The harness is the OS

A model answers. A harness keeps it alive - runtime between tasks, gateway in and out, memory across sessions, tools to act, identity, skills, policy. Coding assistants cover a slice and lose context when the session ends. A harness runs unattended. Jensen called the open gateway "the operating system for personal AI." The analogy sticks. Model is CPU. Harness is OS. Memory is the filesystem. Platform vendors are shipping the hypervisor.

This is the open-source playbook on a new layer. Linux commoditized the OS and value moved up. Kubernetes commoditized orchestration. Now open weights commoditize the model and value moves into the control plane around it. OpenClaw showed breadth can draw OpenAI sponsorship, Nvidia NemoClaw, and Microsoft Scout with Entra identity. Memory-first open agents showed depth can lead daily inference without the same stage time. Foundations keep the reference honest. Vendors wrap governance. Same rhyme as Red Hat around the kernel.

Watch the absorption wave, not the star count. February: OpenClaw founder joins OpenAI, project to an independent foundation. Same month: memory-first agent runtime ships open. March: Nvidia sandboxes agents under OpenShell policy. May: memory-first leads OpenRouter daily tokens. June: OpenClaw native on Windows containers at Build. The prize for platforms is not picking one agent forever. It is sitting beneath whichever you choose - policy, identity, observability. That layer outlasts any single foundation model.

Parallels worth keeping honest. OpenClaw maps to Linux for reach - runs everywhere, explosive contributor energy, corporate absorption of key people, foundation for legitimacy. Enterprise wrappers map to Red Hat - you do not sell the kernel, you sell the auditable distribution. The harness layer maps to Kubernetes - boring control plane, standards bodies racing (MCP, agent skill portability, agentic foundations). Memory-first labs map closer to Postgres: packagers of upstream that become infrastructure operators build on, with value in state, not just execution. What did not repeat: open source rarely beat proprietary on UX. It won on control, portability, and cost at scale. Same here. Open harnesses will not out-reason the frontier APIs. They win when the buyer's question is what happens to my agent's memory if the vendor changes terms.

Near-term I expect convergence, not consolidation. Migration paths already pull gateway setups into memory runtimes. Nvidia blueprints already run more than one agent under one shell. Skills want a portable standard the way containers wanted OCI. Agent-as-PID-1 experiments will stay early - most production stays userspace harnesses for years, same as containers did not kill VMs overnight.

Stake: in two to three years the harness is boring infrastructure. Memory is the durable lock-in. Governance is where hyperscalers compete. The operator who owns keys, compounds judgment into skills, and can export what the agent learned is playing the open-source game correctly. The team that only rents seats is still on the CPU and wondering why nothing compounds.