LangChain has expanded LangSmith with tools that test agents before deployment, isolate each user's memory and turn production sessions into training data. The releases cover both the runtime that executes an agent and the evaluation layer used to find failures.
Engine v2 can derive potential failure cases from production traces and code repositories, run those tests and surface confirmed problems for review. For agents on LangSmith Deployment, it can reproduce an offending input, test proposed fixes against a broader evaluation set and present a pull request once a candidate passes. It also looks for latency, cost and error trends as well as wasteful repeated tool calls.
Managed Deep Agents 0.8 adds memory scoped to an authenticated user, separate from shared agent memory. Access policies apply at each layer, and external connections can use either agent-owned or user-owned credentials. The managed runtime also gains built-in web search, Slack file transfer and an HTTP channel for services that send JSON webhooks.
A new Trajectories view arranges human messages, model replies, tool calls and subagent handoffs into a readable session. Reviewers can annotate those sessions, and the SmithTune command-line workflow can convert selected trajectories into supervised fine-tuning datasets, train open models through supported providers and evaluate the result. These automation features reduce manual work, but teams still need independent tests and human review before deploying generated fixes.