Harrison Chase

本文介绍了如何通过Auth Proxy来保护LangSmith代理沙箱的网络访问，确保在大规模部署代理时的安全性。Auth Proxy通过在网络层控制和管理代理与外部服务的交互，实现了凭据的安全管理、网络访问的显式控制以及团队职责的清晰分离。

AI coding costs are rising sharply; using open-source models can significantly reduce expenses. Kimi K2.6 on BaseTen is ~5x cheaper than Opus 4.7 with comparable performance on most tasks.

企业需定制AI以实现业务价值，而非依赖通用模型。控制AI行为与演进的系统才是核心竞争力。

Harrison Chase 推出 Eval Engineering Skill，通过 Harbor 工具自动化构建评估流程，但仍需人工参与。

NVIDIA Nemotron 3 Ultra在LangChain Deep Agents Harness上实现行业领先的性能，成本降低10倍，任务完成率提升。

本地优先AI代理通过LangGraph和子代理实现复杂任务编排，提升响应效率。

Deep Agents 框架新增编程调用子代理功能，提供6种实现方式及开源资源。

KV-cache命中率是生产级AI代理最关键的指标，提示缓存可降低41%-80%的推理成本。

SmithDB 是为代理追踪设计的数据库，通过对象存储实现高效全文搜索索引。

LangChain 提供了部署智能代理的完整指南，涵盖前端集成、后端架构和生产级持久化。

LangSmith 引入了‘睡眠时间计算’机制，通过后台分析代理轨迹并更新记忆存储，提升智能代理的长期记忆能力。

通过微调开源模型，可以实现比前沿模型便宜10-100倍的生产代理追踪问题检测，同时保持高精度。

可观测性对于智能代理是必不可少的，需结合 LangGraph 和 LangSmith 工具实现。

Harrison Chase is excited about an open-source AI agent designed for actual continual learning.

本文介绍了 LangChain 的沙盒 Auth Proxy，这是一种控制代理生成行为与外部世界之间边界的工具。通过使用 Auth Proxy，可以安全地管理代理对网络资源的访问，防止未授权的访问和潜在的安全风险。

构建智能代理的最艰难事实是，只有在生产环境中才能真正了解它们的行为。LangChain 的联合创始人 Harrison Chase 强调了在开发和部署智能代理时面临的挑战，包括不可预测的行为、安全性和责任问题。他建议通过在受控环境中进行测试和监控来减轻这些风险，并强调了持续学习和适应的重要性。

LangChain introduces the Agent Development Lifecycle (ADLC), dividing agent development into four phases—build, test, deploy, and monitor—emphasizing that its fundamental difference from traditional software lies in infinite input/output spaces and non-determinism, with successful teams adopting a "ship early, iterate fast" pattern.

Agent observability isn't just for debugging — to enable continuous learning, you must collect or generate feedback data directly within your observability platform.

Harrison Chase指出切换大模型供应商容易，但切换AI应用层的‘harness’（如LangChain）更难；厂商正通过私有harness锁定用户，亟需开放标准。

Nemotron 3.5 Lightning模型集成Switchyard基准后，成本降低74%，准确率仅下降6点，推动代理系统优化。

改进智能体的核心在于大规模数据挖掘与实验，强化学习、Harness Engineering等方法均需依赖数据驱动的优化。

Verifiers are crucial for reinforcement learning, but their costs are high. Through team collaboration and optimization methods, verifier costs can be significantly reduced.

Harrison Chase recommends using LangChain’s LangSmith Fleet tool to build no-code agents via natural language, accelerating real-world automation with free courses available today.

Harrison Chase and AWS co-publish a deep dive guide on evaluating DeepAgents using LangSmith, enabling observability and reliability for long-horizon AI systems through structured data points and evaluators.

LangSmith Engine currently only supports LangSmith traces, but it's easy to trace to LangSmith via OTEL and 30+ framework integrations.

Harrison Chase emphasizes that agent improvement relies on combining observability with feedback; logging alone is insufficient—teams must actively integrate direct, indirect, and generated feedback into their observability platforms.

LangChain proposes binding user feedback with agent traces to transform observability from a debugging tool into a self-learning system, enabling continuous optimization of AI agents.

Harrison Chase highlights a key AI trend: LLMs now mature enough to power autonomous web-browsing agents, citing DeepAgents + BrowserBase as an early working example.

为智能体添加记忆功能可通过三个步骤实现，包括运行智能体、分析其行为轨迹、更新记忆。

Leve 是一个基于 LangGraph 的文件系统优先的持久化代理框架，通过目录结构定义代理并编译运行。