# Replit Cloud Agents 开放企业代码库贡献,非技术人员可直接提交经测试的 PR
> 原标题:Anyone can now bring their ideas into the company codebase
- 来源:Replit:Blog
- 发布时间:2026-10-27T15:30:00.000Z
- AIHOT:https://aihot.news/items/qoz4pwew5rsp1idirgbhtpj9d
- 原文:https://replit.com/blog/anyone-can-bring-ideas-to-company-codebase
## 摘要
Replit 推出 Cloud Agents,面向 Pro 与 Enterprise 客户开放,让设计、客服等非工程人员用自然语言向公司核心代码库提交经测试的 pull request。
## 推荐理由
官方介绍 Cloud Agents 如何让非工程人员以经测试的 PR 参与公司代码库,含沙盒测试、skills 沉淀与共享环境等机制。
## 正文 · 中文译文
自2024年推出首个正式可用的编码智能体以来,几乎任何人都能构建应用。但改变公司的核心产品,对工程团队之外的人来说仍然遥不可及,而随着越来越多的人参与构建,这一瓶颈只会愈发严重。
知道产品该改什么的人,往往身处工程团队之外。设计师注意到某处间距已经不对好几个月了。支持团队负责人清楚哪个设置每周都让客户困惑。这些修复往往被写进工单,然后等待一个永远不会到来的迭代。
但当他们尝试自己动手修改时,事情就卡住了。他们无法向工程团队证明改动确实有效。工程师根本没有时间教他们——以及公司里其他所有人——团队是如何写代码的。而让应用在正确的环境中跑起来,本身就是一个独立项目。
今天,**你团队中的任何人都可以使用 Replit Cloud Agents,直接、安全地为你的核心代码库贡献代码**。Cloud Agents 会在运行中的应用里测试每一项改动,将评审反馈持续传递下去,并在整个团队间共享同一套环境配置。工程师始终掌握控制权,而其他人的贡献门槛则被降低了。
我们将这一部分称为[自驱动公司](https://replit.com/blog/rise-of-self-driving-company),现已面向所有 Pro 与 Enterprise 客户开放。
[预留席位](https://replit.com/discover/self-driving-company-webinar),在11月5日的网络研讨会上直接听取我们产品团队的分享
## 借助 Cloud Agents,将想法变成经过评审的代码改动
领域专家不再需要知道代码在哪里,或应用是如何构建的。现在任何人都可以用自然语言描述应该改什么。随后 Cloud Agents 会找到相关代码,进行安全的修改,同时应用的实时预览会在聊天窗口旁同步更新。你可以对所见内容做出反应,并在任何内容提交给工程团队之前要求调整。

当效果看起来没问题时,智能体会在 GitHub 中发起拉取请求(PR),或在 GitLab 中发起合并请求(MR),并遵循你团队的技能与贡献规范。如果评审方——无论是人还是智能体——要求修改,智能体会完成修改并更新请求。非技术专家不再需要自己把评审意见转化为代码。
## 功能请求现在以经过测试的改动形式送达,并附带视频证据
如今,当领域专家尝试交付改动时,评审很快就会变成一个来回拉锯的瓶颈。他们无法自己运行应用,于是请工程师截图,发现哪里不对,再要求修复。每一轮评审都耗时且徒增繁琐工作。
现在,Cloud Agents 可以在拉取请求送达任何人之前,先在隔离的云沙箱中检查改动。智能体会在浏览器中打开你运行中的应用,像客户一样点击走查这项改动。如果出现问题,它会修复代码并重新测试。然后,它将最终版本的截图和视频附加到拉取请求中。
这意味着,一个功能请求现在可以以经过测试的拉取请求形式送达。任何人都可以描述改动,合上笔记本电脑,回来时看到它正常运行的视频。工程师收到的拉取请求无需自己运行就能理解,它们已经过测试、带着改动正常运行的视频抵达 GitHub 或 GitLab,随时可以批准或提供反馈。
## 评审反馈变成由工程师批准的技能
贡献者越多,同一位评审者收到的拉取请求就越多,评审者最终不得不一遍又一遍地重复自己。
Replit 拥有**自我改进的工作区记忆**,可将评论转化为 Cloud Agents 遵循的指导。当评审反馈值得保留时,智能体会提议更新你团队的技能(即智能体在代码库中遵循的书面指令),或起草一份新的技能。每次更新都需工程师批准后才会生效。
此后,下一个拉取请求将从上次评审结束的地方开始。Cloud Agents 开始消除团队之间的协作负担,使双方都能更具战略性。该功能处于早期访问阶段。

更广泛的工作区记忆是 Replit 的一项关键功能,也将成为 Self-Driving Companies 的关键要素。在 Replit 内部,我们已用它减少了 51% 的 token 用量,并将智能体完成一项常见数据任务的时间从 16 分钟缩短到 1 分钟,且未降低输出质量。
## 轻松访问代码,做出更好的决策
对许多公司而言,代码是服务如何运作的基本事实之一。充分理解你的产品,你能做的事情非常多。借助 Replit,公司里的任何人都能获得基于实际构建内容的答案。支持负责人可以查明客户为何反复遇到错误。产品营销人员可以核实某项功能是否与发布文案所述一致。销售负责人可以在提议变更前,了解哪个团队最熟悉某项功能。
这也为每天回答这些问题的工程师和产品负责人节省了时间。
Replit 面向公司代码库的智能体已于 2026 年 3 月初推出。访问代码是 2026 年 Replit 内部在 Replit 上完成的工作量增加的主要驱动因素之一:

## 一次配置环境,人人皆可贡献
让公司的应用运行起来需要正确的仓库、运行时版本、数据库、后台服务和凭据。在此之前,每位新贡献者都必须重复这一配置过程,这对非技术用户来说是一项艰巨的任务。他们因此望而却步,转而单独构建,或基于过时的代码构建,结果要么让工程师去协调这些工作,要么在原型与现实之间留下巨大鸿沟。
Replit 拥有**共享环境。**管理员可以完成配置,所有人都能从那里开始。一切开箱即用。Cloud Agent 会探索你的代码库,并配置一台运行你应用的初始虚拟机(VM)。最终得到一个即用型环境,包含你的仓库、运行时、技能、工具和依赖项,任何人都可以使用。它会从你的主分支自动更新,因此每项新变更都从当前代码开始。
贡献者看不到这些。他们只需描述想要的变更,智能体就会从一个已在运行的应用开始。对工程师而言,环境可处理真实应用所需的一切:数十种编程语言、多个仓库协同运行的微服务架构、Docker 镜像、数据库、队列和缓存。凭据通过凭据代理从你的[外部密钥管理器](https://docs.replit.com/build/code/adding-secrets-to-environments)获取。
## 在团队已有的协作场所发起变更
大多数产品想法始于一次对话。客户在支持渠道报告 bug,或产品经理在讨论串中提出修复建议。但必须有人将其复制到工单中,而上下文却留在了原处。
**现在,对话串本身就能变成请求。**在 Slack 中提及或私信 @Replit,即可调查问题或将提案转化为拉取请求。你可以在对话串中跟进进度并给出反馈,需要时再回到 Replit 继续操作。最终产出的是同样经过测试的拉取请求,附带视频,位于 GitHub 或 Gitlab 中。
Microsoft Teams 支持即将推出。当你不在工位时灵光一现,也可以从 Replit 移动应用发起拉取请求。
## 这对你的团队意味着什么
你的代码库决定了产品如何运作,以及客户如何体验它。借助 Replit 的 Cloud Agents,工程团队仍然决定什么可以发布,但最贴近问题的设计师、产品经理、支持负责人和运营人员可以以经过测试的拉取请求的形式提出变更,而不是提交工单。
面向现有代码仓库的 Replit Cloud Agents 现已面向所有 Pro & Enterprise 客户开放。在 [replit.com/code-repos](http://replit.com/code-repos) 连接代码仓库,或与[销售团队](https://replit.com/enterprise)洽谈在整个组织内推广。文档见[此处](http://docs.replit.com/build/code/code-with-replit)。
Slack 集成面向所有客户开放。要进行设置,请前往 Workspace Settings,关联你的 Replit 账户,并选择一个工作区。文档见[此处](https://docs.replit.com/features/platforms/slack)。
11 月 5 日,我们将举办一场关于如何开始使用这两项功能的网络研讨会。\[[预留席位](https://replit.com/discover/self-driving-company-webinar)\]。
## 正文 · 原文
Since we launched the first generally available coding agent in 2024, almost anyone can build an app. But changing your company’s core product is still out of reach for anyone outside engineering, and that bottleneck has only grown as more people build.
The people who know what should change in a product often sit outside engineering. A designer has noticed spacing that has been off for months. A support lead knows which setting confuses customers every week. These are fixes that often go into a ticket and wait for a sprint that never comes.
But when they try to make the change for themselves, it breaks down. They can’t show engineering that the change works. The engineer simply doesn’t have time to teach them, and everyone else in the company, how the team writes code. And getting the app running in the right environment is a project of its own.
Today, **anyone on your team can use Replit Cloud Agents to contribute code directly and safely to your core codebase**. Cloud Agents test each change in your running app, carry review feedback forward, and share one setup across the team. Engineers stay in control, and the barriers to contributing are lowered for everyone else.
We call this part of the [Self-Driving Company](https://replit.com/blog/rise-of-self-driving-company), and it’s now available for all Pro & Enterprise customers.
[Save a seat](https://replit.com/discover/self-driving-company-webinar) to hear directly from our product team in a webinar on November 5
## Turn an idea into a reviewed code change with Cloud Agents
Subject matter experts no longer need to know where the code lives, or how the app is built. Anyone can now just describe what should change in plain language. Cloud Agents then find the relevant code, and make a secure change while a live preview of your app updates next to the chat. You can react to what you see, and ask for adjustments before anything goes to engineering.

When it looks right, the agent opens a pull request (PR) in GitHub or a merge request (MR) in GitLab that follows your team’s skills and contribution guidelines. If a review, human or agent, asks for changes, the agent makes them and updates the request. Non-technical experts no longer have to turn review comments into code themselves.
## A feature request now arrives as a test change, with video evidence
When subject matter experts try to ship changes today, review quickly becomes a bottleneck that goes back-and-forth. They can’t run the app themselves, so they ask an engineer for a screenshot, spot something off, and ask for a fix. Each round of review takes time and adds busywork.
Now, Cloud Agents can check the change in isolated cloud sandboxes before the pull request reaches anyone. The agent opens your running app in a browser and clicks through the change the way a customer would. If something fails, it fixes the code and tests again. Then, it attaches screenshots and video of the final version to the pull request.
What this means is that a feature request can now arrive as a tested pull request. Anyone can describe the change, close their laptop, and come back to a video of it working. Engineers get pull requests they don’t have to run themselves to understand, arriving in GitHub or GitLab already tested, with the video of the change working, ready to approve or provide feedback.
## Review feedback becomes skills engineers approve
More contributors means more pull requests for the same reviewers, and reviewers end up repeating themselves over and over again.
Replit has **self-improving workspace memory**, turning comments into guidance for Cloud Agents to follow. When review feedback is worth keeping, the agent proposes an update to your team’s skills, the written instructions agents follow in your codebase, or drafts a new one. Engineers approve each update before it takes effect.
After that, the next pull request starts where the last review left off. Cloud Agents start to remove the collaboration burden between teams, enabling both to be more strategic. This feature is in early access.

Broader workspace memory is a key function of Replit, and will be a key element of Self-Driving Companies. Internally at Replit, we’ve used it to reduce token usage by 51%, and cut the time agents take to complete a common data task from 16 minutes to 1 minute, without reducing output quality.
## Make better decisions with easy access to code
For many companies, code is one of the ground truths for how services work. There’s so much you can do with a full understanding of your product. With Replit, anyone in your company can now get answers grounded in what’s actually built. A support lead can find out why a customer keeps hitting an error. A product marketer can check that a feature works the way the launch copy says. A sales lead can see which team knows a feature best before proposing a change.
That also gives time back to the engineers and product leads who answer these questions every day.
Replit’s agent on the company codebase was rolled out at the start of March 2026. Access to code is one of the main drivers of increased work at Replit done on Replit in 2026:

## Set up the environment once, and everyone can contribute
Getting a company’s app up and running takes the right repos, runtime versions, databases, background services, and credentials. Until now, every new contributor had to repeat that setup, a daunting task for a non-technical user. Deterred, they would build separately, or on outdated code, leaving engineers to reconcile the work, or creating a yawning gap between prototype and reality.
Replit has **shared environments.** An admin can set it up, and everyone can start from there. It just works. A Cloud Agent explores your codebase and configures an initial virtual machine (VM) that runs your app. The result is a ready-to-use environment with your repos, runtimes, skills, tools, and dependencies that anyone can use. It updates automatically from your main branch, so every new change starts from current code.
Contributors don’t see any of this. They just describe the change they want, and the agent starts from an app that’s already running. For engineers, environments handle what real apps need: dozens of programming languages, microservice architectures with several repos running together, Docker images, databases, queues, and caches. Credentials come from your [external secrets managers](https://docs.replit.com/build/code/adding-secrets-to-environments) through a credential broker.
## Start changes where your team already collaborates
Most product ideas start in a conversation. A customer reports a bug in a support channel, or a PM suggests a fix in a thread. But someone has to copy that into a ticket, and the context stays behind.
**Now, the thread can become the request.** Mention or DM @Replit in Slack to investigate an issue or turn a proposal into a pull request. You can follow progress and give feedback in the thread, then continue in Replit when you need to. The result is the same tested pull request, with video, in GitHub or Gitlab.
Microsoft Teams support is coming soon. You can also start a pull request from the Replit mobile app when an idea comes up away from your desk.
## What it means for your team
Your codebase shapes how your product works and how customers experience it. With Replit’s Cloud Agents, engineering still decides what ships, but the designers, PMs, support leads, and operators closest to the problem can propose changes as tested pull requests instead of tickets.
Replit Cloud Agents for existing repositories are now available for all Pro & Enterprise customers. Connect a repository at [replit.com/code-repos](http://replit.com/code-repos), or talk to the [sales team](https://replit.com/enterprise) about rolling it out across your organization. Documentation is [here](http://docs.replit.com/build/code/code-with-replit).
The Slack integration is available to all customers. To set it up, go to Workspace Settings, link your Replit account, and choose a workspace. Documentation is [here](https://docs.replit.com/features/platforms/slack).
On November 5, we’re hosting a webinar on getting started with both. \[[Save a seat](https://replit.com/discover/self-driving-company-webinar)\].