LangChain AI Updates: October 14, 2026

1. LangSmith Adds AI Search for Natural Language Queries Over Agent Traces

LangChain. LangChain added AI Search to LangSmith, which lets users describe a behavior in plain language, such as “did the user get frustrated?” or “did PII leak?”, and get back the traces where it occurred. Rather than using vector embeddings, a fast decision model reads each candidate trajectory, the deduplicated per-session view that SmithDB builds from traces produced by the OpenAI and Anthropic SDKs, LangChain, LangGraph, the Vercel AI SDK, and OpenTelemetry. Each trajectory is packed into an 8K token budget filled in priority order, with user messages and agent replies first, then intermediate reasoning, tool results, and system messages. Existing structured filters such as error status or tags narrow candidates first, and matches stream back as they are found. Source

2. langchain-scaledown Integration Brings Task-Specific Context Models to Deep Agents

LangChain. LangChain released a langchain-scaledown integration and added LangSmith LLM Gateway support for ScaleDown’s constrained decoding models (Summarize, Compress, Extract, and Classify), which are trained for single tasks and accept up to 1M tokens of input, including images and scanned PDFs. The package ships three middlewares for LangChain and Deep Agents, including ScaledownSummarizationMiddleware, a drop-in replacement for Deep Agents’ built-in summarization that takes the same trigger and keep options, plus middleware for tool result compression and structured extraction. On ScaleDown’s own benchmarks against the average of GPT-5.5, GPT-5.4 Mini, and GPT-5.4 Nano, Summarize is reported as 41x cheaper and 2.7x faster and Compress as 138x cheaper and 1.8x faster. Calls routed through the LLM Gateway are traced and count toward spend and rate limits. Source