1 of 30

Mastering Multi-Agent Orchestration using ADK

Cloud - Edmonton

Suvaditya Mukherjee

ML GDE | Magnopus | USC

@halcyonrayes

2 of 30

Cloud - Edmonton

$whoami

Google Developer Expert - Machine Learning & Cloud

ML Engineer @ Magnopus

Master’s - Computer Science (Artificial Intelligence) @ USC

3 of 30

Agents

Chapter One

The proliferation of agents and why you need them

4 of 30

Generative AI vs. Agentic AI: Moving from static text generation to dynamic, goal-oriented execution.

The Core Loop: Perceive (Observation) -> Think (Reasoning) -> Act (Execution).

Autonomy: The ability to iteratively plan and execute sub-tasks without continuous human prompting.

Moving beyond Chat

5 of 30

Interactivity

Systems need to interact with external environments (databases, APIs, local file systems).

Complexity

Real-world workflows require multi-step reasoning that exceeds single-prompt capabilities.

Resilience

Agents can self-correct, catch exceptions, and retry execution paths when initial attempts fail.

Why Agents?

6 of 30

Chapter Two

Are they any different?

LLMs v/s Agents

7 of 30

Large Language Models

Stateless prediction engines mapping inputs to probability distributions.

Agents

Stateful systems that use the LLM as a reasoning engine to drive external actions.

Engine v/s Vehicle

8 of 30

Architectural Differences

LLMs generally are standalone, whereas Agents will have other non-LLM tools or software components attached to it.

9 of 30

Chapter Three

One for all, all for one

Single-Agent Systems

10 of 30

Monolithic Agents

A single agent equipped with a massive system prompt and access to all available tools.

Use Case: Excellent for narrow, well-defined tasks (e.g., a simple code generation assistant).

Implementation: Often relies heavily on the ReAct (Reason + Act) prompting framework.

11 of 30

Context Window Pollution: Keeping track of tool outputs, past reasoning, and system instructions degrades performance over time.

Persona Conflict: Forcing one LLM to act as a planner, executor, and critic simultaneously leads to poor task execution.

Error Cascading: A single hallucination early in the loop can derail the entire execution pipeline.

Context Collapse

12 of 30

Chapter Four

Why have one when you can have multiple?

Multi-Agent Systems

13 of 30

Decomposing complex workflows into specialized, narrow-scope agents.

  • Dedicated System Prompts: Each agent has a single responsibility (e.g., a "Researcher" agent vs. a "Synthesizer" agent).
  • Modularity: Easier to test, debug, and upgrade individual components of the system.

Divide and Conquer

14 of 30

Each agent is executed in a predetermined order. This is the simplest form of a multi-agent pattern.

Multi-Agent Patterns

Sequential

15 of 30

Multi-Agent Patterns

Multiple agents can run parallel to each other. They can be independent to each other.

Parallel

16 of 30

Multi-Agent Patterns

Agents are triggered at intermediate steps based on results of previous steps. Can also trigger multiple agents.

Decision

17 of 30

Multi-Agent Patterns

Agents run a loop until a certain condition is met, at which point it exits the loop.

Loop

18 of 30

The Agentic Stack

19 of 30

Google ADK

Chapter Five

A toolset for 0-to-100 builders

20 of 30

Edges

The routing logic that determines the flow of execution.

Nodes

The individual agents or functional blocks in your graph.

State

The shared memory object passed between nodes.

Agents Galore

21 of 30

Multi-Agent Systems as State Machines

Each node receives the current state, performs its task, mutates the state, and passes it along.

Predictability as a feature

Ensures predictable execution paths and simplifies debugging complex Python workflows.

State-driven Orchestration

22 of 30

5 Core ADK Skills

01

Scaffolding

Initializes projects via agent-starter-pack with 50+ validated templates.

02

Dev Guide

Development lifecycle, coding guidelines, and operational rules.

03

Cheatsheet

API quick reference: patterns, tool definitions, and state management.

04

Evaluation

Testing, benchmarking, and quality validation using adk eval.

05

Deployment

Packaging, production rollout, and observability on Agent Engine.

23 of 30

Chapter Six

Giving your agents access to the world

ADK Tools

24 of 30

Familiar API surfaces

Binding Python functions (e.g., executing a PyTorch script, querying a local database, or running shell commands) to the agent.

Hands for the Agent

Tools are the interfaces between the LLMs’ reasoning and external execution.

Easy to integrate across apps

ADK handles the JSON schema generation and argument parsing automatically.

Connecting Agents to Reality

25 of 30

Source: ADK Documentation

Manage issues, search pages, and update team content

Connect with BigQuery to retrieve data and perform analysis

Atlassian

BigQuery

Store and retrieve information using semantic vector search

ChromaDB

Access models, datasets, research papers, and AI tools

Generate speech, localize voices, and create audio content

Hugging Face

Cartesia

Manage payments, send invoices, and handle subscriptions

PayPal

And many more!

26 of 30

Chapter Seven

Create to understand

Build with ADK

27 of 30

Chapter Eight

Where to go from here

Advanced Patterns

28 of 30

Human-in-the-Loop (HITL)

Designing graphs that pause execution before taking irreversible actions (e.g., merging code to main, executing a transaction).

The system waits for explicit human validation or modification of the state before proceeding via a defined edge.

29 of 30

Parallelization and Optimization

Running independent agent branches concurrently (e.g., dispatching three different "Research" agents simultaneously).

Merging the parallel states back together via a "Reducer" node, optimizing latency and token usage in heavy workflows.

30 of 30

Thank You!

Merci beaucoup!

Cloud - Edmonton

Suvaditya Mukherjee

ML GDE | Magnopus | USC

@halcyonrayes