What you'll learn

• 1 (04:30) **What LangChain Builds & Why Agents Are So Hard** - Harrison defines LangChain’s two products: an open-source framework for connecting models to tools/data, and LangSmith for observability and evals. He then explains the core difficulty of building agents: LLMs are non-deterministic black boxes with infinite input spaces, leading to cascading errors in multi-call agent systems.
• 2 (06:29) **How Model Improvement Forces Product Erasure** - Harrison explains that as models get smarter, LangChain must constantly delete large parts of its product that are no longer needed, similar to Cursor’s “delete the product” operating principle.
• 3 (08:50) **Why the World Needs Neutral Harnesses** - Harrison argues that neutral harnesses are essential because agents build up state (memory, context) over time, and you don’t want that state locked to a single model provider.
• 4 (10:46) **What It Means to “Own Your Intelligence”** - Harrison breaks down owning your intelligence into controlling three layers: the model, the harness, and the context. Most companies should start at the context/harness layer before touching model weights.
• 5 (11:56) **Where Agent Value Really Lives: Verification & Vertical Agents** - Harrison argues that if value is derived from verification (knowing whether the agent did well), the winners will be domain experts who can write those verifiers.
• 6 (13:16) **The Reality of “Building Your Own Model”** - Harrison clarifies that most companies like Harvey and Ramp are not training models from scratch — they are post-training open-weight models (often Chinese open-source models), which is a valuable but later-stage optimization.
• 7 (16:29) **Model Routing: Where It Belongs and Why It’s Not a Commodity** - Harrison dives into the debate: should model routing live at the gateway layer (simple API call, portable) or the harness layer (more state-aware)?

+ Full timestamped outline available in the app