TechPrufer
Get started
LangChain
LangChain logoLangChain · Feature trail

Agent Memory Management

12 videos · Sep 2026

What this trail covers

LangSmith Custom Apps allows users to build, publish, and run custom UIs directly within LangSmith, leveraging LangSmith data. This feature enables the creation of focused interfaces for reviewing agent behavior, feedback, experiments, and traces. Enterprise customers can use LangSmith Chat to generate these apps from descriptions. This enhances the agent development lifecycle by providing tailored visualization and interaction layers on top of agent data and traces.

All LangChain trails
  1. How to go from your agent's traces to a fine-tuned model in one workflowThis post introduces LangSmith Fine-Tuning and the `smithtune` CLI, a tool that enables a direct workflow from LangSmith agent traces to custom fine-tuned models. It automates dataset creation from trajectories, integrates with training platforms like Fireworks and Baseten, and leverages LangSmith for evaluation. The workflow allows for iterative improvement by comparing base models against fine-tuned versions within LangSmith experiments.
  2. Introducing LangSmith Trajectories: A readable view of every agent sessionIntroduces 'Trajectories' in LangSmith, a new chronological, conversational view of agent sessions. Trajectories aggregate messages from humans, AI, and tools across the main agent and subagents, presenting them in their order of appearance. This feature aims to improve root cause analysis by providing a readable path of agent behavior, reduce friction for expert feedback by offering a less technical view than raw traces, and enable continuous improvement by allowing scoring with online evaluators and saving sessions to datasets for fine-tuning.
  3. Introducing LangSmith Engine v2: Detect and ship fixes to agents fasterEngine v2 introduces Red Teaming for proactive issue detection by generating and testing hypotheses about potential problems using production traces and repository context. It expands issue detection to include performance trends (error rate, latency, cost) and inefficient agent work (repetitive tool calls, long trajectories). For LangSmith Deployments, it automates the validation of proposed fixes by testing them against offending inputs and a broader eval set before presenting them for human review and one-click PR deployment.
  4. Introducing LangSmith Custom Apps: Build custom interfaces around your agent dataIntroduces LangSmith Custom Apps, a new capability allowing users to build and deploy custom user interfaces directly within LangSmith. These apps can be built using templates and AI-guided coding, and are designed to visualize and interact with LangSmith data such as traces, feedback, and agent behaviors. Enterprise users can leverage LangSmith Chat to generate these custom UIs via natural language prompts. This directly extends the agent development lifecycle by providing flexible and tailored interfaces for analyzing and interacting with agent data.
  5. Schedules for Managed Deep Agents: Cron jobs, prompts, and Slack deliveryThis post introduces the capability to schedule Managed Deep Agents to run at specific times, such as cron jobs, and deliver their outputs to designated channels like Slack. It demonstrates how to configure a weekly Salesforce pipeline summary that posts to Slack every Monday, while the same agent remains available for ad hoc questions. The process involves defining the schedule syntax, building the cron job and prompt, configuring Slack delivery, and deploying the scheduled agent to LangSmith for inspection and monitoring.
  7. What Is LangChain, Actually?This post clarifies that LangChain has evolved from an open-source framework for wrangling LLMs to a comprehensive platform covering the full agent development lifecycle. It highlights LangSmith as the production-ready platform for building, testing, deploying, and monitoring agents, emphasizing its use by over 7000 organizations for shipping production agents. The post also mentions LangGraph as a low-level orchestration framework and Deep Agents for building complex agents.
  8. How To Build A Harness With Jev | A LangChain x TypeSafe ConversationThis conversation introduces Jev, a System 1 model from TypeSafe AI, and its integration into LangChain for building agent harnesses. It details Jev's capabilities in classification-style tasks, its efficiency compared to traditional LLMs, and its application in model routing and as an evaluator. The discussion includes live demos of safety classification and model routing using Jev within an agent harness, and explores Jev's context length limitations and potential future developments like multimodal support. It also covers other production use cases and community Q&A regarding Jev's training data and threshold settings.
  9. What is LangSmith?This post introduces LangSmith as LangChain's comprehensive platform for agent development. It highlights key features such as tracing agent execution to debug issues, using datasets and evaluators for testing, and deploying/monitoring agents in production. The post also mentions annotation queues, insights, and the LangSmith Engine, positioning LangSmith as a central hub for the agent development lifecycle.
  10. Building a Harness with JevIntroduces Jev, a System 1 model that offers significant speed and cost advantages over LLMs for classification tasks. Details its application in model routing, auto-mode for tool calls, and as an evaluation judge within agentic systems, with integration examples for LangChain.
  11. How Do Agents Remember Anything Between Runs?The post explains the distinction between an agent's context window (short-term memory) and persistent storage mechanisms (long-term memory). It clarifies that the context window is limited by the LLM's token capacity, while long-term memory involves storing and retrieving information across multiple agent interactions. This foundational explanation is key to understanding agent state management.
  12. Why Sierra uses one agent per brand instead of multi-agent systemsThis post argues against the default adoption of multi-agent systems for organizational convenience, proposing that a single agent per brand, coupled with strong context engineering, is often a more effective and less complex approach. The core technical insight is that splitting agents for organizational reasons deprives them of critical context, which is detrimental to value. The emphasis is on maximizing the value of a single agent's context rather than managing inter-agent communication and state.
  13. How to manage shared and user credentials with Managed Deep Agents ConnectionsThis post details three methods for managing credentials and secrets for Managed Deep Agents (MDA) to interact with external tools and data. It covers agent-owned secrets (a single shared key), user-owned OAuth through MCP (using Linear as an example), and user-owned OAuth with a custom application. The talk demonstrates the MDA connections CLI for creating and managing connections, LangSmith Studio for running tools and triggering OAuth flows, and the process of registering OAuth apps in the connections catalog. It also shows how to write custom tools that leverage user tokens for actions like searching and creating issues.