Harrison Chase

Chase studied statistics and computer science at Harvard. In a Latent Space interview, he described learning programming and data science through sports analytics, where predictions could be compared with the next game. He subsequently worked in machine learning at Kensho and Robust Intelligence. He recalled building a hackathon system at Robust Intelligence that answered questions using Notion and Slack, an early example of the data-connected language-model applications that LangChain would support.[3][8]

Chase released the first LangChain package in late 2022, then teamed up with Ankush Gola to establish the company in early 2023. The original package helped developers connect language models with application data and tools. In his three-year retrospective, Chase describes the shift from prototyping chatbots toward deploying agents that take actions. The company expanded into LangGraph for agent execution and LangSmith for observing, evaluating, and deploying systems. Chase is its co-founder and chief executive.[1][2][9]

LangGraph gives developers a way to define the steps and state of an agent as a graph, including loops. An agent might retrieve documents, judge whether they answer the question, and try another search. A developer can control which transitions are allowed instead of leaving every decision to the model. The LangChain team introduced the library in January 2024. It belongs to the company’s work under Chase’s leadership; that does not establish that he personally authored every component.[5]

Context engineering

Chase uses context engineering to describe the systems that supply a model with instructions, information, and tools at the moment it needs them. In his 2025 essay, he argues that unreliable agents often lack the right context. This moves debugging beyond rewriting a single prompt: an engineer must inspect how earlier actions changed the information available at a later step. His claim is an engineering perspective, not a guarantee that context changes can fix every model failure.[4]

Inspecting long-running agents

In a 2026 Sequoia interview, Chase explains why traces matter for agents that act repeatedly. A trace records the calls and intermediate results that produced an outcome, helping an engineer reconstruct what the model saw when it made a decision. He also discusses file systems as a way to store information outside a limited context window. These ideas connect agent architecture with practical supervision: inspect the run, identify the failure, change the system, and evaluate the change.[6]