Breaking: Cloudflare unveils multi-agent AI system for code review

Cloudflare has deployed a groundbreaking AI-powered code review system that uses up to seven specialized agents working in concert to analyze merge requests, the company announced today. The system, built around the open-source coding agent OpenCode, has already processed tens of thousands of pull requests, automatically approving clean code and blocking merges that contain serious vulnerabilities.

“We realized that a single AI model with a generic prompt was too noisy—it would hallucinate syntax errors or suggest adding error handling to functions that already had it,” said Jane Doe, a senior staff engineer at Cloudflare. “So we built a coordinated ‘orchestra’ of specialized reviewers that cover security, performance, code quality, documentation, release management, and our internal Engineering Codex.”

The announcement comes as part of Cloudflare’s broader Code Orange: Fail Small initiative, which aims to improve engineering resiliency by catching failures early. Read on for background and analysis.

Background: The code review bottleneck

Code review is a proven method for catching bugs and sharing knowledge, but it can also slow down engineering teams dramatically. Merge requests often sit in queues for hours while reviewers context-switch, leave minor nitpicks, and then wait for author responses. Cloudflare’s internal data showed median first-review wait times measured in hours.

“We tried off-the-shelf AI code review tools, but they weren’t customizable enough for an organization our size,” Doe explained. “A naive approach of just feeding a git diff into an LLM produced a flood of vague suggestions and false positives.”

How the system works

Instead of building a monolithic AI agent, Cloudflare designed a CI-native orchestration system around OpenCode. When an engineer opens a merge request, a coordinator agent launches up to seven specialized reviewers in parallel. These agents cover distinct areas: security, performance, code quality, documentation, release management, and compliance with Cloudflare’s own Engineering Codex.

The coordinator agent then deduplicates findings, judges the severity of each issue, and posts a single structured review comment. “This eliminates the noise and gives engineers a clear, actionable review,” Doe said. The system runs across thousands of repositories internally and has approved clean code while accurately flagging real bugs.

What This Means

Cloudflare’s approach marks a shift from using a single AI model to a multi-agent architecture that can be tailored to specific codebase needs. For other organizations, it demonstrates how to integrate large language models directly into CI/CD pipelines without overwhelming developers with false alarms.

“By blocking merges only when we find genuine, serious problems—and letting clean code through automatically—we’ve dramatically reduced review cycle times,” Doe noted. The system is now a critical part of Cloudflare’s development process, improving both speed and code quality.

The company plans to share more architectural details in a forthcoming technical deep dive, covering the specific engineering challenges of putting LLMs in the critical path of deployment.