Cursor 构建 Origin 来托管来自大批 AI 代理的代码，今日上线

Why it matters

Origin would give Cursor control of the repository layer beneath its editor, cloud agents and code review, turning a coding tool into an integrated software-production stack.

Michael Truell (@mntruell), Cursor's co-founder and CEO, introduced Origin on June 16th as a Git platform built to handle code produced by teams of AI agents. Cursor is set to launch Origin on August 11th after initially limiting the product to a work-email waitlist.

该发布将 Cursor 的覆盖范围扩展到从将工作分配给代理、到编写、测试、审查、存储和合并由此产生的代码的完整路径。Origin 为已包含以代理为中心的编辑器、云端执行、自动化拉取请求审查、工作流自动化和用于构建自定义代理的 SDK 的产品线增加了仓库托管功能。

Tomas Reimers, a co-founder of code-review developer Graphite, is leading Origin inside Cursor. Cursor agreed to acquire Graphite in December 2025 after identifying review and safe merging as the next bottlenecks created by faster code generation. Reimers had originally built Graphite with co-founders Merrill Lutsky and Greg Foster as an internal tool for keeping engineers unblocked before turning it into a code-review product.

Tomas Reimers 在 Cursor 内部负责领导 Origin。Cursor 在 2025 年 12 月同意收购 Graphite，此前发现审查和安全合并将成为更快代码生成带来的下一个瓶颈。Reimers 最初与联合创始人 Merrill Lutsky 和 Greg Foster 一起将 Graphite 打造成一个用于避免工程师受阻的内部工具，随后将其转变为代码审查产品。

At Cursor's Compile conference, Reimers presented Origin as a new Git platform for teams and agents. Coverage of the demonstration reported a Git-compatible service with agent-facing APIs, Model Context Protocol support, automated responses to merge conflicts and failed continuous-integration runs, and infrastructure designed for many agents pushing changes in parallel. Cursor's demonstration claimed throughput of 22.6 commits per second in one repository, but Cursor has not published a benchmark methodology or production results for that figure.

在 Cursor 的 Compile conference 上，Reimers 将 Origin 作为面向团队和代理的全新 Git 平台进行展示。演示报道显示这是一个兼容 Git 的服务，提供面向代理的 API、Model Context Protocol（MCP）支持、对合并冲突和失败的持续集成运行的自动化响应，以及为大量代理并行推送更改而设计的基础设施。Cursor 的演示宣称在单个仓库中的吞吐量为每秒 22.6 次提交，但 Cursor 尚未发布该数字的基准方法或生产环境结果。

That number describes repository ingest, not software quality. A forge can accept thousands of agent commits without establishing whether those changes are correct, secure or worth merging. Cursor's larger product strategy addresses that gap by placing automated review and policy controls around the agents that generate the code.

该数字描述的是仓库的摄取速率，而非软件质量。一个代码托管平台可以接受数千次代理提交，但并不等于这些更改是正确的、安全的或值得合并的。Cursor 更大的产品策略通过在生成代码的代理周围放置自动化审查和策略控制来弥补这一空白。

Origin is a Git host, not a replacement for Git

Origin 是一个 Git 托管服务，而非替代 Git

Origin is best understood as a forge: the service around Git that stores repositories and manages collaboration, pull requests, permissions, checks and merges. Git remains the underlying version-control system. Origin is positioned against the repository and collaboration layer occupied by GitHub and GitLab.

Origin 最好被理解为一个代码托管服务：围绕 Git 的服务，用于存储仓库并管理协作、拉取请求、权限、检查和合并。Git 仍然是底层的版本控制系统。Origin 的定位是在 GitHub 和 GitLab 所占据的仓库与协作层次之上。

The distinction matters for portability. Standard Git history, branches and tags can generally move between compatible hosts. Pull-request discussions, issue trackers, access policies, CI configurations and application integrations are host-specific. Any serious Origin evaluation will therefore depend on Cursor's import tools, identity controls and compatibility with existing development pipelines.

这种区分对可移植性很重要。标准的 Git 历史、分支和标签通常可以在兼容的主机之间迁移。拉取请求讨论、问题跟踪器、访问策略、CI 配置和应用集成则是特定于主机的。因此，对 Origin 的任何认真评估都将取决于 Cursor 的导入工具、身份控制以及与现有开发流水线的兼容性。

Ahead of the August 11th launch, Cursor had not published Origin's pricing, security architecture, data-handling terms or migration tooling, and the public product page continued to offer only a work-email waitlist. Cursor had said around the June announcement that Origin was expected in fall 2026, making the August launch earlier than its initial public timeline.

在 8 月 11 日 发布之前，Cursor 尚未公布 Origin 的定价、安全架构、数据处理条款或迁移工具，公开的产品页面仍然只提供工作邮箱候补名单。Cursor 在 6 月的公告附近曾表示 Origin 预计将在 2026 年秋季推出，因此 8 月的发布早于其最初的公开时间表。

What an agentic Cursor team can use today

使用代理的 Cursor 团队今天可用的功能

Origin is one layer in a stack Cursor has assembled throughout 2026. Teams do not need Origin to begin running agents in parallel, although their repositories and pull requests can remain on an existing host.

Origin 是 Cursor 在 2026 年构建的技术栈中的一层。团队并不需要 Origin 就能开始并行运行代理，尽管他们的仓库和拉取请求可以继续保留在现有主机上。

Cursor 3, released on April 2nd, introduced a workspace that puts local and cloud agents in one interface. Agents started from desktop, web, mobile, Slack, GitHub or Linear can appear together, and a task can move between a local machine and Cursor's cloud. Worktrees isolate simultaneous changes on separate branches, while multi-root workspaces let one agent modify several repositories in a single session.

Cursor 3 于 4 月 2 日 发布，推出了一个将本地和云端代理放在同一界面的工作区。从桌面、网页、移动端、Slack、GitHub 或 Linear 启动的代理可以同时出现，任务可以在本地机器与 Cursor 的云之间移动。Worktrees 将同时发生的更改隔离到不同的分支上，而多根工作区（multi-root workspaces）则允许一个代理在单次会话中修改多个仓库。

Cursor also supports asynchronous subagents. A parent agent can split work into narrower jobs with separate context, models and tool permissions. The /multitask workflow can break a larger request into parallel streams, while nested subagents can delegate further. This is Cursor's closest current implementation of an agent team: a human or parent agent decomposes the work, specialized agents execute it, and their outputs return for integration and review.

Cursor 还支持 asynchronous subagents。父代理可以将工作拆分为具有独立上下文、模型和工具权限的更小任务。/multitask 工作流可以将较大的请求拆分为并行流，嵌套的子代理可以进一步委派任务。这是 Cursor 目前最接近代理团队的实现：由人类或父代理进行工作分解，专门化代理执行任务，然后将产出返回以供集成和审查。

Cloud agents run in dedicated virtual machines with their own dependencies and network access. They can continue after a developer closes a laptop and return artifacts such as logs, screenshots and demonstrations. That changes the operator's job from watching every tool call to specifying the task, defining acceptance criteria and checking the result.

Cloud agents 在独立的虚拟机中运行，具有自己的依赖和网络访问。即使开发者关闭笔记本，它们也可以继续运行并返回诸如日志、截图和演示等产物。这将操作员的工作从监视每一次工具调用转变为指定任务、定义验收标准并检查结果。

Truell wrote in February that 35% of pull requests merged inside Cursor were created by agents running autonomously in cloud virtual machines. That is a Cursor-reported internal metric, but it explains the timing behind Origin: Cursor is already experiencing the review, environment and coordination problems that appear when agent output becomes a meaningful share of production work.

Truell 在二月写道，在 Cursor 内合并的拉取请求中有 35% 是由在云虚拟机中自主运行的代理创建的。这是 Cursor 报告的内部指标，但它解释了 Origin 推出的时机：当代理产出成为生产工作中有意义的组成部分时，Cursor 已经在经历审查、环境和协调方面的问题。

Rules, automations and custom agents form the operating layer

规则、自动化和自定义代理构成操作层

Cursor's team product turns individual prompting habits into shared infrastructure. The current Teams plan starts at $40 per user per month and includes a private team marketplace for distributing rules, skills and plugins. Teams can encode repository conventions, testing requirements and recurring workflows instead of asking every engineer to configure agents independently.

Cursor 的团队产品将个人的提示习惯转化为共享基础设施。当前的 Teams plan 起价为每位用户每月 40 美元，并包含一个用于分发规则、技能和插件的私有团队市场。团队可以对仓库约定、测试要求和经常性工作流进行编码，而无需让每位工程师独立配置代理。

Automations create always-on agents triggered by schedules or events from services including GitHub, Linear, Slack, PagerDuty and webhooks. Each run starts a cloud sandbox and follows the configured instructions, model choices and MCP connections. Appropriate jobs include triaging bug reports, investigating failed builds, updating dependencies and preparing routine pull requests.

Automations 创建始终在线的代理，这些代理由计划任务或来自包括 GitHub、Linear、Slack、PagerDuty 和 webhooks 在内的服务事件触发。每次运行都会启动一个云沙箱，并遵循已配置的指令、模型选择和 MCP 连接。适合的任务包括对错误报告进行分类、调查失败的构建、更新依赖项以及准备常规的拉取请求。

For workflows that need application logic, the public-beta Cursor SDK exposes the same agent runtime used across Cursor's desktop, CLI and web products. Developers can launch local or cloud agents from TypeScript or Python, stream their work and incorporate them into CI systems or internal tools. June updates added custom tools, configurable state stores, nested subagents and an auto-review layer that can hold risky tool calls for approval.

对于需要应用逻辑的工作流，公开测试版的 Cursor SDK 暴露了在 Cursor 的桌面、CLI 和 Web 产品中使用的相同代理运行时。开发者可以从 TypeScript 或 Python 启动本地或云端代理，流式传输它们的工作并将其整合到 CI 系统或内部工具中。6 月的更新增加了自定义工具、可配置的状态存储、嵌套子代理以及可以将风险较高的工具调用暂存以供批准的自动审查层。

The review layer is Bugbot, Cursor's usage-priced pull-request reviewer. Teams can set review effort and connect additional context through MCP. Cursor says Bugbot can learn review rules from reactions and human comments, creating a feedback loop between a team's standards and future automated reviews. Those performance and resolution figures remain Cursor's own measurements.

审查层是 Bugbot，这是 Cursor 按使用量计费的拉取请求审查工具。团队可以设定审查力度并通过 MCP 连接额外的上下文。Cursor 表示 Bugbot 可以从反应和人工评论中学习审查规则，在团队标准与未来自动化审查之间形成反馈回路。这些性能和解决率数据仍然是 Cursor 自行测量的结果。

Governance becomes the limiting factor

治理成为限制因素

Running several agents is straightforward. Giving them production credentials, network access and permission to merge code is the harder organizational decision.

运行多个代理并不复杂。真正更难的组织决策是授予它们生产凭证、网络访问权限和合并代码的权限。

Cursor Enterprise now supports organizations, teams and groups with different budgets, approved models and agent permissions. Administrators can separate experimental users from production teams, restrict repository and MCP access, apply network policies, control command execution and inspect usage. Cursor's auto-review mode also routes uncertain shell, MCP and fetch operations through a classifier before allowing execution or requesting human approval.

Cursor Enterprise 现在支持具有不同预算、批准模型和代理权限的organizations, teams and groups。管理员可以将实验性用户与生产团队分离，限制仓库和 MCP 访问，应用网络策略，控制命令执行并检查使用情况。Cursor 的自动审查模式还会在允许执行或请求人工批准之前，将不确定的 shell、MCP 和 fetch 操作通过分类器进行路由。

A workable agentic team therefore needs explicit boundaries: isolated branches or worktrees, reproducible environments, written acceptance criteria, restricted credentials, deterministic tests and human ownership of final merges. Parallel agents increase output and can also multiply flaky tests, conflicting changes and unnecessary pull requests.

因此，一个可行的代理化团队需要明确的边界：隔离的分支或 worktrees、可重现的环境、书面的验收标准、受限的凭证、确定性的测试以及对最终合并的人类所有权。并行代理会增加产出，但也可能放大不稳定的测试、冲突的更改和不必要的拉取请求。

Origin is Cursor's proposed answer to the coordination pressure created by that output. Its strategic value will depend on review quality, policy enforcement and migration support rather than raw commit throughput. The August 11th launch shifts the immediate question from when Origin will arrive to whether Cursor can provide the operating details and compatibility that engineering teams need before moving their system of record.

Origin 是 Cursor 针对由这些产出带来的协调压力所提出的解决方案。其战略价值将取决于审查质量、策略执行和迁移支持，而非原始的提交吞吐量。8 月 11 日 的发布将眼下的问题从 Origin 何时到达，转为 Cursor 是否能在工程团队迁移其记录系统之前提供所需的操作细节和兼容性。