Progressive disclosure is the practice of keeping an AI agent's baseline context minimal and loading instructions, tools, and knowledge only at the moment a task actually calls for them.
The term comes from interface design, where a system reveals advanced options only as a user needs them; in agentic collaboration the surface being managed is the context window rather than a screen. 📝Anthropic named the shift explicitly in the 📝Context Engineering Rules for Claude 5 Generation, reporting that more than eighty percent of 📝Claude Code's system prompt could be removed for advanced models with no measured performance loss — a figure Anthropic reports on its own product and that has not been independently replicated.
In practice the pattern operates at several layers. A 📝Skill (Claude Code) exposes only its name and description at session start, injecting its full body when invoked. 📝Progressive Discovery (MCP) applies the same logic to tool schemas, holding them behind a search tool instead of loading every connected server's definitions upfront. Instruction files and knowledge bases follow it by naming where context lives rather than carrying it, deferring retrieval to the point of need.
The cost is an extra inference hop and a quieter failure mode: unloaded context fails silently, because a model that never triggers the load behaves exactly like one working without the material. That moves the engineering burden onto the index — descriptions, triggers, and naming — which is where progressive disclosure succeeds or fails. It is distinct from 📝Compaction (Claude Code), which reclaims context after it has been spent; progressive disclosure avoids spending it. The operative rule for 📝Context Engineering is to design the catalog with more care than the payload.
Every layer of my stack runs on this — a global instruction file that points instead of carries, 📝MythOS scopes that load as a conversation narrows. The bet is always the trigger: context that never gets pulled is indistinguishable from context that was never written.