Progressive Disclosure
/prəˈɡrɛsɪv dɪsˈkloʊʒər/
Also known as: lazy context loading, on-demand context, progressive context loading
What Is Progressive Disclosure?
Progressive disclosure is a context management pattern used by AI agent platforms to enable a single agent to access thousands of skills without overloading its context window. Instead of loading all skill instructions upfront, only lightweight metadata (name and description) is stored in the agent’s active memory.
The loading happens in stages:
- Always loaded — Skill names and descriptions (metadata)
- On trigger — The
skill.mdprocess instructions load when the user activates the skill - On demand — Reference files, code scripts, and assets load only when the skill’s process explicitly calls for them
Why It Matters
Modern AI agents have finite context windows. Loading thousands of full skill instructions simultaneously would be impossible. Progressive disclosure solves this by treating skills like programs on a computer — they’re installed and discoverable, but only run (loaded into memory) when needed.
This architecture enables a fundamentally different agent experience: one general-purpose agent that can do thousands of specific things, rather than many isolated agents each doing one thing.
In Practice
When you give an AI agent access to a new skill, the agent only stores a short description like “Infographic Generator — creates branded infographics for LinkedIn and newsletters.” When a user says “make me an infographic,” the agent matches this intent to the metadata, loads the full skill.md, and begins execution — loading additional reference files only as each step requires them.
Related Reading
- Agent Skills — The capabilities being progressively disclosed
- Skill Engineering — Designing skills for effective progressive loading
Mentioned In
Ben (Ben AI)
Only the metadata — the description and the name — are stored in the agent memory. And then only when the skill is triggered, the skill MD will be loaded into the context window.