Cloudflare's Codex: How AI Agents Enforce Engineering Standards at Scale

Cloudflare has been running an internal experiment in AI-enforced engineering standards, and the numbers are hard to ignore. Over the past four months, their AI code reviewer flagged nearly a quarter of a million deviations from company standards and blocked 16,000 merges. A separate spec reviewer evaluated close to 600 technical designs before implementation began. Both systems draw from the Cloudflare Codex, a shared source of engineering guidance built for both humans and agents.

Before the Codex, developer guidance at Cloudflare lived in scattered places: formal docs, repo files, chat threads, and the heads of senior engineers. Finding the right standard often meant spelunking through outdated wikis or pinging someone who might know. As the company grew, that model broke down. No one could read every standard, and reviewers couldn't reliably check every requirement.

Codex: A governed RFC pipeline

The Codex is organized into domains—architecture, security, reliability, specific languages like TypeScript and Rust—each with an owner responsible for content quality. Standards are written as RFCs using RFC 2119's SHOULD and MUST keywords. Any employee can propose an RFC via merge request; it goes through rounds of feedback before domain owner approval. Approved RFCs become part of the Codex and are published to an internal Astro-powered site.

But approval isn't enforcement. An RFC must be explicitly promoted from approved to enforced before its MUST requirements can block merges. This two-stage process gives teams time to absorb new requirements and accommodates cases where enforcement needs extra work.

Feeding the LLM without blowing the context window

With 60+ RFCs and growing, feeding the entire Codex to an LLM would stress context windows and degrade results. So Cloudflare built an agent that extracts and compacts SHOULD and MUST statements into a structured JSON format, enriched with metadata for lazy discovery and progressive disclosure. Each statement gets a stable slug identifier that persists across RFC updates, enabling tracking across systems.

The JSON format includes the RFC number, domain, statement level, and a link back to the full RFC section. This lets agents filter precisely without loading entire documents.

Three agents, one Codex

The AI code reviewer evaluates merge requests against Codex compliance. It retrieves relevant RFCs, parses statements, and decides based on SHOULD vs. MUST and RFC status. Since inception, it's flagged ~230,000 violations, with ~16,000 causing approval to be withheld.

To address latency complaints, Cloudflare added two alternatives: custom linter configuration packages for language-specific requirements (TypeScript first, Rust in development, Go later) that surface problems in milliseconds, and a CLI that runs the same OpenCode-based agents locally, cutting out the CI round trip.

The spec reviewer evaluates design documents before implementation. It runs as a Cloudflare Worker, stores state in D1, routes model requests through AI Gateway, and scans for new specs via Cron Trigger. Since May 2026, it's reviewed ~600 unique specs across 3,200+ invocations. Findings are rated by severity: 65% major, 29% minor, 6% critical.

The incident report reviewer applies the same pattern to postmortems, checking completeness and quality of incident explanations, contributing factors, resolutions, and follow-up actions.

The takeaway

Cloudflare's Codex is a serious attempt to make engineering standards executable—not just documented. The key insight is the separation between approval and enforcement, and the structured extraction that keeps LLM context windows sane. The linter approach for mechanical checks is pragmatic; the CLI option respects engineer preferences. This is a blueprint other large orgs will likely crib from.