LangChain Skills Progressive Disclosure: Dynamic SKILL.md Loading Bumps Claude Code from 17% to 92% on LangSmith Tasks Without Token Bloat

LangChain Blog·medium signal

LangChain's skills architecture uses progressive disclosure — agents see skill names and one-line descriptions at start, then load full instructions only when relevant to the current task. This solves the historical problem where too many tools degrades agent performance. Result: Claude Code accuracy on LangSmith tasks jumps from 17% to 92%, and LangChain tasks from 29% to 95%, with no increase in baseline token usage.