LangSmith Engine v2 validates proposed agent fixes before presenting them

• LangSmith Engine v2 shifts its agent-improvement loop toward proactive failure discovery, with LangChain's Engine v2 announcement saying it finds and resolves agent problems before users encounter them.
• The release adds red teaming, proposed-fix validation, and detection of inefficient work plus cost and latency trends, according to LangChain's v2 feature list.
• LangSmith Trajectories place human, AI, tool, and system messages into a single chronological session view, as LangChain's explainer describes.
• The Interrupt bundle also includes Managed Deep Agents 0.8, LangSmith Fine-Tuning, and Custom Apps, which LangChain's Interrupt recap lists alongside Engine v2 and Trajectories.
• LangChain's trace-data workflow centers on monitoring every trace, clustering errors, and turning them into evals, Vtrivedy10's trace workflow says.

The Engine architecture post says a production failure can become a future-catching evaluator, a regression dataset example, and a code fix. An August Engine update claims more than twice the internal-benchmark performance on issue identification and 25% better issue fixing on industry benchmarks. Managed Deep Agents schedule docs label that product public beta and available only on LangSmith Cloud in the US region.

LangChain named three additions to Engine v2:

1. Red Teaming identifies agent issues before they reach production.
2. Validated Fixes tests Engine's proposed changes before presenting them.
3. Improved issue identification surfaces inefficient trajectories and trends in cost and latency.

The product framing starts from a production trace rather than a static test suite. LangChain's v2 announcement describes Engine as an in-platform agent that analyzes traces, finds issues, proposes fixes, and monitors for regressions autonomously.

The [link:1|Engine announcement] describes a three-step fix loop tied to the inputs that exposed an issue:

1. Reproduce the failure on the offending inputs using LangSmith Deployment.
2. Develop, test, and refine a proposed change against those same inputs.
3. Present the results and validated fix, with a quick PR available for deployment.

That gives the word “validated” an operational meaning: a candidate change has to pass a reproduction-based test before it reaches the review surface.

A trajectory pulls together four kinds of events:

• Human messages
• AI messages
• Tool messages
• System prompts

LangChain's follow-up says the view spans the main agent and any subagents, ordered by when each item first appeared. It is a readable session-level surface alongside the nested traces Engine analyzes.

Vtrivedy10 described LangSmith's longer feedback loop as four trace operations:

• Monitor every trace with custom models, Jev, low-cost small models, or built-in rubrics.
• Cluster traces and errors into human-readable views.
• Turn errors into evals that can be improved iteratively.
• Open-source Harbor Environment tooling for evaluations or reinforcement learning from trace data.

The operating premise is that production trace data supplies both incident evidence and post-training material, not merely a debugging log.

Custom Apps was another product in LangChain's Interrupt recap. LangChain's follow-up named three internal views teams can build:

• Annotation queues for reviewing and labeling one or many runs.
• Experiment comparisons for side-by-side outputs and metrics.
• Trace reviews for inspecting trace or thread history.

A schedule declaration in a project-level schedules/ directory can supply a cron expression and prompt, according to LangChain's schedule demo. The official schedule documentation says mda deploy provisions each declared schedule as a LangSmith cron after the deployment is live.