Skip to content

Skills for ADK Agents: Optimizing Context with Progressive Disclosure

Learn how to build efficient AI agents using ADK SkillToolset and progressive disclosure instead of massive system prompts. Discover how to reduce token consumption while optimizing agent context through layered knowledge architecture.

Type
Workshop
Category
AI
Level
Intermediate
Duration
45 mins
Language
English

Events

Name Organizer Date Location Attendees Links
Build with AI Manila 2026: Beyond the Prompt GDG Manila 2026-04-11 Accenture People Hub, Uptown Mall, Taguig, Philippines 80 📊 Slide Deck

Resources

AI - Agent Development Kit (ADK) AI - Agents AI - Gemini Google Cloud Build with AI
QR Code for this session
QR Code

Abstract

As AI agents take on more tasks, developers often rely on massive, token-heavy system prompts to provide domain knowledge. This session presents a highly efficient alternative using the ADK SkillToolset. By breaking knowledge down into L1 metadata, L2 instructions, and L3 resources, we can implement progressive disclosure to drastically reduce token consumption. We will walk through building inline, file-based, external, and meta skills, culminating in a hands-on codelab where attendees build a fully optimized agent that pulls in expertise exactly when a task requires it.

Outline

  • The Problem: Monolithic Prompts

    • What AI agents are and how they work (model + tools)
    • Prompt engineering vs. context engineering: why the distinction matters
    • The U-shaped attention curve and the "lost in the middle" problem

  • The Solution: Progressive Disclosure

    • Static context (traditional) vs. skill approach (progressive disclosure)
    • Three-tier architecture: L1 Metadata, L2 Instructions, L3 Resources
    • Structure of a SKILL.md file: YAML frontmatter + Markdown body
    • Token impact (approximate; varies by model and skill content): 20 skills = ~20,000 tokens (monolithic) vs. ~1,000 tokens (L1 only)

  • Implementation: Skill Patterns with SkillToolset

    • Agent Development Kit (ADK): build fast, multimodal, open, scalable
    • Four skill patterns: Inline, File-Based, External, Meta
    • SkillToolset wiring: auto-generates list_skills, load_skill, load_skill_resource tools

  • Demo: Building Optimized Agents

    • Hands-on codelab: Build a Skill-Powered ADK Agent
    • Requirements: Python 3.10+, Google AI API key, terminal and text editor

  • Closing: Building Your Own Optimized Agents

    • References and further reading
    • Community resources and next steps

Key Takeaways

  • TLDL (Too Long Didn't Listen): short context = better quality, lower cost, faster responses
  • Progressive disclosure loads knowledge in layers — L1 metadata always, L2 instructions on demand, L3 resources only when needed
  • SKILL.md structure: YAML frontmatter (name + description) for routing, Markdown body for instructions, references/ for deep resources
  • SkillToolset automatically generates three tools that implement progressive disclosure: list_skills, load_skill, load_skill_resource
  • Four skill patterns for different needs: Inline (quick), File-Based (production), External (community), Meta (self-extending)
  • Skills follow the agentskills.io spec — portable across any compatible agent or toolchain that supports the specification

References