LangChain Skills: 29% to 95% Task Success

LangChain just released skills for building agents with LangChain, LangGraph, and Deep Agents.

Claude Code goes from 29% to 95% task success on LangChain projects, and from 17% to 92% on LangSmith. The benchmarks are public and open source.

As a reminder, a skill is a reusable block of expert knowledge that tells an agent how to use a specific tool or framework correctly. Skills are structured instruction sets living in a file called skills.md and loaded dynamically as needed. An agent first sees only the name and description of a skill, then pulls the full content on demand. This really matters because flooding an agent with context kills its performance.

v1 ships with 11 skills across three categories.

• LangChain core for standard agent patterns • LangGraph for human-in-the-loop workflows and durable execution • DeepAgents for multi-agent systems with middleware and FileSystem tools.

LangSmith skills are included in the same release.

If your coding agent supports the standard, you can plug it into your existing LangChain project as is.