Harrison Chase of LangChain on Deep Agents, LangSmith, and Earning Trust | NVIDIA AI Podcast Ep. 297

In short

In this episode of the NVIDIA AI Podcast, Harrison Chase, CEO and co-founder of LangChain, discusses the evolution of LLMs and agentic systems. He covers the genesis of LangChain, the concept of deep agents, the importance of observability and evaluation, and the future of agents. Harrison also shares insights on mixing frontier and open models, the impact of Open Claw, and LangChain's involvement with the NVIDIA NeMoTron Coalition.

Key takeaways

• LangChain provides tools to simplify the creation of systems and agents around LLMs.
• Deep agents are a new type of agent harness that provides LLMs with more autonomy and a structured environment for interaction.
• LangSmith is LangChain's platform for observability and evaluation, addressing the need for transparency and control in agent behavior.
• The future of agents includes asynchronous sub-agents, agent memory, and agent identity.
• Open-source models are becoming increasingly important for driving agent harnesses and enabling cost-effective solutions.

Chapters

1. 00:00

   Introduction to Harrison Chase and LangChain

   Noah Kravitz introduces Harrison Chase, CEO and co-founder of LangChain, highlighting LangChain's rapid growth and focus on helping developers build applications with LLMs and agentic frameworks.

2. 01:43

   The Genesis of LangChain

   Harrison discusses the initial vision behind LangChain, which stemmed from observing the common patterns and increasing complexity in applications built on top of LLMs. LangChain aims to provide tools to simplify the creation of systems and agents around LLMs.

3. 03:02

   Deep Agents Explained

   Harrison defines deep agents as a new type of agent harness that provides LLMs with more autonomy and a structured environment for interaction. These agents share common architectures and patterns, making them general-purpose and customizable.

4. 05:24

   Enterprise Conversations and Autonomous Systems

   Harrison describes conversations with enterprise leaders about the balance between risk and reward when implementing autonomous agents. He emphasizes that not everything requires an autonomous agent and introduces LangGraph as an alternative for more directed workflows.

5. 07:26

   LangSmith: Observability and Evaluation

   Harrison introduces LangSmith, LangChain's platform for observability and evaluation, addressing the need for transparency and control in agent behavior. He discusses the importance of evaluation-driven development and building eval datasets.

6. 09:29

   Agent Development Lifecycle

   Harrison outlines LangChain's agent development lifecycle: build, test, run, and manage. He explains that the open-source tools cover the 'build' phase, while LangSmith focuses on testing, running, and managing agents.

7. 10:08

   Skills and Coding Agents

   Harrison explains how skills are used to package knowledge and tools for agents, particularly in coding agents. He discusses the importance of the environment in which agents run and the role of tools like NVIDIA's OpenShell.

8. 13:10

   The 'Aha' Moment for Deep Agents

   Harrison shares the 'aha' moment for deep agents, which came from recognizing common patterns in projects like Claude Code, MANAS, and Deep Research. He describes how the first version of Deep Agents was created during a weekend hackathon.

9. 14:38

   Evaluation-Driven Development and Trust

   Harrison discusses the importance of auditability and traceability in building trust in enterprise agents. He explains how evaluation-driven development helps ensure agents perform as expected and adapt to unexpected use cases.

10. 18:52

    Frontier vs. Open Models

    Harrison discusses the approach to mixing frontier and open models to achieve cost performance ratio. He explains how sub-agents can use specialized agents and the importance of open-source models driving the harness.

11. 24:50

    The NVIDIA NeMoTron Coalition

    Harrison discusses LangChain joining the NVIDIA NeMoTron Coalition. He explains the importance of open models and harnesses that they can run in, and how LangChain can provide the harness and work with NVIDIA to provide a model that can work with that harness.

12. 26:17

    The Future of Agents

    Harrison shares his vision for the future of agents, including asynchronous sub-agents, agent memory, and agent identity. He emphasizes the importance of proactive, always-on agents and the need for human-in-the-loop learning.

13. 32:50

    Open Claws Impact

    Harrison discusses the impact of Open Claw and how it has set a new objective for what agents can and should be able to do. He also discusses the importance of agent identity and how it is changing the way people think about agents.

14. 34:59

    Where to Learn More

    Harrison shares where to learn more about LangChain and the work they are doing. He recommends the LangChain blog and Twitter.

Get the next The AI Podcast recap

A short recap of every new episode, by email. Free for up to 5 shows.