LangChain AI Updates: July 1, 2026

1. LangChain Details Two Approaches for Running Untrusted Agent Code

LangChain. LangChain published guidance on securely executing agent-generated code through two mechanisms: LangSmith Sandboxes, which run code in remote isolated containers, and Code Interpreters for Deep Agents, which provide an in-process runtime. The post, authored by Hunter Lovell, addresses the risk of agents executing code influenced by potentially compromised inputs, offering options to limit access and pause for human approval. It matters because agents that orchestrate multiple tools increasingly need safe, controllable execution environments. Source

2. LangChain Introduces the Wiki Memory Pattern for Agents

LangChain. In an In the Loop post, Harrison Chase introduced “wiki memory,” a pattern where agents transform unstructured source material into a compact, persistent, agent-readable knowledge layer maintained as files. Unlike traditional RAG, which retrieves raw chunks at query time, wiki memory precomputes and maintains synthesized knowledge for more efficient reasoning. The approach reframes agent memory as a curated knowledge base rather than raw data retrieval. Source

3. LangChain and Harbor Launch a Unified Stack for Evaluating Agents

LangChain. LangChain announced an integration between Harbor, Deep Agents, and LangSmith aimed at evaluating long-running, stateful agents. Authored by Nicholas Bohm and Nick Hollon, the partnership lets developers run hundreds of agent trials in parallel within isolated cloud sandboxes while capturing detailed traces that explain why evaluations pass or fail. It targets the gap that standard eval runners leave when testing complex, stateful agent behavior. Source