How CloudTalk built CodeRabbit into its multi-agent development workflow

CloudTalk’s internal AI agents were helping the company build features and fix bugs faster. A request in Slack could become a code change, a pull request, and a working development environment. As agents took on more implementation, CloudTalk needed reliable review to check their work against its architectural and security requirements.

The company wanted developers to receive feedback locally and its agents to act on review findings before handing a change to a developer for approval.

After evaluating several AI code review tools, CloudTalk chose CodeRabbit. Its CLI and VS Code extension brought review into local development, while its pull request reviews gave CloudTalk’s agents instructions they could use to fix issues.

CloudTalk used local checks and configurable PR review rounds to address findings earlier and manage the back-and-forth.

CodeRabbit also caught a missing configuration annotation that could have made an internal endpoint publicly reachable.

“So it definitely speeds up the code review process.”

Filip Prošovský, Head of infrastructure, CloudTalk

Company snapshot

CloudTalk provides business calling software for small and midsize businesses and mid-market customers. It operates its own voice infrastructure, giving its engineers control over call performance and latency as they develop calling features and AI products.

Its engineering environment includes a Go monorepo, individual service repositories, and infrastructure repositories containing Terraform modules and GitOps configuration.

CloudTalk has also built internal agents for development, support, and other operations. Its engineering team connects those agents to the tools and documentation they need to carry out tasks, including creating features and fixing bugs.

Choosing a reviewer its developers and agents could use

CloudTalk evaluated code review tools against the needs of both its developers and its internal agents. Engineers wanted to check work before pushing it. Agents needed feedback they could use to make corrections. The company also wanted to supply its own architectural guidance and control review spending.

Prošovský said CodeRabbit stood out because “it matched our setup with the AI agents we built internally.”

CodeRabbit’s fix prompts gave those agents instructions they could evaluate and use to address findings, while the CLI let them check code before opening a pull request.

From a Slack request to a reviewed implementation

CloudTalk’s bug and feature agents can be invoked through Slack or Jira. They gather information from the company’s architecture repository and the relevant service, determine how to approach the task, and create a pull request.

CodeRabbit reviews the proposed change. The internal agent waits for the review to finish, then evaluates the findings. It can implement a suggested fix using CodeRabbit’s guidance, defer a question to a human, or determine that a proposed change is unnecessary for the task.

The exchange can run through three or four rounds before the agent summarizes the work and hands it to a developer for review.

CloudTalk’s agents can also provision an isolated development environment, deploy the change, and provide a URL for someone to inspect the result. A designer can request a button through Slack and see it running while the proposed code goes through review.

Catching potential exposure of an internal endpoint

Prošovský described an internal endpoint that needed to be routed through Cloudflare so CloudTalk’s VPN access restriction could be enforced. That routing depended on a configuration annotation.

The coding agent missed the annotation. The omission could have made the endpoint publicly reachable outside its intended protection. CodeRabbit flagged the issue during review.

Prošovský said useful catches were a regular part of the experience.

“Oh, I think like every week we find something like, okay, this was a good catch.”

Filip Prošovský, head of infrastructure, CloudTalk

Reviewing changes against CloudTalk’s architecture

Reviewing a change against CloudTalk’s requirements takes context about the service and its place in the wider system. CloudTalk supplies that context through shared architecture documentation and guidance in individual repositories.

A shared architecture repository describes the services, their connections, the infrastructure they use, and the events they send. Individual service repositories provide more detailed guidance about internal behavior, including how endpoints should be created.

Linked repositories make the shared documentation available during CodeRabbit reviews. Path instructions direct the reviewer to relevant context files.

This supports CloudTalk’s context-driven development initiative. Its agents have guidance for implementing changes, and CodeRabbit can use the same architectural and service-level documentation during review.

Catching issues while developers are still working locally

CloudTalk wanted that feedback early enough for developers and agents to act on it before opening a pull request.

In the previous review setup, developers had described repeated rounds of PR comments. They would address findings, commit changes, and receive additional feedback. The team wanted to catch more issues locally and better manage the subsequent review rounds.

Explaining the appeal of CodeRabbit’s CLI, Prošovský said “every developer likes to first check the stuff locally and then push it.”

The CLI and VS Code extension give developers that option. CloudTalk also connected the CLI to its own Model Context Protocol tooling so internal agents could run local reviews before pushing their work.

CloudTalk uses CodeRabbit’s assertive mode to obtain a broader set of findings. Its agents evaluate which suggestions to implement, which to leave unchanged, and which to refer to a developer.

Managing review access and spending

Repeated agent activity also made cost control part of CloudTalk’s evaluation.

Prošovský said earlier experience with usage-based AI tools had shown how individual developers could consume substantial portions of an organization’s budget. CloudTalk valued CodeRabbit’s seat-based setup and the ability to monitor review activity within its plan.

Per-user controls let the team select which contributors received reviews, including in repositories where occasional contributors also made changes. Configuring how many commits were reviewed gave CloudTalk another way to manage usage as developers and agents iterated.

Before and with CodeRabbit

Bringing CodeRabbit into the work of building software

CloudTalk has incorporated CodeRabbit into local development and pull request review. Developers can check code before pushing it, and internal agents can use review findings to make fixes before a developer assesses the result.

That gives developers more to work with when a change reaches them. The implementation can already have gone through several rounds of review and correction, with an isolated environment available to test the result. Developers can examine the code, understand how findings were addressed, and assess whether the change meets the original request.

The missing endpoint annotation shows the value of that process. CodeRabbit identified a configuration issue with potential security consequences in code produced by an AI agent. Its feedback gave CloudTalk a specific problem to investigate and correct.

CloudTalk built the agents and infrastructure that move an idea toward a working implementation. By incorporating CodeRabbit into that process, it also gave those agents a way to receive review, respond to findings, and bring revised work back to a developer for approval.