Build an ADK agent that uses four different skill patterns: inline skills defined in Python, file-based skills loaded from directories, external skills pulled from community repos, and a meta skill that generates new skills on demand. Each pattern demonstrates a different tradeoff between simplicity and reusability.
- How ADK Skills solve context window bloat through progressive disclosure (L1/L2/L3)
- Four skill patterns: inline, file-based, external, and meta (skill-creator)
- How to wire multiple skills through a single
SkillToolsetwith three auto-generated tools - When to use each pattern based on reusability vs simplicity tradeoffs
- Python 3.11+
- Google ADK (
pip install google-adk) - A Google API key (get one here)
# Clone the repo
git clone https://github.com/GoogleCloudPlatform/adk-samples.git
cd adk-samples/python/agents/agent-skills-tutorial
# Set up environment
python3 -m venv .venv && source .venv/bin/activate
pip install -e .
# Configure API key
cp .env.example app/.env
# Edit app/.env with your GOOGLE_API_KEY
# Run with ADK Web UI
adk web
Test these queries to see the agent in action:
| # | Query | What It Demonstrates |
|---|---|---|
| 1 | "I have a blog post titled 'Getting Started with Kubernetes'. Can you review it for SEO?" | Inline skill (seo-checklist) loaded on demand |
| 2 | "Help me write a short introduction for a blog about Python async programming. Make it SEO-friendly." | Multi-skill: blog-writer + seo-checklist loaded in parallel |
| 3 | "Can you use your video-editing skill to create a thumbnail?" | Edge case: agent handles nonexistent skill gracefully |
| 4 | "OK, then use your content research skill to help me research async Python" | External skill (content-research-writer) with resource loading |
| 5 | "I need a new skill for reviewing Python code for security vulnerabilities. Can you create a SKILL.md?" | Meta skill: skill-creator generates a new skill on demand |
The agent uses a single SkillToolset that auto-generates three tools: list_skills (L1 — returns names and descriptions), load_skill (L2 — loads full instructions), and load_skill_resource (L3 — fetches reference files). This progressive disclosure pattern means the agent sees only ~200 tokens of skill metadata per LLM call, loading full instructions only when needed.
agent-skills-tutorial/
├── app/
│ ├── __init__.py
│ ├── agent.py # Root agent with SkillToolset wiring 4 skill patterns
│ └── skills/ # File-based skills (blog-writer, content-research-writer)
├── assets/ # Screenshots, diagrams, and demo GIF
├── .env.example # Environment variable template
├── pyproject.toml # Project configuration and dependencies
└── README.md