🔶 GrassLobster：用于参数化几何工作流的 AI Agent 生成

英文标题：🔶 GrassLobster: AI Agentic Generation of Parametric Geometry Workflows

GrassLobster 通过将 Grasshopper 画布上的几何流程镜像为可读、可执行的项目文件，让外部 AI Agent 能够理解并针对性修改参数化几何工作流。其核心组织方式是连接式“Geometry Stations”，将输入、主几何、结构逻辑、次级几何和输出等任务拆分为具备明确依赖关系的模块。项目还可通过 Agent Instructions、领域知识和几何参考资料向 Agent 提供上下文，但并非自动索引的知识系统。该方案面向可替换的兼容 Agent，并为基于可测量结果的参数迭代与优化提供了潜在路径。

打开原文相关 5 · 机会 1

中文

几何过程始终对你可见；同一过程也会以文本形式镜像给 Agent。

画布上的几何站点

GrassLobster Session Components 将模型组织为相互连接的 Geometry Stations（几何站点）。每个 Station 都会运行针对一项有意义任务的代码。你可以看到任务序列及其依赖关系，而无需追踪每一个单独操作。

一个项目可能依次经历输入、主几何、结构逻辑、次级几何和输出等阶段。结构遵循任务本身。GrassLobster 隐藏的是实现复杂性，而不是几何逻辑。

镜像的逻辑与输入

Station 定义和可执行代码会镜像到易读的项目文件中。Agent 可以定位相关逻辑、理解连接的 Stations，并通过受支持的工作流进行针对性修改。

受支持的输入值同样会被镜像。尺寸、数量或间距可以在不重写 Station 代码的情况下修改；Grasshopper 会重新计算相关几何。

包含项目上下文的文件夹

Agent Instructions（Agent 指令） 指导 Agent 如何提出问题以及如何与你协作。Domain knowledge（领域知识） 可加入预先准备的专业指引、你自己的标准或材料说明。Geometric references（几何参考） 则通过截图、研究资料或示例模型传达形状信息。

参考资料可以来自聊天内容或指定的项目文件夹。Agent 能够使用哪些资料，取决于其工具以及文件格式。这是项目上下文，而非一个自动建立索引的知识系统。

保持外置的 Agent

该项目围绕可互换的 Agent 设计，这些 Agent 需要具备必要的文件与工具能力。它并不绑定某一个内置模型。随着兼容 Agent 在编程、规划和几何推理能力上的进步，同一项目结构也可以从中受益。

通往 Agentic 优化的路径

当一个模型能够提供可衡量的结果时，Agent 可以改变输入、比较输出结果，并朝着某个目标逐步优化——例如，在保持建筑面积的同时减少材料体积。

这是一种可能的发展方向，而不是内置的通用优化器。它取决于模型、可用输出以及 Agent 工具。结构目标还需要采用恰当的分析方法。

“我想和我的文件对话。”

在使用 Grasshopper 大约十二年后，Miro Bannwart 开始使用 AI Agent，并希望将这种对话带入自己的参数化项目中。

Grasshopper 已经为人们提供了一种以可视化方式理解几何的方法。Agent 则需要一种能够读取和修改的表示形式。这促成了文本镜像：在画布上保留几何过程，同时通过文件让 Agent 访问同一个项目。

English / 原文

The geometric process stays visible to you. The same process is mirrored as text for the agent.

Geometry Stations on the canvas

GrassLobster Session Components organise the model into connected Geometry Stations. Each Station runs the code for a meaningful task. You see the sequence and dependencies without following every individual operation.

A project might move through inputs, primary geometry, structural logic, secondary geometry and output. The structure follows the task. GrassLobster hides implementation complexity, not geometric logic.

Mirrored logic and inputs

Station definitions and executable code are mirrored in readable project files. The agent can locate the relevant logic, understand connected Stations and make targeted changes through the supported workflow.

Supported input values are mirrored too. Dimensions, counts or spacing can be changed without rewriting the Station code; Grasshopper recalculates the related geometry.

A folder with project context

Agent Instructions guide how the agent asks questions and works with you. Domain knowledge adds prepared subject guidance, your own standards or material notes. Geometric references communicate shape through screenshots, studies or example models.

References can come from chat or designated project folders. What the agent can use depends on its tools and the file format. This is project context, not an automatically indexed knowledge system.

An agent that stays external

The project is designed around interchangeable agents, with the necessary file and tool capabilities. It is not tied to one built-in model. As compatible agents improve at coding, planning and geometry reasoning, the same project structure can benefit.

A path toward agentic optimization

Where a model exposes measurable results, an agent could vary inputs, compare outcomes and refine toward a goal — for example, maintaining floor area while reducing material volume.

This is a possible direction, not a built-in general-purpose optimizer. It depends on the model, available outputs and agent tools. Structural goals also require appropriate analysis methods.

“I wanted to talk to my files.”

After around twelve years with Grasshopper, Miro Bannwart began working with AI agents and wanted to bring that conversation into his parametric projects.

Grasshopper already gave people a visual way to understand geometry. Agents needed a representation they could read and change. That led to text mirroring: keeping the geometric process on the canvas while giving the agent access to the same project through files.