Replit Cloud Agents Open Tested Pull Requests

Replit says staff outside engineering can now change their company’s core product by describing what they want in plain language. Its Cloud Agents find the code, make the change, test it in the running app and then open a pull request for engineers to review, with screenshots and video attached. The company announced the feature in a blog post on 27 October.

The pitch is aimed at designers, product managers and support leads who know what should change but cannot write the code. “Engineers stay in control,” write Alexandre Cuoci and Nick McLachlan of Replit. Nothing reaches the main codebase until a reviewer approves it, and Replit’s own documentation says that opening a pull request “does not merge or deploy it”.

In brief

Replit’s Cloud Agents turn a plain-language request into a tested GitHub pull request or GitLab merge request, with video.
Replit says the feature is available to all Pro and Enterprise customers; its docs still call the workflow early access.
A memory feature that turns review comments into rules for the agent is in early access, and engineers approve each change.

From a chat request to a pull request

The process starts in Replit or in Slack. A user describes a change, and a live preview of the app updates beside the chat so they can ask for adjustments. When it looks right, the agent opens a pull request in GitHub or a merge request in GitLab “that follows your team’s skills and contribution guidelines”, the blog says. If a reviewer, human or agent, asks for changes, the agent makes them and updates the request.

Testing happens before anyone sees the code. According to Replit, the agent runs the change in isolated cloud sandboxes, opens the app in a browser and clicks through it “the way a customer would”. If something fails, it fixes the code and tries again. It then attaches screenshots and a video of the final version, so that an engineer does not need to run the branch to understand it.

The blog’s example contributors are a designer who has noticed spacing that has been off for months, or a support lead who knows which setting confuses customers every week. Replit says such fixes often go into a ticket and “wait for a sprint that never comes”.

Shared set-up and team memory

Getting a company’s app running is often the hardest step for an outsider. Replit says an admin sets up a shared environment once, after a Cloud Agent explores the codebase and configures a virtual machine that runs the app. That environment updates automatically from the main branch, and credentials come from the company’s external secrets managers through a credential broker. Replit’s security documentation says each sandbox runs inside a microVM, a lightweight virtual machine that acts as the isolation boundary.

The second piece is memory. When review feedback is worth keeping, the agent proposes an update to the team’s “skills”, the written instructions agents follow in that codebase, or drafts a new one. Engineers approve each update before it takes effect. Replit says this feature is in early access. It also says broader workspace memory has cut its own token use by 51% and reduced a common data task from 16 minutes to one minute “without reducing output quality”. Those are Replit’s internal figures.

Users can also start a change by mentioning @Replit in a Slack thread, which the company says is available to all customers, or from its mobile app. Microsoft Teams support is “coming soon”. Replit says its own staff have used the agent on the company codebase since March 2026.

Who can use it today

Replit’s blog says Cloud Agents for existing repositories are “now available for all Pro & Enterprise customers”, connected through replit.com/repos. Its documentation is more cautious: “Repository-based development is an early-access workflow. Availability and setup can differ by workspace.” Customers checking whether they have it should expect some variation, at least at first. Replit is running a webinar on 5 November.

Engineers keep the merge button, and the review load

Replit’s design puts the safeguard in the familiar place: the pull request. That matters because the risk with letting more people change production code is not bad intentions but plausible-looking changes nobody fully checks. The company’s docs are blunt about this. “A passing test or an Agent summary does not replace a code review,” the security page says, and “isolation does not remove the permissions you grant to code running inside the sandbox.”

So the real shift is in who writes the first draft, not who decides. A video of the change working answers the first question a reviewer asks, but it says nothing about edge cases, data handling or code the agent touched that the demo never exercises. If the feature works as described, engineering teams could face many more small pull requests from colleagues who cannot explain the code they are submitting. The memory feature is Replit’s answer: turning each repeated review comment into a rule so the next request arrives cleaner. Whether that cuts review time or simply moves it will show up only in customer use, not in Replit’s internal numbers.

Replit is also competing on the infrastructure underneath. Sandboxing agents in microVMs is the same approach Google described for its Agent Compute Engine, and Cursor has gone the other way, letting developers build agents as testable code with its Agent SDK. Replit has not said which AI models power Cloud Agents, a question that drew attention this month after its chief executive said it no longer works with Anthropic, a claim an insider disputed.