LangChain CEO & co-founder, agent engineering platform
Harrison Chase
Profile
Harrison Chase is the closest thing the LLM application layer has to a standards body — not because anyone appointed him, but because he shipped first and kept shipping. In October 2022, while working as an ML engineer at Robust Intelligence, the Harvard statistics-and-CS grad (previously at Kensho) open-sourced a Python package called LangChain. It was a weekend-project-shaped thing that wrapped prompt templates, output parsers, and tool calls into composable “chains.” Two months later ChatGPT landed, tens of thousands of developers went looking for scaffolding, and LangChain was the only scaffolding there. He and colleague Ankush Gola spun it out into a company in 2023; Benchmark and Sequoia funded it in quick succession.
What makes Chase interesting to developers isn’t the framework so much as the vocabulary. “Chains,” then “agents,” then “RAG,” then “agent harness,” then “context engineering” — a striking amount of how the AI engineering community talks about building things traces back to a LangChain blog post. His April 2025 essay How to think about agent frameworks is the best single piece of writing on the topic by anyone, and notably it argues against heavy abstraction: the hard part, he says, is controlling exactly what context reaches the model at each step, and frameworks that hide that from you are actively harmful. Coming from the person who sells the most-used framework, that’s an unusually honest position. Two months later he coined the framing that stuck industry-wide in The rise of context engineering.
The company has grown up considerably. LangGraph — a low-level, graph-based, durable execution runtime — replaced the old chain abstractions as the real product, and LangSmith (tracing, evals, monitoring) is the commercial engine. Both LangChain and LangGraph hit 1.0 in October 2025, alongside a $125M Series B led by IVP at a $1.25B valuation on roughly $16M ARR. In 2026 the focus moved to Deep Agents, an open-source, model-agnostic harness — planning, subagents, virtual filesystem, memory, all batteries-included — plus a managed hosted version and LangSmith Engine, announced at their own Interrupt conference in May 2026.
If you’re learning to build with AI in 2026, the honest read is this: you probably shouldn’t reach for LangChain reflexively, and Chase would half-agree with you. Model providers absorbed a lot of what the original framework existed to paper over, and plenty of production teams are better served writing the loop themselves. But read everything he publishes. His In the Loop essays and his Interrupt keynotes are where the field’s operating concepts get argued out first — currently the claim that generic models are a commodity and the durable advantage lives in the harness, the context, and the feedback loop you own. He’s also a genuinely good teacher: his DeepLearning.AI short courses with Andrew Ng introduced more developers to LLM application development than any textbook.
Key Articles & Papers
How to think about agent frameworks The rise of context engineering Introducing ambient agents LangChain and LangGraph Agent Frameworks Reach v1.0 LangChain raises $125M to build the platform for agent engineering What does it mean to 'own your intelligence'? 3 Years of Graph Engineering with LangGraph How We Benchmark Deep Agents Introducing Interrupt: The AI Agent Conference by LangChainVideos
Controversies
The abstraction backlash. LangChain is one of the most criticized libraries in modern Python. The complaints are consistent and largely fair: dependency bloat, breaking changes across minor versions, documentation that lagged the code, and layers of abstraction that made it hard to see the actual prompt being sent. A well-circulated critique, Let’s Talk about LangChain, captured the mood — and many teams did rip it out in favor of raw OpenAI or Anthropic SDK calls or lighter tools like LlamaIndex. To Chase’s credit, the response was substantive rather than defensive: LCEL, then LangGraph’s explicit low-level control, then the 1.0 rewrite that stripped the framework back. His own writing now concedes the core point about hidden context.
Open source versus the business. LangSmith is closed source and commercial, which has drawn recurring suspicion that the open-source packages get designed to funnel users toward it. Chase’s counter is that the framework is Apache-licensed, model-agnostic, and increasingly thin, while observability and evals are genuinely hard infrastructure people are willing to pay for. It’s a reasonable defense, but the tension is real and worth knowing about before you build a company’s agent stack on the ecosystem.
Hype-cycle vocabulary. Terms like “ambient agents” and “deep agents” get accused of being marketing dressed as taxonomy. Sometimes true — but the concepts have generally outlived the launch posts, which is more than most AI coinages manage.
Spotify Podcasts
YouTube
