Why the gate lives inside the protocol

Every company deploying agents that transact eventually builds the same slide: a diagram of the agent stack with a box labelled "human oversight" drawn next to it. The box is sincere. It is also, in most architectures, optional — a notification, a dashboard, a Slack ping that arrives after the commitment is already made.

We started Subchain from the opposite premise: if you are answerable for what your agents commit to, the approval step cannot be a courtesy. It has to be a state in the machine — one the runtime cannot skip, rush, or interpret away.

A gate beside the protocol is a suggestion. A gate inside it is a fact.

What "inside" means concretely

In Subchain, a negotiation that crosses a threshold does not produce a warning. It halts. The state becomes pending-approval, the commitment routes to a named governor, and there is no API call that moves it forward — not with any key, not with any scope. The only verb that unblocks it is a human decision in the console, recorded with that person’s identity.

This is an unfashionable design. It makes demos slower. It means a benchmark of our runtime shows a hard stop where other systems show throughput. We consider that the feature.

Speed where speed belongs

Below the thresholds, agents run at machine speed — proposing, countering, accepting against shared templates without a human in sight. The gate is not a brake on everything; it is a boundary humans draw once, in advance, about what "everything" may include.

Agents negotiate. People decide. The order of those sentences is the architecture.