callback
Observe & Debug:
Log detailed information at
critical steps for monitoring
and troubleshooting
Control:
Modify data flowing
through the agent or
bypass certain steps.
Implement Guardrails
Enforce safety rules,
validate inputs/outputs,
or prevent disallowed
operations
Manage State:
Read or dynamically
update the agent's
session state
Caching:
Temporary data store
Notification:
Tracing, or notify
human
How ADK Works
https://google.github.io/adk-docs/runtime/#how-it-works-a-simplified-invocation
https://google.github.io/adk-docs/runtime/#how-it-works-a-simplified-invocation
Proprietary & ConfidentialGoogle Cloud
Agent2Agent (A2A) Protocol
Agent
Tool
Tool
Agent
Tool
Tool
Workflow Agent
Agent
Tool
Tool
Agent
Tool
Tool
Agent
Tool
Tool
Agent
Tool
Tool
Agent
Tool
Tool
Tool
Tool
?
?
?
?
?
?
Agent2Agent (A2A) Protocol
Agent
Tool
Tool
Agent
Tool
Tool
Workflow Agent
Agent
Tool
Tool
Agent
Tool
Tool
Agent
Tool
Tool
Agent
Tool
Tool
Agent
Tool
Tool
Tool
Tool
?
?
?
?
?
?
Proprietary & ConfidentialGoogle Cloud
A2A - Agent Card
Name: Agent ABC
Provider: Google (URL)
Capabilities:
I can do streaming
I can do Push Notifications
Skills:
● Skill A
○Description
○Input
○Output
● Skill B
○Description
○Input
○Output
Authentication:
Scheme: Bearer
Key: xxx
Preferred input/output: text
A2A - Agent Card
Name: Agent ABC
Provider: Google (URL)
Capabilities:
I can do streaming
I can do Push Notifications
Skills:
● Skill A
○Description
○Input
○Output
● Skill B
○Description
○Input
○Output
Authentication:
Scheme: Bearer
Key: xxx
Preferred input/output: text
Proprietary & ConfidentialGoogle Cloud
A2A Agent Discovery
A2A Server
Workflow Agent
Model
Instruction
get Agent Cards
Get Task
Send Task
Agent A
Agent B
Agent C
Agent D
Tasks
+ AgentCards
Server Initialization:
All Agent Cards are
initialized and made
available by the server.
1
Client Discovery:
Discovers all available agents by retrieving
and storing their Agent Cards in its list of
available sub-agents.
2
Agent
Selection:
Based on
instructions
processed by a
Large Language
Model (LLM), the
client determines
which agent to
interact with.
.
3
Subagents
A2A Agent Discovery
A2A Server
Workflow Agent
Model
Instruction
get Agent Cards
Get Task
Send Task
Agent A
Agent B
Agent C
Agent D
Tasks
+ AgentCards
Server Initialization:
All Agent Cards are
initialized and made
available by the server.
1
Client Discovery:
Discovers all available agents by retrieving
and storing their Agent Cards in its list of
available sub-agents.
2
Agent
Selection:
Based on
instructions
processed by a
Large Language
Model (LLM), the
client determines
which agent to
interact with.
.
3
Subagents
Proprietary & ConfidentialGoogle Cloud
A2A Task Execution Flow
A2A Server
Workflow Agent
Model
Instruction
Sub Agents
get Agent Cards
Send Task
Agent A
Agent B
Agent C
Agent D
Task
Registry
T T T
Task Initiation(Client):
Serializes(JSON) and
sends a task request.
1
Task Validation:
Deserializes and
validates the incoming
task request
2
Task Store:
Stores the task's ID, status,
data, and context for
tracking and persistence.
3
Task Execution:
Loads and initializes
the appropriate
agent, then executes
it within the created
context, performing
the actions defined in
the task request.
4
Get Task
Response:
serializes and sends a response
containing the task's status and
results back to the client.
5
A2A Task Execution Flow
A2A Server
Workflow Agent
Model
Instruction
Sub Agents
get Agent Cards
Send Task
Agent A
Agent B
Agent C
Agent D
Task
Registry
T T T
Task Initiation(Client):
Serializes(JSON) and
sends a task request.
1
Task Validation:
Deserializes and
validates the incoming
task request
2
Task Store:
Stores the task's ID, status,
data, and context for
tracking and persistence.
3
Task Execution:
Loads and initializes
the appropriate
agent, then executes
it within the created
context, performing
the actions defined in
the task request.
4
Get Task
Response:
serializes and sends a response
containing the task's status and
results back to the client.
5
Proprietary & ConfidentialGoogle Cloud
Model Context Protocol (MCP)
Agent
(MCP Client)
MCP Server
List Tools
Call Tools
Model
Instruction
Tools
doABC
fetchDEF
genGHI
sendJKL
Model Context Protocol (MCP)
Agent
(MCP Client)
MCP Server
List Tools
Call Tools
Model
Instruction
Tools
doABC
fetchDEF
genGHI
sendJKL
Proprietary & ConfidentialGoogle Cloud
Agent
(MCP Client)
MCP Server
Stdio
TOOL TOOL
TOOL TOOL
Model Context Protocol (MCP) Deployment
Agent
(MCP Client)
MCP Server
Stdio
TOOL TOOL
TOOL TOOL
Agent
(MCP Client)
sse
Agent
(MCP Client)
sse
MCP Server
sse
TOOL TOOL
TOOL TOOL
Agent
(MCP Client)
Agent
(MCP Client)
MCP Server
Stdio
TOOL TOOL
TOOL TOOL
Model Context Protocol (MCP) Deployment
Agent
(MCP Client)
MCP Server
Stdio
TOOL TOOL
TOOL TOOL
Agent
(MCP Client)
sse
Agent
(MCP Client)
sse
MCP Server
sse
TOOL TOOL
TOOL TOOL
Agent
(MCP Client)
Proprietary & ConfidentialGoogle Cloud
TEST IN
PRODUCTION
THEY SAID….
TEST IN
PRODUCTION
THEY SAID….
Proprietary & ConfidentialGoogle Cloud
Evaluation (Testing)
Test Cases
&^)**%%
&^)**%%
&^)**%%
&^)**%%
Expected
Result
Design
Evaluation
Development 2
This slide is an editable slide with
all your needs.
Deployment
Deploy
Evaluation (Test)
Design
Develop
Report
Agentic Development Life cycle
Evaluation (Testing)
Test Cases
&^)**%%
&^)**%%
&^)**%%
&^)**%%
Expected
Result
Design
Evaluation
Development 2
This slide is an editable slide with
all your needs.
Deployment
Deploy
Evaluation (Test)
Design
Develop
Report
Agentic Development Life cycle
Proprietary & ConfidentialGoogle Cloud
Deploying on Google Cloud
Artifact Registry
Cloud Run, Cloud Run Function, GKE
Platform
Dockerfile,
Google
Cloud Build
Container
Model
VertexAI, AI Studio
Self host (GKE, GCE)
Agent Engine
Vertexai
Libraries (ADK,MCP, A2A)
Python, LangGraph
Google GenAI, ADK
Your Code
Agent
Runner
Deploying on Google Cloud
Artifact Registry
Cloud Run, Cloud Run Function, GKE
Platform
Dockerfile,
Cloud Build
Container
Model
VertexAI, AI Studio
Self host (GKE, GCE)
Agent Engine
Vertexai
Libraries (ADK,MCP, A2A)
Python, LangGraph
Google GenAI, ADK
Your Code
Agent
Runner
Google Cloud Proprietary & Confidential
Accelerate your journey to production-ready agents
Reduce time-to-production from months to weeks.
Learn how to build production-ready Agents in GCP
With a real-time chat
interface, event streaming,
and auto-generated docs.
Prod-ready API
server
UI Playground for
experimentation
Ready-to-use
AI patterns
GCP Native
Observability
CI/CD and
Terraform
Offering a diverse set
of examples: Conversational
chain, custom RAG,
LangGraph agent.
With Cloud Tracing
and Logging. Includes
a pre-built Looker
dashboard for monitoring.
For quick deployment.
Including a set of unit,
integration and load tests.
Including chat curation, user
feedback collection,
multimodal input, and more.
?????? goo.gle/agent-starter-pack
Deployment & Operations Customization Observability
Evaluation Evaluation
Agent Starter Pack
A collection of prod-ready templates
Accelerate your journey to production-ready agents
Reduce time-to-production from months to weeks.
Learn how to build production-ready Agents in GCP
With a real-time chat
interface, event streaming,
and auto-generated docs.
Prod-ready API
server
UI Playground for
experimentation
Ready-to-use
AI patterns
GCP Native
Observability
CI/CD and
Terraform
Offering a diverse set
of examples: Conversational
chain, custom RAG,
LangGraph agent.
With Cloud Tracing
and Logging. Includes
a pre-built Looker
dashboard for monitoring.
For quick deployment.
Including a set of unit,
integration and load tests.
Including chat curation, user
feedback collection,
multimodal input, and more.
?????? goo.gle/agent-starter-pack
Deployment & Operations Customization Observability
Evaluation Evaluation
Agent Starter Pack
A collection of prod-ready templates
Proprietary & ConfidentialGoogle Cloud
●The customer's refund request starts at a
Cloud Run frontend.
●A Coordinator Agent manages the core
logic and agent orchestration.
●Specialized Subagents execute tasks and
call necessary internal tools.
●The AI model, such as Gemini, is hosted
and run on Vertex AI.
●Agents are deployed to the Google
Cloud platform via Cloud Run, Agent
Engine, or GKE.
●System is deployed within a specific
Google Cloud Region for low latency.
●Cloud Run services manage user
interaction and scale automatically to
handle load spikes.
●Vertex AI hosts the models, providing
high availability for inference requests.
●The Agent runtime can utilize GKE for
custom orchestration and scaling needs.
●Tools and databases within the system
are secured and managed in Google
Cloud.
MVP to PROD
●The customer's refund request starts at a
Cloud Run frontend.
●A Coordinator Agent manages the core
logic and agent orchestration.
●Specialized Subagents execute tasks and
call necessary internal tools.
●The AI model, such as Gemini, is hosted
and run on Vertex AI.
●Agents are deployed to the Google
Cloud platform via Cloud Run, Agent
Engine, or GKE.
●System is deployed within a specific
Google Cloud Region for low latency.
●Cloud Run services manage user
interaction and scale automatically to
handle load spikes.
●Vertex AI hosts the models, providing
high availability for inference requests.
●The Agent runtime can utilize GKE for
custom orchestration and scaling needs.
●Tools and databases within the system
are secured and managed in Google
Cloud.
MVP to PROD
Proprietary & ConfidentialGoogle Cloud
Takeaways: Building Autonomous Agentic Systems ??????
Agents vs Autonomous Agents → Agents act, autonomous agents adapt & decide.
Core Challenges → Planning the plan, memory wars, critique cadence, tool strategies,
debugging chaos.
Evaluation is Key → Don’t just measure if it works → measure how well it works (success
rate, coherence, cost, robustness).
Design Patterns Emerging → “Deep Agent” (LangChain), Agent Swarms.
Mindset Shift → Building agents is not about coding bots, it’s about architecting
mini-organisms.
Final Thought → Today’s agents may stumble, but tomorrow’s will collaborate like human
teams—your design choices will shape that future.
Takeaways: Building Autonomous Agentic Systems ??????
Agents vs Autonomous Agents → Agents act, autonomous agents adapt & decide.
Core Challenges → Planning the plan, memory wars, critique cadence, tool strategies,
debugging chaos.
Evaluation is Key → Don’t just measure if it works → measure how well it works (success
rate, coherence, cost, robustness).
Design Patterns Emerging → “Deep Agent” (LangChain), Agent Swarms.
Mindset Shift → Building agents is not about coding bots, it’s about architecting
mini-organisms.
Final Thought → Today’s agents may stumble, but tomorrow’s will collaborate like human
teams—your design choices will shape that future.
Proprietary & ConfidentialGoogle Cloud
Happy Building!
“Civilization advances by extending the
number of operations we can perform
without thinking about them”
- Alfred North Whitehead
Mathematician and Philosopher
Happy Building!
“Civilization advances by extending the
number of operations we can perform
without thinking about them”
- Alfred North Whitehead
Mathematician and Philosopher
Download
Download Slideshow Get the original presentation fileQuick Actions
Statistics
- Views 15
- Slides 36
- Age 7 days
Related Slideshows
11
8-top-ai-courses-for-customer-support-representatives-in-2025.pptx
0 views
10
7-essential-ai-courses-for-call-center-supervisors-in-2025.pptx
0 views
13
25-essential-ai-courses-for-user-support-specialists-in-2025.pptx
0 views
11
8-essential-ai-courses-for-insurance-customer-service-representatives-in-2025.pptx
0 views
21
Know for Certain
0 views
17
PPT OPD LES 3ertt4t4tqqqe23e3e3rq2qq232.pptx
0 views