An agent fleet needs a new kind of OS, not a bigger harness
5 hours ago
- An agent harness is a control plane for a single agent, managing one model's interaction with the world via seven authorities: context, tools, execution, orchestration, verification, observability, and governance.
- An operating system for agents addresses contention and scarcity when multiple agents share resources, with components like scheduling, memory protection, filesystem, accounting, and supervisor mode.
- Scaling from one to hundreds of agents transforms each harness authority: context becomes shared memory with permissions, tools become capabilities, execution becomes process isolation, orchestration becomes scheduling, observability becomes unified logging, and verification/governance converge into a single supervisor mode.
- Two critical facts missing from a harness but essential for a fleet are accounting (whose budget is spent) and identity (which agent acts for whom), both existing only between agents.
- A harness cannot evolve into an operating system because it lives inside the agent run, while an OS must exist outside all agents to own facts that no single agent can assert, such as fleet-wide policy compliance.