An agent fleet needs a new kind of OS, not a bigger harness

The piece argues that individual agent deployments need a harness - a layer that wraps a model and provides context, tools, execution control, orchestration, verification, observability, and governance - because models alone cannot safely assert facts about the world. Those seven authorities each own evidence the model cannot provide, and a harness functions as the control plane for a single run: what the run can see, call, do, and claim. That arrangement suffices when one agent is working in isolation, because the harness mediates model behavior and bounds damage within that run.

When many agents share enterprise data, scarcity and contention change the problem: shared memory requires protection and management; tool access becomes a system-call surface governed by capabilities; each agent needs an isolated process and a global scheduler; logs must join into a single, ordered record; and governance must sit outside any run in supervisor mode. Two additional fleet facts - accounting (whose budget was spent) and durable agent identity - exist only between agents and so cannot live in per-run harnesses. Because a harness is necessarily inside runs and therefore rewriteable by them, it cannot become the kernel-like authority the fleet needs. The solution is an operating system for agents: an external, non-rewriteable supervisor that owns contention resolution, identity, accounting, and fleet-wide policy.