How To Build a Self-Improving Agent with LangSmith Engine and Context Hub

Most agents don't learn from their mistakes, they just leave traces. 

Jake Broekhuizen walks through how to close that loop using LangSmith Engine and Context Hub, turning production traces into durable memory updates that continuously improve your agent's behavior. 

You'll see a full end-to-end build with NOVA, a financial assistant Deep Agent, going from local memory to versioned context in Context Hub to automated issue detection and fixes via Engine.

How LangSmith Engine Turns Agent Traces Into Durable Memory
0:00 Why agents repeat the same mistakes
0:27 What is agent memory?
0:55 Working memory vs long-term memory
1:21 The three types of long-term memory
1:52 The read/write loop between memory layers
2:18 How LangSmith Engine and Context Hub fit together
3:35 Demo: Creating a Context Hub repo for NOVA
4:45 Building NOVA locally with Deep Agents
5:57 State backend vs composite backend
7:44 Connecting NOVA's traces to LangSmith
8:44 Setting up Engine: linking GitHub and Context Hub
9:16 How Engine scans traces and surfaces issues
10:07 Reviewing an Engine-detected issue: banned filler words
11:07 Applying the fix directly from Engine
12:14 Confirming the memory update worked
12:32 Wrap-up: the continual learning loop in action

Extra resources:
- LangSmith Engine: https://www.langchain.com/langsmith/engine
- Deep Agents: https://www.langchain.com/deep-agents 
- LangSmith Context Hub: https://www.langchain.com/blog/introducing-context-hub