LangChain AI Updates: September 26, 2026

1. LangSmith Engine v2 Adds Red Teaming and Validates Its Own Fixes

LangChain released Engine v2, the in-platform agent in LangSmith that diagnoses issues from traces. It now reproduces a failure, proposes a change, re-runs the same inputs, and iterates until the fix works before a human reviews it, with one-click deployment via PR. A new red teaming mode, in private beta for LangSmith Deployment users, probes agents for hallucinations and prompt violations before they reach production. Engine v2 is available on Plus and Enterprise SaaS plans, with self-hosted support coming later. Source

2. LangSmith Fine-Tuning Turns Agent Traces Into LoRA Models

LangChain launched LangSmith Fine-Tuning in public beta, built around a smithtune CLI that curates agent trajectories into datasets, runs supervised fine-tuning with LoRA, and compares the result against the baseline. Training runs on Fireworks or Baseten, so users bring an API key from one of those providers rather than managing GPUs, with open models such as Kimi K3 and Qwen-3.8-27B named as targets. LangChain reports a code review task where F1 rose from 48.9% to 53.7% with 29.8% fewer model calls. Source

3. Managed Deep Agents v0.8 Adds Per-User Auth, Two-Layer Memory, and HTTP Channels

LangChain shipped Managed Deep Agents v0.8 with user-owned and agent-owned credentials across 23 OAuth services including Linear, GitHub, and Google Workspace. Memory is now split into agent-level memory shared across users and user-level memory scoped to the authenticated caller, with defaults that vary by channel. A new HTTP channel accepts JSON webhooks for internal tools and portals, and Parallel-powered web search is built in and free during the beta. Source

4. Trajectories Give LangSmith a Readable, Chronological View of Agent Sessions

LangChain added Trajectories to LangSmith, which flatten nested trace runs into an ordered sequence of human, AI, and tool messages across a main agent and its subagents. They work with traces from LangChain, LangGraph, Deep Agents, the OpenAI and Claude SDKs, and coding agents such as Codex and Claude Code. Trajectories can be scored with online evaluators and exported as supervised fine-tuning data, and are available on all LangSmith plans in the US. Source

5. LangSmith Custom Apps Lets Teams Build Internal UIs on Trace Data

LangChain introduced Custom Apps, which run custom interfaces such as annotation queues, experiment comparison views, and trace review screens natively inside a LangSmith workspace. Apps can be generated from a chat prompt on Plus and Enterprise plans or built in code from templates against the LangSmith API, using the open-source Macaw design system. Plus plans get one Custom App per organization and Enterprise plans get unlimited apps. Source