Cloudflare OS: How Cloudflare Is Turning AI Agents into an Internal Platform

Cloudflare's CIO, Sam Rhea, describes how the company moved from a cautious AI stance to building an internal platform called Cloudflare OS. The catalyst: a sales team member asked for API keys to build a 'SuperApp' that would transform go-to-market, prompting a realization that AI agents could actually do things—and needed guardrails.

The platform, built on Cloudflare's own Workers and Access, gives employees a browser-based harness to run skill files and workflows, with security controls that ensure agents only access what they need. The post outlines five principles: AI should serve customers, everyone deserves superpowers, humans own the output, organizational context matters more than the model, and AI should never grant more permissions than the user has.

For engineers, Cloudflare built the 'Engineering Codex,' a context layer that agents use to review merge requests, designs, and incident reports. In four months, these agents flagged nearly 250,000 potential problems and blocked 16,000 merges. For non-engineers, the team started with a 'magic email alias'—a manually staffed bot that automated mundane work—which revealed common jobs to be done and informed the skills library in Cloudflare OS.

The key insight: meet users where they are. Giving everyone a coding harness leads to a flood of vibe-coded apps; instead, Cloudflare worked backwards from real pain points. The platform now runs in an ephemeral cloud environment, giving security audit visibility and network control, while users can close their laptops and let work happen.