How LangSmith and LangChain OSS Help You Meet EU AI Act RequirementsLangSmith与LangChain开源版如何助力您符合欧盟《人工智能法案》要求

Observability and tracing: Full execution capture可观测性与链路追踪：全链路执行捕获

Regulators want a record of the actions an AI system takes. For agents making multi-step decisions, good practice is to trace the full thread, including inputs, reasoning, tool calls, and outputs.监管机构要求留存人工智能系统所有操作的记录。对于做出多步决策的智能体，最佳实践是追踪完整的执行链路，包括输入、推理过程、工具调用和输出结果

What the Act requires:法案要求：

• Article 9 requires a living risk management system across the development lifecycle第9条要求在整个开发生命周期内建立动态风险管理系统
• Article 12 requires automatic event logging over the system's lifetime, sufficient to identify risks, support post-market monitoring, and enable operational oversight by deployers第12条要求在整个系统生命周期内进行自动事件日志记录，日志需足以识别风险、支持上市后监测，并便于部署方开展运营监督
• Article 13 requires traceable, interpretable decisions第13条要求决策可追溯、可解释

LangSmith gives you full observability and evaluation tools for every step of your agent's execution.LangSmith为您提供全链路的可观测性与评估工具，覆盖智能体执行的每一个步骤

What LangSmith provides:LangSmith提供的功能：

• End-to-end tracing captures every LLM call, tool invocation, and reasoning step with structured metadata: inputs, outputs, timestamps, and agent context.端到端链路追踪可捕获每一次大语言模型调用、工具调用和推理步骤，附带结构化元数据：输入、输出、时间戳和智能体上下文
• LangSmith Studio visualizes the full execution graph, including state transitions and tool calls, so you can inspect the agent's decision-making process step by step.LangSmith Studio可可视化完整的执行图谱，涵盖状态转换和工具调用，支持您逐步核查智能体的决策过程
• LangSmith Insights Agent processes trace data to automatically identify and cluster recurring patterns, surfacing failure modes and usage trends that would otherwise require manual review.LangSmith Insights Agent可处理链路追踪数据，自动识别并聚类重复出现的模式，呈现原本需要人工审核的故障模式和使用趋势
• Custom dashboards track risk scores and trigger alerts through PagerDuty or webhooks when a metric crosses your threshold.自定义仪表盘可追踪风险评分，当指标超出您设定的阈值时，可通过PagerDuty或webhooks触发告警

Retention and storage:留存与存储：

Self-hosted, BYOC, and managed cloud deployment options give you control over where logs live and how long they're retained.自托管、BYOC（自带云账号）和托管云部署选项可让您自主控制日志的存储位置和留存时长

In managed cloud, base traces are retained for 14 days, designed for short-term debugging and ad-hoc analysis. Extended traces are retained for 400 days, intended for ongoing model improvement, evaluation, and human feedback. You can upgrade base traces to extended at any time, and bulk export trace data for long-term archival.在托管云环境中，基础追踪数据留存时长为14天，适用于短期调试和临时分析；扩展追踪数据留存时长为400天，用于持续的模型优化、评估和人工反馈。您可以随时将基础追踪升级为扩展追踪，也可批量导出追踪数据用于长期归档

For EU data residency requirements specifically, LangSmith EU keeps all trace data in-jurisdiction. With self-hosted and BYOC options, the entire stack runs in your Kubernetes cluster or cloud region. Your data never leaves your perimeter.针对欧盟数据驻留要求，LangSmith EU会将所有追踪数据存储在法域范围内。选择自托管或BYOC选项时，全栈技术栈将运行在您的Kubernetes集群或云区域中，您的数据绝不会离开您的私有网络边界

Evaluators: Continuous quality and safety scoring评估器：持续质量与安全评分

The EU AI Act requires ongoing measurement, with evaluations on production traffic.《欧盟人工智能法案》要求持续开展测量，并对生产流量进行评估

What the Act requires: Several articles demand ongoing measurement of your agent's outputs:法案要求：多项条款要求对智能体的输出进行持续测量：

• Article 10 requires data governance and bias examination across development and testing datasets第10条要求对开发和测试数据集进行数据治理与偏见审查
• Article 13 requires that systems be transparent enough for deployers to interpret outputs and use them appropriately第13条要求系统具备足够的透明度，便于部署方解读输出结果并合理使用
• Article 15 requires declared levels of accuracy and relevant accuracy metrics, adversarial resilience, and protection against common attack surfaces第15条要求公开声明的准确率水平及相关准确率指标、对抗韧性，以及对常见攻击面的防护

LangSmith's online evaluators continuously score a configurable sample of production traces, with filters you define. Each score is logged with full trace context, giving you an evidence trail. When a metric crosses a threshold, alerts fire through PagerDuty or webhooks.LangSmith的在线评估器可对您自定义筛选的生产流量追踪样本进行持续评分，每次评分都会记录完整的追踪上下文，形成可追溯的证据链。当指标超出阈值时，会通过PagerDuty或webhooks触发告警

LangSmith provides prebuilt evaluators across all of these areas:LangSmith在上述所有领域均提供预置评估器：

• Bias and fairness based on characteristics like race, gender, age, religion, nationality, disability, and sexuality基于种族、性别、年龄、宗教、国籍、残疾状况、性取向等特征的偏见与公平性评估
• Toxicity toward individuals or groups针对个人或群体的有害内容检测
• Sensitive imagery and explicit content敏感图像与露骨内容检测
• Hallucination and answer relevance to catch outputs that mislead users幻觉与答案相关性检测，识别误导用户的输出内容
• PII leakage to flag accidental exposure of sensitive attributes个人敏感信息（PII）泄露检测，标记敏感属性的意外泄露
• Prompt injection and jailbreaking for adversarial input detection提示词注入与越狱检测，识别对抗性输入
• API leakage and code injection covering common attack surfaces in tool-calling agentsAPI泄露与代码注入检测，覆盖工具调用类智能体的常见攻击面
• Correctness, exact match, plan adherence, and task completion for accuracy measurement正确性、精确匹配、计划遵循度和任务完成度评估，用于准确率测量
• Tool selection and plan adherence to score agent decision quality工具选择与计划遵循度评估，用于衡量智能体决策质量

Every evaluator is customizable, and you can create new ones for behaviors specific to your use case.所有评估器均可自定义，您还可以针对自身用例的特定行为创建新的评估器

Human oversight: Interrupt, review, and escalate人工监督：中断、审核与升级

Human oversight is one of the Act's core principles. Consequential decisions made by AI systems should remain contestable and correctable by people. In practice, that means building oversight into the architecture with defined escalation paths, structured review workflows, and audit evidence that intervention happened.人工监督是该法案的核心原则之一。人工智能系统做出的重大决策应始终可由人类提出异议并进行纠正。实践中，这意味着需将监督机制嵌入架构中，配备明确的升级路径、结构化的审核流程，以及留存人工干预的审计证据

For agentic systems, this carries extra weight. An agent making multi-step decisions can compound errors before a human has a chance to catch them. In some cases, oversight mechanisms need to be embedded in the execution graph itself.对于智能体系统而言，这一要求尤为重要。智能体做出多步决策时，可能在人类有机会发现问题前就累积错误。部分场景下，监督机制需直接嵌入执行图谱中

What the Act requires: Article 14 requires that humans can understand, intervene on, override, and interrupt the system.法案要求：第14条要求人类能够理解系统、对系统进行干预、覆盖系统决策，以及中断系统运行

What LangSmith provides:

• LangGraph's interrupt primitive makes human-in-the-loop (HITL) a first-class part of the agent graph. You can pause execution, inspect state, modify it, and resume at any node.LangGraph的中断原语将人在回路（HITL）设为智能体图谱的一等公民组件。您可以在任意节点暂停执行、检查状态、修改状态后恢复运行
• LangSmith Deployment provides the durable runtime underneath: automatic checkpointing, exactly-once execution, and resume-from-exact-point recovery for paused runs. This ensures reliable HITL interrupts in production.LangSmith Deployment提供底层的高可用运行时：支持自动检查点、恰好一次执行，以及暂停任务从精确节点恢复，确保生产环境中人在回路（HITL）中断的可靠性
• Annotation queues route production traces to human reviewers for structured feedback.标注队列可将生产流量追踪数据分配给人工审核员，以获取结构化反馈
• Webhooks fire when evaluators exceed defined thresholds or interrupt events occur, so you can page the right person through PagerDuty, or your preferred incident response system.当评估器超出设定阈值或发生中断事件时，webhooks会触发告警，您可通过PagerDuty或您偏好的事件响应系统通知相关负责人

Where to start入门指引

August 2 is close. For teams running high-risk AI systems, here's how LangSmith helps you meet the Act's core technical requirements.距离8月2日的截止日期已非常接近。对于运营高风险人工智能系统的团队，LangSmith可通过以下方式帮助您满足法案的核心技术要求

Observability and tracing are the foundation. Full tracing across every tool call, retrieval step, and reasoning node gives you the audit trail and the foundation to run evaluations.可观测性与链路追踪是基础。对每一次工具调用、检索步骤和推理节点进行全链路追踪，可为您提供审计追踪，也为开展评估奠定基础

Evaluations on production traffic, including scoring for bias, hallucination, toxicity, accuracy, and adversarial inputs, address Act's post-market monitoring requirements.对生产流量开展评估，包括偏见、幻觉、有害内容、准确率和对抗性输入的评分，可满足法案的上市后监测要求

Human-in-the-loop is an architectural requirement. The Act requires that humans can intervene on, override, and interrupt the system. LangGraph's interrupt primitive and LangSmith's annotation queues make that mechanism auditable.人在回路是架构层面的要求。法案要求人类能够干预系统、覆盖系统决策并中断系统运行，LangGraph的中断原语和LangSmith的标注队列可使该机制可审计

To meet EU data residency requirements, deployment matters too. LangSmith's EU SaaS, BYOC, and full self-hosted options are designed for agent workloads in production. The right choice depends on how much operational control you need, and we're happy to walk through the tradeoffs.为满足欧盟数据驻留要求，部署方案的选择也至关重要。LangSmith的欧盟SaaS、BYOC和全自托管选项均针对生产环境中的智能体负载设计，具体选择取决于您对运营控制权的需求，我们很乐意为您详解各方案的利弊

These are the same practices that teams already follow to run agents well in production.这些都是团队在生产环境中稳定运行智能体时已经遵循的最佳实践

• Get started with LangSmith开始使用LangSmith
• Speak to a LangSmith expert咨询LangSmith专家
• Explore the trust center访问信任中心

Article crosswalk条款对照表