[DevFest Pisa 2026] Build your own executive assistant with ADK and Gemini

Talk presented at DevFest Pisa 2026

Transcript

1. 18/Apr/2026 petruzalek@google.com Daniela Petruzalek Developer Relations Engineer Google Cloud Build your own executive assistant with ADK and Gemini
2. A little about myself… DevRel at Google UK Originally from Brazil Previously Backend / Data Engineer Currently obsessed with AI Love Games, Anime and Cats =^_^=
3. 1 Types of Agents 2 Agent Development Kit (ADK) 3 Components 4 Build your own! Contents
4. 4 Types of Agents An unofficial classification
8. 8 The Agent Development Kit And a few other tools…
9. Agent Development Kit (ADK) An open source framework for development and deployment of AI agents. Optimised for Gemini and Google products, but also model-agnostic and environment-agnostic google.github.io/adk-docs
10. Gemini CLI Open source coding agent that uses the Gemini model family. Can be used standalone or in combination with an IDE geminicli.com
11. Developer Knowledge API Provides programmatic access to Google's public developer documentation, enabling you to integrate this knowledge base into your own applications and workflows. Supports MCP. developers.google.com/knowledge/api
12. CodeWiki Chat grounded in open source. Powered by Gemini codewiki.google
13. 13 OpenPaws The Executive Assistant
14. Architecture v1: Stateless root agent Dev UI (runner) user Gemini 3 Flash Preview calendar agent todo agent research agent
15. MARCH 2025 # root_agent is the entry point for an ADK agent root_agent = Agent( model="gemini-3-flash-preview", name="sharon", instruction=f""" You are Sharon, a professional Executive Assistant. Use the following tools and agents to perform specialised tasks: - Calendar: For all scheduling and meeting management. - Research: For deep-dive web investigations and synthesis. - Todo: For tracking tasks and actionable lists. """, tools=[ AgentTool(calendar_agent), AgentTool(todo_agent) ], sub_agents=[research_agent] ) agent.py
16. MARCH 2025 # 1. The Calendar Specialist (Agent Definition) calendar_agent = Agent( name="calendar_specialist", model="gemini-3-flash-preview", instruction=""" You manage the user's schedule. Always query the calendar before adding events to avoid conflicts. Confirm details (time, attendees) before finalizing. """, tools=[ list_events, # Queries Google Calendar API create_event, # Adds new events update_event # Modifies existing events ] ) calendar.py
17. MARCH 2025 calendar.py # 2. Tool Implementation (Google Calendar Integration) import google.auth from googleapiclient.discovery import build async def list_events(time_start: str, time_end: str) -> dict: """Lists calendar events within a time range.""" credentials, _ = google.auth.default( scopes=["https://www.googleapis.com/auth/calendar"] ) service = build("calendar", "v3", credentials=credentials) events_result = service.events().list( calendarId='primary', timeMin=time_start, timeMax=time_end ).execute() return events_result.get('items', [])
18. 18 Agentic Memories
19. Short Term Memory Uses session state. Lost after each session unless it is user scoped (magic key “user:”) or app scoped (“app:”). Persistence depends on session service Long Term Memory Remembers facts about the conversation itself. Saved in a database or specialist service like Vertex AI Memory Bank
20. # Context Injection via Callbacks async def setup_agent_context(callback_context, **kwargs): """ Runs once per turn via before_agent_callback. """ now = datetime.now(ZoneInfo(SHARON_TIMEZONE)) callback_context.state["current_time"] = now.strftime("%A, %Y-%m-%d %I:%M %p") callback_context.state["timezone"] = SHARON_TIMEZONE # Register the callback in the Root Agent root_agent = Agent( name="sharon", # ... before_agent_callback=[setup_agent_context] )
21. Vertex AI Memory Bank Automatically retrieves relevant memories at the beginning of the turn Automatically persist memories at the end of the turn (or session) Allows ad hoc search docs.cloud.google.com/agent-builder/agent-engine/memory-bank
22. from google.adk.tools.preload_memory_tool import preload_memory_tool # 1. Persistence: index every interaction at the end of the turn async def auto_save_memory(callback_context, **kwargs): memory_service = VertexAiMemoryBankService( project=PROJECT_ID, location=MEMORY_LOCATION, agent_engine_id=AGENT_ENGINE_ID ) await memory_service.add_session_to_memory(callback_context.session) # 2. Agent Configuration root_agent = Agent( name="sharon", tools=[preload_memory_tool], # 3. Implicit Retrieval after_agent_callback=[auto_save_memory] )
23. Architecture v2: Stateful root agent Dev UI (runner) user Gemini 3 Flash Preview calendar agent todo agent research agent Vertex AI Memory Bank
24. A side note on memories… github.com/MemPalace/mempalace
25. 25 Build your own assistant
26. goo.gle/daniela-devfest-pisa-2026
27. Thank you! Questions?