How Progressive Disclosure Cuts AI Agent Token Bloat by 90% and Enables Self‑Generated Skills

Google Developers Blog published a deep‑dive guide on the Agent Development Kit (ADK) and its SkillToolset capabilities, focusing on the problem of knowledge explosion in AI agents. When many domain‑specific rules (SEO, code review, API specs, compliance, data handling) are packed into a single system prompt, token usage can sky‑rocket, wasting up to 80% of the context and degrading response accuracy.

Progressive Disclosure Architecture

The solution is a three‑layer progressive disclosure design that loads knowledge only when needed, reducing base‑context consumption by about 90%.

L1 Metadata Layer (~100 tokens per skill) contains only the skill name and a short description, acting like a menu.

L2 Instruction Layer (up to 5,000 tokens) holds the full step‑by‑step instructions; it is fetched via load_skill after L1 indicates relevance.

L3 Resource Layer holds external reference files (style guides, API docs, templates) loaded on demand with load_skill_resource.

For a 10‑skill agent, traditional prompts consume ~10,000 tokens per call, whereas progressive disclosure keeps the base prompt around 1,000 tokens and only adds L2/L3 tokens for the actually used skills, saving roughly 70% of tokens in typical scenarios.

Four Skill Construction Modes

Google describes four increasingly complex ways to define skills:

Inline Skill : Defined directly in code as a Python object with name, description, and instruction strings. Ideal for small, stable rule sets (e.g., an SEO checklist).

File‑based Skill : Stored in a directory with a SKILL.md (YAML front‑matter for L1 and markdown instructions for L2) and optional subfolders for L3 resources, enabling version control and reuse.

External Skill : Same structure as file‑based but sourced from community repositories (e.g., awesome-claude-skills), loaded via load_skill_from_dir.

Skill Factory (Meta Skill) : A special skill that generates new SKILL.md files at runtime. It reads the agentskills.io specification and example skills, then creates a compliant skill definition based on user requests, allowing agents to extend themselves without human intervention.

Meta‑skill implementation relies on a resources field that embeds the full agentskills.io spec and an example skill, guiding the agent to produce correctly formatted output.

End‑to‑End Workflow

A user asks the agent to create a new skill for scanning Python code for security vulnerabilities. The agent lists existing skills, finds none matching, activates the meta‑skill, reads the spec and example, and generates a new SKILL.md with proper naming, structured instructions, and resource references. The generated skill can then be saved locally and reused across sessions and platforms (Gemini CLI, Claude Code, Cursor, etc.).

Google recommends a final human review of auto‑generated skills and using ADK’s built‑in evaluation tools to verify correctness.

The architecture encourages starting with inline skills, moving to file‑based when external resources or cross‑agent reuse are needed, and only adopting meta‑skills for truly novel scenarios. This progressive approach mitigates token bloat, preserves response quality, and opens the door to self‑extending AI agents.

For developers, the workflow reduces to creating a SkillToolset instance, adding the desired skills (inline SEO, file‑based blog writer, external research skill, meta‑skill), and passing it to the agent constructor. The agent’s system prompt simply states its role (e.g., a blog‑writing assistant) and instructs it to load and explain the relevant skills during execution.

Overall, progressive disclosure solves the token‑bloat issue, while the four skill‑building modes and the meta‑skill enable agents to scale from static instruction execution to dynamic self‑extension.