Harrison Chase is the co-founder and CEO of LangChain. What began as an open-source library for LLM applications has expanded into LangGraph for agent orchestration and LangSmith for observability and evaluation. These lessons trace his thinking about frameworks, context engineering, agent reliability, memory and product design. — Reflections on Three Years of Building LangChain.

Visual summary of operating lessons from Harrison Chase.

Part 1: The Philosophy of Orchestration and Frameworks

  1. On the Role of Frameworks: LangChain began by capturing recurring patterns Chase saw in early LLM applications, so developers could reuse abstractions instead of rebuilding them each time. — Reflections on Three Years of Building LangChain.
  2. On Composability: LangChain's stack evolved from a general library into LangGraph for stateful orchestration and LangSmith for observing and evaluating agent behavior. — Reflections on Three Years of Building LangChain.
  3. On Orchestration Layers: A framework provides reusable interfaces for combining models, data, and tools while leaving developers control over the application's workflow. — Latent Space Interview.
  4. On Open-Source Roots: Chase started LangChain as a side project in 2022 and folded repeated patterns from early LLM experiments into an open-source library. — Reflections on Three Years of Building LangChain.
  5. On Stable Interfaces: LangChain separated its stable core abstractions from a fast-growing integration ecosystem so providers could evolve without destabilizing the central package. — LangChain v0.1.0.
  6. On Harnesses and Frameworks: Chase distinguishes a framework's reusable primitives from a more opinionated agent harness, and expects most teams to adopt rather than build the latter. — Sequoia Context Engineering Interview.
  7. On Rapid Iteration: As models and the ecosystem changed quickly, Chase says LangChain's tooling also evolved from quick-start chains toward production agents, LangGraph and LangSmith. — LangChain's Second Birthday.
  8. On the Right Abstraction: A useful framework abstracts recurring components—models, prompt templates, vector stores and tools—while allowing developers to assemble the specific application they need. — Latent Space Interview.
  9. On Learning from Users: Chase describes LangChain's evolution as a response to what developers needed in practice: moving beyond prototypes toward reliable production applications. — LangChain's Second Birthday.
  10. On Ecosystem Integration: LangChain's model, data-store and tool integrations give developers reusable components across an expanding provider ecosystem. — LangChain v0.1.0.

Part 2: The Art and Science of Context Engineering

  1. On Context Engineering: Chase defines context engineering as building dynamic systems that provide a model the right information and tools in the right format for a task. — The Rise of Context Engineering.
  2. On Dynamic Context: For an agent, context must be assembled and updated across the workflow rather than treated as one static prompt. — The Rise of Context Engineering.
  3. On Retrieval as a Workflow: A retrieval pipeline can inspect whether an initial answer has the needed information and query again when the retrieved material is insufficient. — LangGraph.
  4. On Context Quality: Good context engineering selects and formats task-relevant material; simply adding more raw information does not ensure the model receives what it needs. — The Rise of Context Engineering.
  5. On Prompt Templates: Prompt templates are reusable components that organize variable inputs before they are sent to a model. — Latent Space Interview.
  6. On Finite Context Windows: Even as context windows grow, long-running agents must manage accumulated history, including by compacting it when necessary. — Sequoia Context Engineering Interview.
  7. On Retrieval Decisions: An agentic retrieval workflow can decide whether to search again when the first retrieved context does not answer the question. — LangGraph.
  8. On Few-Shot Guidance: Chase's agents lecture uses examples and explicit tool-use guidance to help agents choose the right action, including when no tool is needed. — Full Stack Deep Learning Agents Lecture.
  9. On Structured Outputs: Raw model strings need parsing before downstream programs can reliably use them; structured formats and output parsers make that handoff more dependable. — Full Stack Deep Learning Agents Lecture.
  10. On Graph-Grounded Context: In a LangChain–Neo4j–AWS collaboration, Chase highlighted grounding LLM applications in enterprise data; graph and vector retrieval can provide complementary context. — Neo4j–AWS–LangChain Partnership.

Part 3: Building Reliable Agentic Systems

  1. On What Makes an Agent: Chase describes agents as systems in which a language model helps decide which actions or tools to use next in pursuit of a task. — Sequoia Orchestration Interview.
  2. On the Action–Observation Loop: In ReAct-style agents, a model selects a tool, receives its observation, and uses that result to decide whether another step is needed. — Full Stack Deep Learning Agents Lecture.
  3. On Reliable Production Control: For critical workflows, Chase argues that domain-specific steps should remain explicitly controlled rather than left entirely to a nondeterministic model loop. — Sequoia Orchestration Interview.
  4. On Observability: LangSmith traces expose agent inputs, tool calls and outputs so teams can diagnose behavior they could not infer from a final answer alone. — Reflections on Three Years of Building LangChain.
  5. On Evaluating Trajectories: Chase says traces reveal what actually happened at intermediate steps and can guide human evaluation of agent quality. — Sequoia Context Engineering Interview.
  6. On Human Approval: For consequential actions such as refunds, ambient agents can pause for a person's approval instead of acting without review. — Sequoia Ambient Agents Interview.
  7. On Planning and Execution: Separating higher-level planning from step-by-step execution can help agents handle more complex objectives and revisit the goal. — Plan-and-Execute Agents.
  8. On Tool Restraint: Agent design must teach a model when not to call a tool as well as which tool to call. — Full Stack Deep Learning Agents Lecture.
  9. On Resumable Agents: LangGraph's statefulness and durable execution let a long-running agent resume after interruption instead of restarting the whole task. — Reflections on Three Years of Building LangChain.
  10. On Smaller Models: LangChain's plan-and-execute design separates planning from execution, leaving room to substitute smaller, faster models for suitable components. — Plan-and-Execute Agents.

Part 4: Memory, Tools, and the Agent Stack

  1. On Episodic Memory: In Chase's agent-memory taxonomy, episodic memory stores examples of past behavior that an agent can use to guide later actions. — Memory for Agents.
  2. On Procedural Memory: Procedural memory captures instructions and learned ways of performing a task, distinct from stored facts or examples. — Memory for Agents.
  3. On Stateful Workflows: LangGraph adds cycles and conditional transitions so an agent can return to earlier steps as its state changes. — LangGraph.
  4. On Tool Sandboxing: Coding agents can run untrusted or destructive commands, so their execution environment should isolate those actions from sensitive systems. — LangSmith Sandboxes.
  5. On Retrievable Memory: Chase's agents lecture describes vector stores as one way to retrieve relevant earlier information for an agent's long-running work. — Full Stack Deep Learning Agents Lecture.
  6. On Subagent Handoffs: Chase identifies a specific multi-agent failure: a subagent may finish work but return a final message that does not convey the useful result to its parent. — Sequoia Context Engineering Interview.
  7. On Semantic Memory: Semantic memory represents facts and knowledge an agent can recall, distinct from episodic examples and procedural instructions. — Memory for Agents.
  8. On Persistent State: LangGraph persists state across agent steps so a workflow can pause and resume with its context intact. — Reflections on Three Years of Building LangChain.
  9. On Parsing Tool Calls: Agent systems need to turn model output into valid tool calls; structured output and parsers reduce this failure mode. — Full Stack Deep Learning Agents Lecture.
  10. On Graph-Shaped Orchestration: LangGraph moves beyond fixed linear chains by supporting loops, branches and state-dependent transitions. — LangGraph.

Part 5: The Future of Deep Agents and Product UX

  1. On Long-Horizon Agents: Chase's deep-agent discussion emphasizes work that spans longer periods, with planning and context management beyond a short tool-call loop. — Sequoia Context Engineering Interview.
  2. On Human-Correctable UX: Ambient agents need interfaces that let people inspect, correct and approve work as it unfolds. — UX for Agents: Ambient.
  3. On Agents as Collaborators: Chase describes agents as systems that can do an initial tranche of work and hand results back for human judgment, rather than replacing that judgment. — Sequoia Context Engineering Interview.
  4. On Asynchronous Research: Long-horizon research agents can work in the background and return a draft after sustained investigation rather than responding instantly. — Sequoia Context Engineering Interview.
  5. On Product Quality: Chase argues that proprietary context, memory, evaluation and feedback loops—not generic model access alone—can make an agent product improve with use. — Own Your Intelligence.
  6. On Personalization through Memory: Chase says his email agent became less useful after a platform move lost its accumulated memory, illustrating why persistent memory matters for personal agents. — Sequoia Context Engineering Interview.
  7. On Human-Reviewed Research: Chase points to research across documents and logs as a useful long-horizon task because a human can review and improve an agent's first draft. — Sequoia Context Engineering Interview.
  8. On Domain-Specific Control: Chase expects generic planning to improve inside models, while domain-specific workflows and required steps still need explicit orchestration. — Sequoia Orchestration Interview.
  9. On Human Judgment in Evaluation: Chase says evaluation of agents must incorporate human judgment about the work they do, not just conventional software pass/fail tests. — Sequoia Context Engineering Interview.
  10. On the Mission of the Stack: Chase describes LangChain's mission as making it easier to build useful context-aware, agentic applications; LangGraph and LangSmith developed as production needs grew. — LangChain's Second Birthday.

Learn more:

  1. Latent Space Interview
  2. Sequoia Orchestration Interview
  3. LangChain's Second Birthday
  4. Full Stack Deep Learning Agents Lecture
  5. Announcing LangSmith
  6. LangGraph
  7. Sequoia Ambient Agents Interview
  8. Sequoia Context Engineering Interview
  9. The Rise of Context Engineering
  10. Memory for Agents
  11. UX for Agents: Ambient
  12. Reflections on Three Years of Building LangChain
  13. LangChain v0.1.0
  14. LangSmith Sandboxes
  15. Planning for Agents
  16. Own Your Intelligence
  17. Neo4j–AWS–LangChain Partnership
  18. Plan-and-Execute Agents