Share
Embed
[AI Agents LDN] AI Agents from Scratch with Google ADK
Talk presented on November 19, 2026 at the meetup "AI Agents in Production at MongoDB" (https://luma.com/london-zufm)
Daniela Petruzalek
November 19, 2026
Transcript
26 slides
-
AI Agents from Scratch with Google ADK Daniela Petruzalek Developer
Relations Engineer • Google Cloud • 2026 -
Who am I? • Brazilian living in London since 2019
• Specialisation: backend (Python & Go), data engineering and agentic AI • Hobbies: anime, games and cosplay Photo Blog: danicat.dev LinkedIn: linkedin.com/in/petruzalek Twitter: x.com/danicat83 Daniela Petruzalek -
Agenda 01 Intro to ADK 02 Deploying Agents to GCP
03 Securing Agents -
AI Agents Software systems that use AI to perform tasks
and achieve objectives on behalf of their users. A model with access to tools (functions) that allow it to interact with the external world. -
User Agent Tools Model Should I buy GTA VI? user
prompt + system prompt + tools Tool call request: query_database("leaves") query_database("leaves") { "leave_balance": 0, "sick_days": 30} Tool call response { "leave_balance": 0, "sick_days": 30} Final response: "No, you have no leaves left and I don't think you can claim yet another sick day" Google Cloud 06 -
Agent Development Kit (ADK) Open-source agent development framework Supports Gemini
and models from other labs Available for JS/TS, Python, Java and Kotlin Official skills: uvx google-agents-cli setup adk.dev -
Coding Agent Setup Google Skills (incl. GCP products): github.com/google/skills Google
Developer Knowledge MCP: developers.google.com/knowledge/mcp Google Cloud Developer Plugin (MCP + Skills): - Antigravity: agy plugin install github.com/google/skills/plugins/cloud/google-cloud-developer - Claude: claude plugin marketplace add google/skills claude plugin install google-cloud-developer@google-plugins Google Cloud 08 -
Anatomy of Your First Agent council/agent.py Model from google.adk.agents.llm_agent import
Agent root_agent = Agent( model='gemini-4-hydrogen', Identity name='root_agent', description='A helpful assistant for user questions.', instruction='Answer user questions to the best of your knowledge', ) Instruction -
The Council of Ancients Building our own personal board of
advisors Facilitator Plato Frames & synthesizes Virtue & civic good Descartes Machiavelli Method & clarity Incentives & realism -
Defining the Council in ADK counselors/plato.py plato = Agent( name="plato",
model=MODEL_ID, mode="single_turn", static_instruction=PLATO_INSTRUCTION, generate_content_config=GENERATE_CONFIG, ) agent.py facilitator = Agent( 3 Counselors mode="single_turn" Facilitator name="facilitator", model=MODEL_ID, static_instruction=FACILITATOR_INSTRUCTION, sub_agents=[plato, descartes, machiavelli], ) Auto-delegation tools -
Council Architecture App + ModelArmorPlugin • Input & Output Screening
mode="single_turn" Plato User (Prompt / SSE) Facilitator (Coordinator) Descartes Council Verdict (Facilitator) Machiavelli PARALLEL • OPENINGS & REBUTTALS Streams native ADK events over POST /run_sse -
Deployment Targets on Google Cloud Serverless container runtime; Google Cloud
Run agents run as a container responding over HTTP Gemini Enterprise Agent Platform Unified platform to build, deploy, govern & optimize enterprise AI agents (Memory Bank, Sessions, Registry, Gateway) -
Fast Deploys with Cloud Run --no-build gcloud beta run deploy
gcloud beta run deploy council-of-ancients \ --source . --region us-central1 \ --no-build --base-image python313 \ --command python3 --args -m,uvicorn,main:app,--host=0.0.0.0,--port=8080 \ --no-allow-unauthenticated --session-affinity --min-instances=0 Instant --no-build Scale to Zero IAM Protected -
LIVE DEMO 3 Deploy with Cloud Run Fast live deploy
and the Council Chamber UI -
Model Armor Template deploy/model-armor-prompt-template.json MEDIUM_AND_ABOVE { "filterConfig": { "piAndJailbreakFilterSettings": {
"filterEnforcement": "ENABLED", "confidenceLevel": "MEDIUM_AND_ABOVE" }, "sdpSettings": { "basicConfig": {"filterEnforcement": "ENABLED"} }, "maliciousUriFilterSettings": {"filterEnforcement": "ENABLED"} }, "templateMetadata": {"enforcementType": "INSPECT_AND_BLOCK"} } Jailbreak Defense RAI, PII & URIs 4 RAI filters + SDP Inspect & Block Server-side policy -
One Plugin Protects Every Agent agents/council_of_ancients/agent.py app = App( name="council_of_ancients",
root_agent=facilitator, plugins=[ ModelArmorPlugin( config=ModelArmorConfig( prompt_template_name=config.model_armor_prompt_template, response_template_name=config.model_armor_response_template, block_on_screening_failure=True, ) ) ], ) Screens Every Agent Fails Closed -
LIVE DEMO 4 Guardrails in Action Watching Model Armor block
a live prompt injection -
Key Takeaways 1. Build 2. Deploy 3. Protect ADK multi-agent
Cloud Run Google Cloud Model teams & Agent Platform Armor