How we’re rethinking work at Cloudflare with Cloudflare OS

Summary

Cloudflare’s Chief Information Officer Sam Rhea recounts how rapid advances in agentic AI prompted hundreds of employees to request direct access to production systems, exposing the limits of the company’s earlier cautious rollout. In response, the company defined five operating principles before building any tooling: AI must serve defined customer jobs rather than be adopted for its own sake; every employee, not only engineers, should receive usable capabilities; humans remain accountable for all outputs and deployed agents; organizational context outweighs raw model quality; and no user or agent may obtain elevated permissions through AI interfaces.

These rules shaped two parallel tracks. For engineers, Cloudflare assembled an authoritative Engineering Codex that encodes domain-specific standards and review criteria. Agents now apply the Codex to planning, merge-request reviews, design documents, and incident reports, resulting in hundreds of thousands of flagged issues and thousands of blocked merges. For non-engineering teams, an initial “magic email alias” staffed by humans collected repeated workflows, which were later codified into reusable skill files and context layers rather than generic code workspaces.

The resulting internal platform, Cloudflare OS, runs these skills inside ephemeral containers on Cloudflare infrastructure. Users authenticate through Zero Trust, invoke pre-built workflows through a browser, and obtain scoped connections to systems of record via custom Model Context Protocol servers. All inference passes through the AI Gateway for logging and Data Loss Prevention inspection, while Workers enforce rate limits and permission boundaries that mirror each user’s existing access rights. The approach keeps data exposure minimal, preserves audit visibility, and allows teams to automate routine work without granting agents broader privileges than their owners possess.

Why it matters

Directly addresses AI security risks, privacy controls and compliance patterns that Dutch security teams can adapt under GDPR and the EU AI Act. Provides concrete, actionable guidance on Zero Trust integration, permission scoping and auditability for AI agents.