LangChain AI Updates: October 24, 2026

1. LangChain Publishes a Deep Agents Guide to Retrieval as Context Engineering

LangChain. A new Deep Agents conceptual guide from LangChain argues that agent retrieval fails in three ways: inefficient searching, too much context arriving at once, and irrelevant or stale context. The post explains how the Deep Agents harness handles the first two by default. Its filesystem tools search before they read (glob, grep, then paged read_file), and oversized tool results are saved to files so the agent only gets a path and a preview. For custom search tools such as RAG, it recommends a “retrieve, offload, delegate” pattern: results are saved to files and subagents read them, so the main agent sees only what is relevant. Source

2. Deep Agents Code 0.2.1 Adds Skill-Scoped MCP Tools

LangChain. Deep Agents Code 0.2.1 lets MCP tools be scoped to individual skills. The release also makes Claude Sonnet 5.5 the default model for the Anthropic Auto classifier and switches that classifier to native structured output. A bug fix migrates installs to Python 3.14 automatically when no other session is running. Source

3. LangGraph SDK and CLI Add Thread Priority Settings

LangChain. langgraph-sdk 0.4.9 adds a Priority type for thread priority settings in the Python SDK. langgraph-cli 0.4.38, released the same day, adds a matching priority section to langgraph.json so deployments can set it in configuration. Source Source