Organize Your Agents Like Data Products像管理数据产品一样组织您的智能体
By: Matt Marzillo AI Product Specialist and Elliott Botwick Principal AI/ML Architect作者：Matt Marzillo（AI 产品专家）和 Elliott Botwick（首席 AI/ML 架构师）
When adding Cortex Agents to an account that already contains business data, create a clear boundary between the data layer and the AI layer.在包含业务数据的账户中添加 Cortex 智能体时，请在数据层和 AI 层之间建立清晰的界限。
Keep raw and curated data in the databases that own it. Organize agents and their supporting objects in domain-focused schemas using Snowflake’s standard database.schema.object hierarchy:将原始数据和经整理的数据保留在各自所属的数据库中。使用 Snowflake 标准的 database.schema.object 层级，在以领域为中心的 schema 中组织智能体及其支持对象：
Keep objects together by domain. Put an agent and the semantic views, Cortex Search services, custom tools, and related stages it uses in the same domain schema wherever practical. Much like your data products, your agent products should be organized similarly. This simplifies ownership, permissions, deployment, and troubleshooting.按领域将对象归类在一起。在切实可行的情况下，将智能体及其使用的语义视图（semantic views）、Cortex Search 服务、自定义工具和相关阶段（stages）放在同一个领域 schema 中。正如您的数据产品一样，您的智能体产品也应以类似方式组织。这简化了所有权管理、权限分配、部署和故障排查。
Model use cases, not source systems. Build focused semantic views around business questions and organizational needs such as sales performance or customer support.A semantic view should ideally map 1:1, or 1:few, with data schemas. A single semantic view ideally should not span several unrelated schemas.对用例建模，而非对源系统建模。围绕业务问题和组织需求（如销售业绩或客户支持）构建专注的语义视图。理想情况下，一个语义视图应与数据 schema 保持 1:1 或 1:few 的映射关系。单个语义视图不应跨越多个不相关的 schema。
Use the right tool for the job. Use semantic views for structured business analysis. Use Cortex Search services for unstructured content and fuzzy text matching. Add skills for repeatable workflows that should be reused across agents. And MCPs to connect to external systems. Use CoCo’s skills catalog to accelerate development. Meaningful tool descriptions in the Agent specification goes a long way to improve accuracy and efficiency.选择合适的工具。使用语义视图进行结构化业务分析。使用 Cortex Search 服务处理非结构化内容和模糊文本匹配。为需要在多个智能体间复用的可重复工作流添加技能（skills）。使用 MCP 连接外部系统。利用 CoCo 的技能目录来加速开发。在智能体规范中提供有意义的工具描述，将极大提升准确性和效率。
Control access with roles. Create domain-specific roles and grant only the required access to the agent, database, schema, warehouse, semantic views, search services, and custom tools. Remember that agent requests use the user’s default role. Apply masking and row access policies to the underlying tables.通过角色控制访问权限。创建特定领域的角色，并仅授予智能体、数据库、schema、仓库、语义视图、搜索服务和自定义工具所需的访问权限。请记住，智能体请求使用的是用户的默认角色。对底层表应用掩码和行访问策略。
Separate development from discovery. Let domain teams own and maintain their agents. Use the account-level Snowflake CoWork object to control which agents users discover and access through CoWork.将开发与发现分离开来。让领域团队拥有并维护他们自己的智能体。使用账户级的 Snowflake CoWork 对象来控制用户可以通过 CoWork 发现和访问哪些智能体。
Version and evaluate before release. Test representative questions, evaluate committed agent versions, and use production traces and feedback to guide changes.在发布前进行版本控制和评估。测试具有代表性的问题，评估已提交的智能体版本，并利用生产环境的追踪记录和反馈来指导改进。
The pattern is straightforward: databases own data, schemas define domains, objects provide capabilities, roles control access, and CoWork manages discovery.该模式非常直接：数据库拥有数据，schema 定义领域，对象提供能力，角色控制访问，而 CoWork 管理发现。
Read the supporting guidance:阅读相关指导：
- Semantic view modeling语义视图建模
- Semantic view development and deployment语义视图的开发与部署
- Cortex SearchCortex Search
- Agent skills智能体技能
- Cortex Agent evaluationsCortex 智能体评估
