LangChain announced three updates to skills in Deep Agents on October 7, 2026: tools can be bound to skills, apps can pin skills before a model call, and long-running threads can reload skills. The changes address enterprise libraries that are growing to thousands of skills shared across teams and agents. Skills are folders of instructions, scripts and reference files; the open standard is supported by dozens of agent products and can be authored as a markdown file.

Deep Agents reveals each skill’s name and description at startup, reads its SKILL.md when a task matches, then loads supporting files as needed. This progressive disclosure keeps the agent’s context limited to information required for the task. A skill’s SKILL.md contains YAML frontmatter with a name and description, and may be accompanied by scripts/, references/ and assets/. LangChain’s example go-to-market agent has more than 50 skills, including meeting-prep, call-transcripts and competitive-intel-card.

With tool binding, a skill can list tools in metadata.include_tools, and those tools become available only after the agent reads that skill. Calling a bound tool before then fails as an unknown tool. In the example, reading call-transcripts unlocks search_calls and get_transcript; the tools arrive in a new system message, leaving the cached prompt prefix unchanged. A label can also stand for a group of tools, such as all tools on an MCP server, or resolve to tools according to runtime permissions. For example, a pipeline-forecast skill can expose CRM tools while allowing only managers to update a forecast.

Pinned skills let an application detect a requested skill, such as /meeting-prep, and pass it through pinned_skills so its instructions are included before the next model call. The agent can then begin on model call 1 instead of model call 2, with the instructions and associated tools already in context. Each pinned skill is added once as a tagged message, preserving earlier messages and the prompt cache; a chat interface can display the skill as a label.

For reloads, an application can set skills_metadata to None when invoking an agent, prompting the next run to rescan the skill library and detect additions or changes without starting a new thread. A newly found skill changes the system prompt and invalidates the prompt cache. LangChain says this is usually less costly after an idle period because provider caches typically expire after minutes to an hour. The latest deepagents includes the updates; LangChain directs developers to its skills documentation and invites feedback through GitHub issues, its forum and X.

LangChain announced three updates to skills in Deep Agents on October 7, 2026: tools can be bound to skills, apps can pin skills before a model call, and long-running threads can reload skills. The changes address enterprise libraries that are growing to thousands of skills shared across teams and agents. Skills are folders of instructions, scripts and reference files; the open standard is supported by dozens of agent products and can be authored as a markdown file.

Deep Agents reveals each skill’s name and description at startup, reads its SKILL.md when a task matches, then loads supporting files as needed. This progressive disclosure keeps the agent’s context limited to information required for the task. A skill’s SKILL.md contains YAML frontmatter with a name and description, and may be accompanied by scripts/, references/ and assets/. LangChain’s example go-to-market agent has more than 50 skills, including meeting-prep, call-transcripts and competitive-intel-card.

With tool binding, a skill can list tools in metadata.include_tools, and those tools become available only after the agent reads that skill. Calling a bound tool before then fails as an unknown tool. In the example, reading call-transcripts unlocks search_calls and get_transcript; the tools arrive in a new system message, leaving the cached prompt prefix unchanged. A label can also stand for a group of tools, such as all tools on an MCP server, or resolve to tools according to runtime permissions. For example, a pipeline-forecast skill can expose CRM tools while allowing only managers to update a forecast.

Pinned skills let an application detect a requested skill, such as /meeting-prep, and pass it through pinned_skills so its instructions are included before the next model call. The agent can then begin on model call 1 instead of model call 2, with the instructions and associated tools already in context. Each pinned skill is added once as a tagged message, preserving earlier messages and the prompt cache; a chat interface can display the skill as a label.

For reloads, an application can set skills_metadata to None when invoking an agent, prompting the next run to rescan the skill library and detect additions or changes without starting a new thread. A newly found skill changes the system prompt and invalidates the prompt cache. LangChain says this is usually less costly after an idle period because provider caches typically expire after minutes to an hour. The latest deepagents includes the updates; LangChain directs developers to its skills documentation and invites feedback through GitHub issues, its forum and X.