WORKSHOP
Building ADK Agents with Skills and Tools
In this 90-minute workshop, participants will build, test, and deploy a Marathon Planner Agent using Google’s Agent Development Kit (ADK). Attendees will replace static prompts with dynamic instructions (PromptBuilder), load domain expertise on demand via Skills (SkillToolset), connect external spatial capabilities via Model Context Protocol (MCP), and deploy the agent to Cloud Run.
- Format: Instructor-led live coding + Hands-on lab
- Prerequisites: Cloud Shell, Google Cloud Project with billing, Python basics
- Agenda
- Architecture & Setup – Overview of ADK, MCP, and Skills. Clone repo, run uv sync, configure Maps API key and Cloud Shell.
- Modular Prompts & Base Agent – Initialize Agent (gemini-3-flash-preview). Test standard LLM response vs. structured PromptBuilder instructions.
- Skills & MCP Integration – Wire directory-based Skills(gis-spatial-engineering, race-director) via SkillToolset. Register Google Maps MCP tools using MapsApiRegistry
- Local Debugging (adk web) – Run uv run adk web. Observe skill load events, MCP tool calls, and trace spans in the browser UI.
- Agent Engine deployment – Deploy to Cloud Run via uv run adk deploy agent_engine. Test the live endpoint with main.py.