The core decision was to build a control plane, not a wrapper. Operators create agents, hand them skills, watch what they do, and shape their environment from the dashboard — the runtime stays an implementation detail. That one boundary shaped everything else.

Nothing risky happens unsupervised. Every action that reaches a real outside system passes a verification agent (GateKeeper) and, when it matters, a human — and it fails safe: if the gate is down, the action is blocked, not waved through.

You can describe a team in plain English and the platform designs it for you — agents, skills, schedules, channels — as a blueprint you review and approve before anything is created. Standing up a new set of agents feels like a conversation, not a config project.

A marketplace of skills and plugins with trust built in: agents can pull the safe, everyday capabilities themselves, sensitive ones need an admin’s sign-off, and every item tracks who made it.

It’s a real operations product, not a demo: live fleet monitoring, cost tracking that separates infrastructure from model spend, full audit history, Microsoft Teams / Slack / Telegram reachability, and a monitoring agent that watches the other agents — all deployed as infrastructure-as-code.