Six specialized AI agents run my business over a shared protocol, with real governance underneath. Not a demo, not a framework I downloaded: an operating system I designed, that operates a real company every day.
A one-person company dies by coordination overhead. Every AI chat starts from zero, so the human becomes the memory: re-briefing context, copy-pasting state between tools, noticing what's stale, chasing dropped threads. AI makes the work faster but makes the coordination worse, and nothing stops a fleet of helpful agents from quietly drifting, duplicating, or doing something irreversible. I wanted the opposite: one human above the loop, many agents in it, and a system that remembers so I don't have to.
An operating system made of plain files and git. One kernel file every agent reads at boot: standing orders, the roster, the current state, the kill switch. Six agent specs with defined lanes and authority tiers. A shared append-only channel where agents hand work to each other instead of routing everything through me. A git-native task board. And a governance layer the agents cannot rewrite. Spawn any agent from any AI tool by pointing it at the kernel: it wakes up already knowing everything, because the memory lives in the system, not the chat window. Built as a working implementation of the ExO 3.0 framework (Salim Ismail / OpenExO).
The hardest problem wasn't intelligence, it was entropy: the moment several agents touch shared files, drift is the default. The answer became a standing rule in the kernel ("fix problems you see, don't walk past them") plus a dedicated ops agent whose whole job is catching rot. Second lesson: safety has to be structural, not polite. The governance docs are protected by a pre-commit guard the agents can't lift, because a kill switch the agent can rewrite is not a kill switch. Third: one agent with good context beats five agents in parallel for almost everything; multi-agent fan-out is reserved for read-side research, never dependent write work. And through all of it, the discipline that made the system real instead of grand-and-unfinished: define "done" before building, and ship the smallest version that runs.