Revamping Skills in Deep Agents

LangChain has revamped skills support in Deep Agents, adding tool binding, runtime pinning and mid-thread reloading. The changes are available in the latest deepagents release.

Skills follow an open standard supported by dozens of agent products and work with any model. A skill is a directory containing a SKILL.md file: YAML frontmatter with a name and description, followed by instructions, plus optional files under scripts/, references/ and assets/. Only each skill's name and description sit in context at startup; the agent reads full instructions only when a task matches, then loads supporting files on demand, a pattern the post calls progressive disclosure. LangChain's example GTM agent carries more than 50 skills covering sales work such as meeting-prep and call-transcripts.

Tools can now be bound to a skill by listing them in the skill's frontmatter under metadata.include_tools and passing them to SkillsMiddleware rather than the agent. A bound tool is not added to context until the agent reads its skill, and calling it earlier fails as an unknown tool. The tools arrive in a new system message, so the cached prefix above it is unchanged. On Anthropic and OpenAI models that accept tools mid-conversation, bound tools are added right after the skill is read and the prompt cache stays intact; on other models they are appended to the request as before. A skill can instead list a label that a function passed to SkillsMiddleware turns into tools, allowing a whole group such as every tool on an MCP server to be disclosed under one name or gated on runtime permissions.

Pinning lets an app pass pinned_skills, so the skill's instructions enter the conversation before the next model call. Deep Agents does not parse messages, so the app chooses the syntax, such as /meeting-prep. The pinned skill, and any tools bound to it, are added once as a tagged message so earlier messages do not change and the prompt cache stays valid.

Skills are loaded at the start of each thread and kept in agent state. Setting skills_metadata to None when invoking the agent invalidates the list so the next run rescans every source, letting long-running threads pick up added, edited or deleted skills. A reload that finds new skills changes the system prompt and invalidates the prompt cache; provider caches typically expire within minutes to an hour of inactivity.

Why it matters: teams with large shared skill registries can keep context small while still guaranteeing the right instructions and tools are loaded, and long-running agents stay current without starting a new thread.