在本期 20VC 节目中，主持人哈里·斯特宾斯（Harry Stebbings）邀请了 LangChain 创始人哈里森（Harrison），共同探讨构建人工智能代理的复杂性、代理工作流的未来以及不断演变的开发者生态系统。 哈里森解释说，LangChain 提供了一个将大语言模型连接到工具、数据和计算的开源框架，以及一个名为 Langsmith 的可观测性和评估平台。构建代理是出了名的困难，因为大语言模型是具有无限输入和输出空间的非确定性黑盒。随着代理执行一系列模型调用和工具交互，输入的微小变化可能会导致级联错误，这使得可靠性成为架构中最难的部分。哈里森指出，随着模型的改进，LangChain 不断删除和调整其产品的部分内容，以匹配原生工具调用等不断演进的 API 级功能。 对话探讨了“拥有你的智能”这一概念，该概念依赖于由模型、控制框架（harness）和上下文组成的开放且可控的代理架构。哈里森认为，中立的控制框架对于防止供应商锁定至关重要，并允许开发者跨不同的模型提供商迁移数据、内存和上下文。虽然像哈维（Harvey）和兰普（Ramp）这样的公司正在成功地针对专业领域对开源权重模型进行后训练，但哈里森建议首先从上下文和控制框架层入手，以便在深入进行后训练之前获得更高的投资回报率。 在模型路由方面，哈里森提出，针对特定领域或垂直代理的智能路由属于控制框架层，因为它需要状态和上下文感知。然而，网关级路由器可以有效地用于通用水平任务或编程。在评估代理时，哈里森强调了特定垂直领域基准、追踪和内部评估的实用性，尽管对数据进行人工检查仍然至关重要。他还谈到了持续学习，将其描述为一个外循环，用于审查代理产生的数据、生成见解并更新内存、控制框架代码或模型权重。 在讨论基础设施和市场趋势时，哈里森坦言，他低估了对推理的大规模、持续需求，以及与编程工具相关的水涨船高的成本。在 LangChain，编程工具的工程支出约占工程薪资的 30%，该公司通过 Langsmith 的可见性、预算控制以及名为 Open SWE 的内部云编程代理来管理这一支出。哈里森预测，代理的内存层目前尚处于萌芽阶段，并在未来两到三年内将在所有应用程序中变得普遍。他展望了这样一种未来：水平消费者代理可以与持有专有用户知识和 API 的垂直特定领域代理（如 Spotify 或 Strava）进行无缝交互。 在谈到风险投资格局时，哈里森分享说，早期获得 Benchmark 以及随后的红杉资本的支持对 LangChain 至关重要，并赞扬了他们深厚的运营和人才支持。他在快问快答环节结束了访谈，将比尔·格利（Bill Gurley）评为他在路演期间遇到的最具压迫感的人物，强调构建代理的难度仍然被严重低估，并表达了对未来五年人工智能创造性应用的兴奋之情。

In this episode of 20VC, host Harry Stebbings is joined by Harrison, founder of LangChain, to discuss the complexities of building AI agents, the future of agentic workflows, and the evolving developer ecosystem. Harrison explains that LangChain provides an open-source framework connecting LLMs to tools, data, and computation, alongside Langsmith, an observability and evaluation platform. Building agents is notoriously difficult because LLMs are non-deterministic black boxes with infinite input and output spaces. As agents execute a series of model calls and tool interactions, minor changes to inputs can lead to cascading errors, making reliability the hardest part of the architecture. Harrison notes that as models improve, LangChain constantly deletes and adapts parts of its product to match evolving API-level features like native tool calling. The conversation explores the concept of "owning your intelligence," which relies on an open and controllable agent architecture comprising the model, the harness, and the context. Harrison argues that neutral harnesses are vital to prevent vendor lock-in and allow developers to port data, memory, and context across different model providers. While companies like Harvey and Ramp are successfully post-training open-weight models for specialized domains, Harrison advises starting at the context and harness layers first to achieve higher ROI before diving into post-training. On model routing, Harrison posits that smart routing for domain-specific or vertical agents belongs at the harness layer because it requires state and context awareness. However, gateway-level routers can work effectively for general-purpose horizontal tasks or coding. When evaluating agents, Harrison highlights the utility of vertical-specific benchmarks, traces, and internal evals, though human inspection of data remains critical. He also touches on continual learning, describing it as an outer loop that reviews agent exhaust, generates insights, and updates memory, harness code, or model weights. Discussing infrastructure and market trends, Harrison admits to underestimating the massive, ongoing demand for inference and the skyrocketing costs associated with coding tools. At LangChain, engineering spend on coding tools sits at roughly 30% of engineering salaries, which the company manages through visibility via Langsmith, budget controls, and an in-house cloud coding agent called Open SWE. Harrison predicts that memory layers for agents are currently in their infancy and will become universally common across all applications in the next two to three years. He envisions a future where horizontal consumer agents interact seamlessly with vertical domain-specific agents, such as Spotify or Strava, which hold proprietary user knowledge and APIs. Addressing the venture landscape, Harrison shares that taking early backing from Benchmark and subsequently Sequoia was pivotal for LangChain, praising their deep operational and talent support. He concludes with rapid-fire insights, naming Bill Gurley as the most intimidating figure he met during pitches, emphasizing that building agents remains drastically underrated in its difficulty, and expressing excitement for the creative applications of AI over the next five years.