Harrison Chase

Harrison Chase is co-founder and CEO of LangChain, the open-source framework and commercial tooling company that helps developers wire language models into applications and agents. He came out of statistics and production machine learning, then turned a late-2022 side project into the default glue stack many teams reach for when models alone are not enough.

Harvard stats into production ML

He graduated from Harvard in 2017 with degrees in statistics and computer science, after finding machine learning through sports analytics and treating the two majors as the same craft in different clothes. At Kensho Technologies he led an entity-linking team that connected messy real-world mentions into structured knowledge for financial products. He then led machine learning at Robust Intelligence, an MLOps firm focused on testing and validating models outside demo conditions. Those jobs left him with a bias toward grounding and reliability when models leave the notebook.

From side project to LangChain

In October 2022 he released LangChain as an open-source Python package that organized emerging patterns for chaining models to data, tools, and memory. ChatGPT's launch the next month turned the package into a rush of downloads. With Ankush Gola he founded the company around that project in early 2023. The commercial line grew into LangSmith for tracing, debugging, and evaluation, and later LangGraph for stateful workflows that mix prescribed steps with model-directed actions and human checkpoints.

Context engineering and traces

Under his leadership LangChain stayed aimed at builders: open packages plus hosted ops for composing prompts, tools, retrieval, and longer-running agents. On Sequoia's Training Data conversation he argued that long-horizon agents make observability existential: "you don't actually know what the context at step 14 will be, because there's 13 steps before that that could pull arbitrary things in." The through-line, he said, is that "everything's context engineering," and that "the source of truth for software is in code, and for agents it's a combination now of code, and traces are where you can see the source of truth."

On autonomy that earns its risk

In the same Sequoia talk he stressed online testing over offline unit tests for agents, because "behavior doesn't emerge until it's actually being used with real world inputs." Conference writing and product work under his watch treat evaluation, human oversight, and reversible actions as first-class parts of shipping LLM software. He has publicly acknowledged early abstraction problems (hidden prompts, breaking changes, heavy dependencies) and pushed redesigns toward a simpler agent loop. The through-line is software engineering discipline applied to systems that are non-deterministic by default.