LangChain AI Updates: October 8, 2026

1. Deep Agents Adds Skill-Bound Tools, Tool Labels, and Pinned Skills

LangChain. The latest deepagents release lets a skill declare its tools in metadata.include_tools in its SKILL.md frontmatter. Those tools are passed to SkillsMiddleware(tools=[...]) and stay out of context until the agent reads the skill. On Anthropic and OpenAI models, the tools are added mid-conversation, which leaves the cached prompt prefix intact. Skills can also list labels that a resolver function maps to tool groups (for example, every tool on an MCP server, or tools filtered by user permissions). A pinned_skills argument to agent.invoke injects a skill’s instructions without a read_file round trip, and setting skills_metadata to None forces a rescan of skills mid-thread. LangChain says this is aimed at large libraries like its own GTM agent, which has more than 50 skills. Source

2. Managed Deep Agents v0.9 Adds Agent-Created Schedules and Per-Run Configuration

LangChain. Managed Deep Agents v0.9 (Public Beta) adds a Schedules SDK that agents can call to create cron-based or one-time (at) tasks during a conversation. Each schedule runs with the requesting user’s permissions and posts results back to the channel or thread where it was created. An agent can now be defined as a callable that returns a define_deep_agent definition at the start of each run, so callers can choose the model, instructions, skills, MCP servers, and sandbox by passing context to client.runs.create. LangChain says this lets one deployment serve multiple teams or repositories while keeping out-of-scope tools hidden from the model. The Slack channel also gains a reactions option to control the emoji the agent adds to acknowledge incoming messages. Source