LangChain Revamps Skills in Deep Agents with Tool Binding and Runtime Pinning

LangChain has revamped skills support in Deep Agents to address scaling needs for enterprise skill registries. The update introduces three key capabilities: binding tools to skills, pinning skills at runtime, and reloading skills mid-thread. These changes aim to optimize context usage and improve agent responsiveness as skill libraries grow to thousands of entries.

Binding Tools to Skills

Previously, skills and tools were disclosed separately, meaning agents could call tools without reading their associated instructions. Now, tools can be bound to a skill via metadata.include_tools in the skill's frontmatter. These tools remain out of context until the agent reads the skill, ensuring the agent understands how to use them before calling them. On models from Anthropic and OpenAI that accept mid-conversation tool updates, adding these tools preserves the prompt cache.

Pinned Skills

Users can now explicitly request a skill (e.g., /meeting-prep), and applications can pin it using the pinned_skills parameter. This loads the skill's instructions into the conversation before the first model call, eliminating the need for a read_file round trip. This reduces latency and guarantees the correct instructions are in context, while keeping the prompt cache valid by adding the skill as a tagged message.

Reloading Skills Mid-Thread

Long-running agents can now reload skills without starting a new thread. By setting skills_metadata to None during invocation, the agent rescans the skill library for added, edited, or deleted skills. While this invalidates the prompt cache, the cost is often negligible for idle threads where provider caches have already expired. This feature allows teams to update skill libraries dynamically without disrupting active sessions.