Cursor 现在可以通过创建画布来直观呈现信息。这让你能够探索和交互自定义界面,而不是在聊天或难以消化的 markdown 文件中阅读大段文字。
借助画布,智能体可以为真实世界数据创建仪表盘,也能根据你的需求构建带有逻辑和交互性的自定义界面。智能体可以利用画布帮你审查 PR、学习新库,甚至在 Cursor 中管理其他智能体。在智能体窗口中,画布是持久的工件,与终端、浏览器和源代码管理等工具并存。
组件作为构建模块
Cursor 使用基于 React 的 UI 库渲染画布,该库内置了表格、方框、图表和流程图等第一方组件。我们让智能体能够访问 Cursor 中已有的组件(如差异对比和待办事项列表),并指示其遵循数据可视化最佳实践。
你可以创建技能来教智能体如何制作不同类型的画布。例如,"文档画布"技能能让 Cursor 生成你代码仓库的交互式架构图。
我们在 Cursor 中如何使用画布
我们发现画布对数据密集型任务特别有用。它们让智能体能够以非线性方式组织信息,比纯文本更容易理解。
事件响应仪表盘
Cursor 中的 Datadog、Databricks 和 Sentry MCP 使我们能够与智能体一起深入分析可观测性数据,这些智能体常常能发现我们自己会忽略的洞察。在画布出现之前,智能体会在 markdown 表格中呈现时间序列数据,这很难解读,还需要额外步骤才能可视化。
现在,智能体可以在画布中创建可视化图表,将来自多个来源(包括本地调试文件)的数据整合到一张图表中。
PR 审查界面
我们正在审查比以往更大的差异对比。传统工具对所有变更一视同仁,需要我们自行判断差异的哪些部分最为重要。
借助画布,Cursor 能够将变更按逻辑分组,优先呈现你最需要审查的内容,并提供丰富的界面供你探索变更集。它甚至能为复杂的算法编写伪代码表示。
评估分析
在 Cursor,我们在修改测试框架或向产品发布新模型时,会投入大量时间研究评估结果。过去,工程师必须逐个检查请求 ID 来识别模式。我们曾考虑构建并部署一个 Web 应用来自动化这一过程,但最终选择直接在 Cursor 中通过技能将其操作化。
该技能允许智能体读取评估中的所有发布结果、对失败案例进行分组,并构建一个画布用于调查评估失败原因和聚类失败模式。这使我们能够发现此前隐藏的框架缺陷,并最近帮助我们以更少的精力在 Cursor 中发布了两个新模型。
自动研究实验
我们一直在将自动研究的思路进行改造,使智能体能够应对与客户端性能相关的复杂优化挑战。借助画布,智能体可以在运行实验时可视化其研究进展,使用户能够查看进度并了解智能体当前正在测试的假设。
提升信息带宽
近期改进如设计模式和升级后的语音输入,都是我们提升信息带宽努力的一部分。我们希望减少人机协作中的摩擦,让你能更轻松地表达意图,而不仅仅局限于纯文本。
在 Cursor 3.1 中尝试画布功能,或查阅我们的文档了解更多信息。
Cursor can now respond by creating canvases to visually represent information. This allows you to explore and interact with custom interfaces instead of reading walls of texts in chats or markdown files that can be hard to digest.
With canvases, agents can create dashboards for real-world data as well as custom interfaces with logic and interactivity tailored to your request. Agents can use them to help you review PRs, learn new libraries, or even manage other agents in Cursor. In the Agents Window, canvases are durable artifacts that live alongside your other tools like the terminal, browser, and source control.
Components as building blocks
Cursor renders canvases using a React-based UI library with first-party components like tables, boxes, diagrams, and charts. We gave agents access to existing components in Cursor like diffs and to-do lists, and we also instructed it to follow data visualization best practices.
You can create skills to teach agents how to create different kinds of canvases. For example, the Docs Canvas skill allows Cursor to generate an interactive architecture diagram of your repo.
How we use canvases at Cursor
We've found canvases particularly useful for data-intensive tasks. They allow agents to arrange information in a way that's non-linear and easier to digest than plain text.
Incident response dashboard
Datadog, Databricks, and Sentry MCPs in Cursor have enabled us to dive into observability data with agents, which often find insights that we'd miss on our own. Before canvases, the agent would represent time-series data in a markdown table, which was hard to interpret and required additional steps to visualize.
Now, the agent can create visualizations in a canvas that join data from multiple sources, including local debug files, into a single chart.
PR review interface
We are reviewing larger diffs than ever before. Traditional tools present all changes equally, requiring us to figure out what parts of the diff are most important.
With canvases, Cursor can logically group changes together, prioritize what's most important for you to review, and present a rich interface for you to explore the change set. It can even write pseudocode representations for tricky algorithms.
Eval analysis
At Cursor, we spend a lot of time investigating eval results as we make changes to our harness and release new models into the product. Previously, engineers had to inspect request IDs one at a time to identify patterns. We considered building and deploying a web app to automate this process, but instead, we operationalized it directly with a skill in Cursor.
The skill allows agents to read all of the rollouts in an eval, group failures, and build a canvas for investigating eval failures and cluster failure modes. This allows us to identify harness bugs that were hidden before, and recently helped us release two new models in Cursor with far less effort.
Autoresearch experiment
We have been adapting the ideas in autoresearch to enable agents to tackle complex optimization challenges related to the performance of our clients. With canvases, the agent can visualize its research progress while running experiments, enabling the user to check on progress and see the hypothesis the agent is currently testing.
Increasing information bandwidth
Recent improvements like Design Mode and upgraded voice input are all part of our effort to increase information bandwidth. We want to remove friction in human-agent collaboration and make it easier to express your intent beyond plain text.
Try canvases in Cursor 3.1, or learn more in our docs.