编码智能体现可通过自然语言提示设计完整的 CARTO 工作流：该工作流为多步骤、可审计且可直接共享。本文介绍 CARTO 智能体技能（CARTO Agent Skills）的实现机制。

Coding agents can now design full CARTO Workflows from a prompt: multi-step, auditable, and ready to share. Here's how CARTO Agent Skills make it work.

工作流设计是一个迭代过程。即使已有扎实的计划，分析师仍需检查中间结果，以验证连接操作（joins）、核查评分分布，并确认增强分析（enrichment）行为是否符合预期，方可继续推进：该连接操作是否正确？结果是否符合预期？ 下面介绍两个整合性示例，展示 CARTO 的命令行工具（CLI）、场景化技能（use-case skills）以及 MCP 应用（MCP Apps）如何赋能编码智能体（coding agents）执行强大的分析任务。 一家位于纽约的连锁超市正规划下一阶段扩张，希望快速甄选出最优的十个候选门店地址，在最大化顾客需求的同时，最小化与现有门店网络的覆盖重叠，从而降低自相蚕食（cannibalization）风险，并确保每一项评分因子对最终结果的贡献均完全透明。分析师仅需向智能体提出如下请求： 随后，该智能体即调用 carto-site-selection 和 carto-trade-area-analysis 两项 Agent Skills，自动构建一个由 CARTO 驱动的工作流，并将整个流程结构化为受管控、可复用的步骤。在工作流运行过程中，分析师与智能体共同审阅中间输出，包括需求面（demand surfaces）、商圈覆盖重叠情况以及各项评分构成要素。这种共享可见性支持快速迭代与优化，最终产出一项可复用的分析资产——该资产可重复执行、跨团队共享，亦可被 AI 智能体调用，从而规模化生成一致的扩张建议。 一名物流规划师需在德克萨斯州选定新建配送中心的最佳位置，在满足客户需求数量的同时，兼顾极端天气风险暴露程度，以实现韧性与运营绩效的最大化。他向智能体提出如下请求： 为此，智能体调用 carto-composite-scoring Agent Skill，将需求信号与风险信号融合为一个透明、可审计的模型，随后自动构建并验证一个 CARTO 工作流。最终成果是一项可复用、受管控的分析资产，其全部 SQL 逻辑均由系统自动生成。规划师随后在 CARTO 中运行该工作流，获得快速、可靠的决策支持，以及一套完全可追溯的方法论——该方法论可直接复用于未来其他规划场景，亦可灵活调整适配。 哈维尔（Javier）现任 CARTO 首席产品官（Chief Product Officer），在全球范围内全面负责 CARTO 的产品组合、客户体验及创新路线图。其职责覆盖完整的产品开发周期与客户体验生命周期。 将 CARTO MCP Server 接入 Microsoft Copilot Studio，即可让每个团队在 Microsoft 365 内，以自然语言方式便捷获取地图、应用与空间分析能力。 我们邀请了七款前沿 AI 模型构建 deck.gl 地图，渲染全部结果，并将其中五份方案提交至苏黎世举办的开放可视化峰会（Open Visualization Summit）。我们从中获得的关键洞见，以及 deck.gl 对 CARTO 至关重要的原因。

Workflow design is an iterative process. Even with a solid plan, analysts need to inspect intermediate results to validate joins, check score distributions, and confirm enrichment behaves as expected before moving forward: does this join look right? Are the results as expected? Let’s look at two prompts that bring everything together and showcase how CARTO’s CLI, use-case skills, and MCP Apps enable coding agents to perform powerful analytical work. A supermarket chain in New York is planning its next phase of expansion and wants to quickly identify the best ten candidate store locations, maximizing customer demand while minimizing overlap with the existing network to reduce cannibalization, with full transparency into how each scoring factor contributes to the final result. The analyst simply asks the agent: Then the agent builds a CARTO Workflow powered by carto-site-selection and carto-trade-area-analysis Agent Skills, automatically structuring the process into governed, reusable steps. As the workflow runs, both the analyst and agent review intermediate outputs such as demand surfaces, overlap between trade areas, and scoring components. This shared visibility enables fast iteration and refinement, resulting in a reusable analytical asset that can be rerun, shared across teams, or invoked by AI agents to produce consistent expansion recommendations at scale. A logistics planner needs to identify the best locations for a new distribution center in Texas, balancing customer demand with exposure to severe weather risk to maximize resilience and performance. They ask the agent: To do this, the agent uses the carto-composite-scoring Agent Skill to combine demand and risk signals into a transparent, auditable model, then builds and validates a CARTO Workflow automatically. The result is a reusable, governed analytical asset with fully generated SQL logic. The planner then runs the workflow in CARTO, gaining fast, confident decision support and a fully traceable methodology that can be reused or adapted for future planning scenarios. Javier is CARTO's Chief Product Officer and is globally responsible for CARTO’s product portfolio, customer experience, and innovation roadmap. Javier's responsibilities span the complete product development and customer experience cycle. Connect the CARTO MCP Server to Microsoft Copilot Studio and give every team maps, apps, and spatial analysis in plain language, inside Microsoft 365. We asked seven frontier AI models to build deck.gl maps, rendered every result, and took five proposals to the Open Visualization Summit in Zürich. What we learned, and why deck.gl matters so much to CARTO.