Intel / article

LangChain October release roundup: LangSmith, Deep Agents v0.9, FastMCP

Source: langchain.com
Author: LangChain
Date: Published Oct 19, 2026

Context

LangChain's October newsletter rounds up what it shipped at its Interrupt events in New York and London, and it groups them under one argument: lasting value goes to companies that own the model, the and the that make a generic model fit their business (langchain.com). Several of the LangSmith additions are about seeing what agents do. AI Search lets teams query traces by meaning, with questions like whether a user got frustrated or personal data leaked, instead of building filters; Trajectories lays out a session chronologically across a main agent and its subagents; and Engine v2 adds and, according to LangChain, tests and validates the fixes it proposes.

The release that most directly carries the ownership argument is LangSmith . It takes the trajectories an agent produced in production, turns them into training data, fine-tunes a model through Baseten or Fireworks, and then evaluates and serves the tuned model, so the record of an agent doing its job becomes the material for a model specialized to that job. An earlier LangChain post described it alongside a smithtune CLI. Elsewhere in LangSmith, Managed Deep Agents reached v0.9 with run-specific personalization, Slack reactions and user-level memory, and on the open-source side LangChain v1.4.0 supports FastMCP through a langchain.mcp namespace. The newsletter describes each item in a sentence, so how well any of them works is not something it establishes.

Key takeaways · AI-distilled

LangSmith AI Search finds traces by semantic questions such as whether a user got frustrated or whether PII leaked, instead of relying on traditional filters.
LangSmith Engine v2 adds red teaming before production and, per LangChain, detects inefficient agent work and metric trends, then tests and validates its own proposed fixes.
LangSmith Fine-Tuning turns production trajectories into training data, fine-tunes through Baseten or Fireworks, then evaluates and serves the tuned models.
LangChain v1.4.0 adds FastMCP support through the langchain.mcp namespace, and Managed Deep Agents v0.9 adds user-level memory, scheduled tasks, Slack reactions and free built-in web search.

Terms in this piece · Glossary

agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
red teaming — Deliberately attacking your own AI system to find what makes it fail before someone else does.
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.

Read the source langchain.com