Agentic OS throughput, in 2026.

The essay

Agentic OS throughput, in 2026.

The Netwit Agentic OS runtime scales horizontally, with audit trail as the load-bearing primitive. This is a VZU essay on the architecture, the throughput, and the cost curve of the runtime in 2026.

The architecture is a control plane and a data plane. The control plane is the orchestrator: it plans, dispatches, observes, and replans. The data plane is the workers: the agents that call the tools. The control plane is small. The data plane is large. The control plane is the bottleneck. The bottleneck is the audit trail.

The audit trail is load-bearing because every decision is logged. Every prompt. Every model call. Every tool call. Every response. Every human approval. The audit trail is not a side-effect of the runtime — it is the runtime. The runtime is the audit trail with an agent loop wrapped around it.

The throughput is a function of the audit trail's write throughput, not the model's inference throughput. The model can take 200ms-2s per call. The audit trail can take 1-10ms per write. The runtime is a streaming system, not a batch system. The streaming is what lets the runtime handle concurrent workflows without losing the audit trail.

The cost curve is dominated by the model inference, not the runtime. The runtime is a small fixed cost per workflow. The model is a variable cost per call. The cost math is: model cost + runtime cost + audit trail storage cost. The first is the operator's choice of model. The second is fixed. The third is a function of retention policy. The total cost is roughly model × 1.2 for a typical workload.

The honest disclosure: the Netwit runtime is in production at a small number of operators. The throughput numbers in this essay are based on the runtime's design, not on a published benchmark. The operators running the runtime in production have not disclosed their throughput numbers publicly. VZU will publish a benchmark when there is enough production data to make the numbers meaningful.