A Developer’s Guide to Building ADK Agents with Skills
Google Developers Blog guide on adding agent-skills to agents built with the adk (Agent Development Kit, google‘s agent framework). Audience: developers building production agents that span multiple complex task domains.
Core argument
Scaling agents means modular knowledge loading, not stuffing everything into one monolithic system prompt. Skills are “modular, reusable instruction sets that agents activate contextually” — each carrying metadata (name, description), a detailed instruction body, and optional reference resources.
What’s new here for this wiki
- An open, cross-vendor standard. Skills here follow the agentskills-spec (agentskills.io) — explicitly universal, working across “Gemini CLI, Claude Code, Cursor, and 40+ other products.” This is the agent-skills form factor crystallizing from per-tool conventions into a shared specification adk-agents-with-skills.
- Progressive disclosure, quantified. The three-level loading model (L1 metadata ~100 tokens at startup → L2 body <5,000 tokens on activation → L3 resources on demand) reportedly cuts baseline context use by ~90% vs. a monolithic prompt — the mechanism detailed on agentskills-spec.
Patterns demonstrated
Four progressive patterns: inline skills (simple tasks) → file-based skills (when reference material is needed) → external imports → self-generating meta-skills (an agent writing its own skills). Recommendation: review AI-generated skills before deployment, “treating them like code reviews.”
Why it routed here
It is the first source to document a second major vendor (Google) standardizing on the same skills format already seen across claude-financial-services, agentic-seo-skill, and gbrain — see agent-skills. (Caveat: vendor blog; the ~90% / token figures are its own claims, not independently measured.)