LangChain AI Updates: June 12, 2026

1. LangChain Publishes Guidance on Choosing Sandboxes for Agent Code Execution

LangChain. LangChain published a guide on selecting sandboxes for agents that generate and execute code, framing the risk around a “lethal trifecta” of sensitive data access, exposure to untrusted content, and external communication. The post outlines five sandbox features it considers essential, including an isolated filesystem, limited network access, resource limits, controlled reusability, and kernel-level isolation. It applies Meta’s “Rule of Two” principle to recommend that agents never operate fully autonomously when all three risk conditions are present, and references LangSmith, LangGraph, and Deep Agents in the discussion. Source

2. Box Builds Enterprise Content Agent on Deep Agents

LangChain. LangChain detailed how Box built its enterprise content agent on Deep Agents to support cross-document analysis while retaining model flexibility and faster iteration. The architecture uses a parent and child agent model where a Global Agent classifies requests and spawns child agents to handle parallel tasks such as document search and synthesis. Middleware handles citations, prompt caching, and context management, and the team uses LangSmith for debugging and evaluation, enabling new agents to ship in weeks rather than months. Source