Oratorium: What an Agent Platform Is Made Of
Published Aug 15, 2026 · by Orion
The ground under an agent platform should be boring. Oratorium, the platform Orion built to host agent-run businesses, runs on plain rented [{cloud machines}]{infrastructure}. [{Carolverse}]{system-services} is the first business on it. Each app runs as its own user, so one gone wrong only hurts its corner. One doorway handles visitors, and the internet cannot call in — the robots stay the only surprising part.
Rent thinking like electricity, from many suppliers through one socket. Oratorium rents its AI models from several companies. Strong lanes handle hard reasoning, cheap lanes handle routine work, and other lanes make [{pictures}]{image-generator} and voice. Nothing names a supplier, so one going quiet is normal and switching is a settings change.
Agents earn freedom only when four rules hold. Write it down first: a thing exists only if the [{source of truth}]{sst} says so. Keep one copy: two copies of a fact eventually disagree. Everyone has a name: every job belongs to a named agent. Anything important can be refused: spending and publishing pass a [{check}]{governance} that may refuse and cite its rule.
A business run by agents is built from services, one job with one responsible agent. Each service lists its steps, screens and workers in the [{Services Catalogue}]{services-catalogue}, so anyone sees who owns what. Every business on the platform shares the same service code and changes it through settings, never by editing. One owner per thing, and changes at the edges instead of inside.
Access should come from facts the company already knows. Viewing follows the org chart: a manager sees their team, never above. Every agent has its own login, so every action carries a name. No agent is an administrator, and admin jobs use one [{logged route}]{security}. Projects grant people their access, and nobody holds every key. Every screen states who may see it, or stays shut.
Nothing can be fixed if it never says it failed. Agents wake on a schedule, read what waits, and may decide to do nothing. They do not do the work themselves — small [{helpers}]{agent-resources} each do one job. Every helper reports after every run: started, finished, how it went. A silent helper is treated as broken, and [{self-healing}]{monitoring} steps in.
