Progressive disclosure

Progressive disclosure is the core idea behind ContextPie: an agent never loads everything up front. It discovers skills, loads a body only when relevant, and fetches a supporting file only when the instructions ask for it. Each step reveals a little more, so the agent's context window stays small.

The flow

The three MCP tools form three stages — search, load, file_load:

1. search      find_skills(query)        →  descriptors (name, description, score)
                  │  agent picks the best match
                  ▼
2. load        get_skill(id)             →  SKILL.md body + file manifest
                  │  body references a file
                  ▼
3. file_load   get_skill_file(id, path)  →  one file's contents

The agent only advances to the next stage if the current one justifies it. A search that returns no good match costs nothing further. A skill whose body the agent reads but never needs a file from stops at stage 2.

Why it keeps context small and cheap

Loading an entire skill library into a prompt is expensive and noisy — most of it is irrelevant to any single task. Progressive disclosure inverts that:

The agent reads only what it commits to. That means fewer tokens, less distraction, and sharper behavior.

It also produces your telemetry

Because each stage is a distinct tool call, ContextPie can see exactly where agents stop. "The query is the telemetry" — see Analytics for how search, load, and file_load events become insight. Discovery quality starts with good trigger phrases.