Coding Agent

|

​

Overview

Coding Agent is the part of CodeRabbit that writes code. It takes a request — typed by you, or raised by a CodeRabbit review finding, a failing check, a merge conflict, or a security finding — clones the repository into an isolated sandbox, works there, and shows you what it did. The results can be delivered as a direct commit or a stacked PR for a review. Find it in the sidebar under Code, or go directly to app.coderabbit.ai/code.

​

What a coding task is

Every agent run is a coding task: one durable unit of work bound to an organization, a repository, and either a pull request or a branch. A task has its own page, its own address, and its own history, so you can leave it, come back to it, and share it with a teammate. A task can be started from different points. A button on a pull request, a request typed in the app, and a security finding’s Fix with AI all create tasks and all land on the app.coderabbit.ai/code page, so whatever started the work, you steer it, read it, and deliver it the same way. Three properties hold for every task:

• It works in its own sandbox. The task gets a prepared checkout of your repository. It does not act on your working copy. Changes reach the remote repository through the delivery workflow you choose. What that sandbox contains is configurable — see Coding environments.
• You choose how it delivers. Manual delivery produces a proposed patch: a diff you read, comment on, revise, reject, or accept. Where available, Autopilot can publish the work and handle follow-up fixes automatically after you enable it.
• Repository permissions still apply. Manual delivery uses your Git provider authorization, and CodeRabbit is named as the commit author. Branch protections still apply to automated delivery. See Who pushes, and who wrote it.

Sending the same request twice does not start two tasks. CodeRabbit identifies a request by the organization, the requester, the repository, and the request itself, so a retried or double-clicked request resolves to the task that already exists.

​

Lifecycle of a task

Turns are also the unit of billing: a few first turns are free, and the rest draw on metered agent minutes. See Access and billing.

​

Where a task can start

See Start a coding task for what each one carries into the sandbox and which providers support it.

​

Code mode and Plan mode

Every task runs in one of two modes, chosen when you start it. Plan mode puts a human approval gate in front of implementation. Once a plan is approved, the agent implements exactly what was approved. See Review a plan.

​

Reuse a skill

Choose a skill to give the task a repeatable procedure. Browse the catalog from the composer, or create private and organization skills under Code > Skills. A skill can guide the first request or a follow-up turn in an existing task.

​

Collaboration

A task is visible to your team by default, so anyone in your organization can open it, follow its progress, and see its prompts, files, and activity. Mark a task private and only you can see it — including anyone who follows a direct link. Only the task’s author can change this, and it can be changed back at any time. Set this when you start a task, or switch it afterwards from the task page. See Control who can see a task.

​

Names that collide

“Coding agent” means several things across this site and across the industry. These distinctions are worth knowing before you read further.

​

What’s next