LangSmith Launches Engine v2, Managed Deep Agents v0.8, and Fine-Tuning Tools

LangChain announced a suite of updates to LangSmith focused on the agent development lifecycle, specifically targeting runtime capabilities and observability. The releases include LangSmith Engine v2, Managed Deep Agents v0.8, LangSmith Trajectories, LangSmith Fine-Tuning, and Custom Apps.

LangSmith Engine v2

Engine v2 enhances the platform's in-platform agent for automating issue detection and fixes. Since its May launch, Engine has analyzed over 60 million traces. New features include:

• Red Teaming: Generates hypotheses about potential issues using production traces and repos, testing them to surface confirmed failures before they impact production.
• Expanded Detection: Identifies performance trends (error rate, latency, cost) and inefficient behaviors like repetitive tool calls.
• Automated Fix Validation: For agents on LangSmith Deployment, Engine tests candidate fixes against eval sets and presents satisfactory solutions for one-click PR deployment.

Managed Deep Agents v0.8

This update addresses production challenges by adding identity-scoped authentication and memory. Key additions include:

• User-Level Memory: Stores caller-specific context separate from shared agent-level memory, with access policies to prevent data leakage.
• Expanded Channels: Slack support now includes file transfer, and new HTTP channel support allows connections to any service sending JSON webhooks.
• Built-in Web Search: Powered by Parallel, this prebuilt tool provides current information without requiring separate vendor accounts or API keys.

Trajectories and Fine-Tuning

LangSmith Trajectories provides a conversational view of agent sessions, aggregating messages from humans, AI, and tools chronologically to simplify debugging and SME reviews. This data feeds into LangSmith Fine-Tuning, which uses the smithtune CLI to conduct supervised fine-tuning (SFT) on open models. The workflow handles dataset preparation, training via Baseten or Fireworks, evaluation, and serving.

Custom Apps

Teams can now build, publish, and run custom interfaces on top of LangSmith data using APIs and templates. This allows for tailored workflows such as annotation queues and experiment comparisons directly within the LangSmith workspace.