I designed and built Cloudex Technologies as an AI-native company. It's staffed by 18 "Digital FTEs": AI agents with written roles, real tools and a reporting line. They sit in four departments: the Office of the CEO, Delivery, Inbound & Demand, and Outbound & Sales. They run on Paperclip with Claude Code as the main lane and Codex as an adversarial reviewer. The whole company is defined in a git repo (constitution, governance, per-agent instructions), so the runtime is a deploy, not a pet.

The rule it's built on: author ≠ verifier ≠ releaser. No agent approves its own work, every build passes five quality gates, and a human principal releases every outbound batch, which can be done by replying SEND or HOLD to an email digest. The sales department runs its own AutoGTM engine. It sources leads from the free Companies House register and researches each prospect's website, keeping only facts quoted word for word. It then drafts short, human emails and sends them within warm-up limits, with auto opt-outs and bounce stops. It runs privately on a VPS over a Tailscale mesh, with budget caps and a full audit trail.