Developer’s Guide to Building ADK Agents with Skills

This guide explains how to use ADK's SkillToolset to build agents that load domain expertise on demand via progressive disclosure.

• •SkillToolset splits knowledge into three levels: L1 metadata (~100 tokens), L2 instructions (<5,000 tokens), and L3 resources loaded only when needed, reducing baseline context by ~90%.
• •Pattern 1 (Inline Skill): A Python object with name, description, and instructions hardcoded directly in agent code, ideal for small stable rules.
• •Pattern 2 (File-based Skill): A directory with SKILL.md and optional references/ subdirectory, enabling reusable skills with on-demand reference loading.
• •Pattern 3 (External Skill): Same API as file-based, but the SKILL.md is sourced from community repositories like awesome-claude-skills or Google's official ADK skills.
• •Pattern 4 (Skill Factory): A meta skill that generates new SKILL.md files at runtime, allowing agents to self-extend their capabilities without human intervention.

This summary was automatically generated by AI based on the original article and may not be fully accurate.