The gate as a switch you can promote

The gate from the last lesson should not be hardcoded as always ask a human. Build it as a switch with two positions, draft-only and automatic, on the same code path. In draft-only the workflow does all its work and stops for your approval. In automatic the identical workflow runs straight through. The only difference is one flag, and that flag is how a capability climbs the autonomy ladder without a rewrite every time your trust changes.

This matters because the right amount of autonomy moves over time, and you do not want a code change every time it does. On day one a new automation has earned nothing, so it sits in draft-only and you are the gate. After it has produced a stack of clean results, forcing yourself to approve the next one is friction with no safety left in it, so you promote it by flipping the flag. If it drifts, you demote it the same way. Behavior and trust move together, by configuration, not by touching the logic. The frontier makes the same point about long-running agents: the harness holds the human checkpoints, and you adjust where they sit as reliability proves out, rather than rebuilding the agent.

Why the switch is the shape of earned autonomy

Because autonomy should be a setting, not a foundation. If flipping a workflow from supervised to unsupervised means rewriting it, you will avoid the change and either stay slow forever or rip the gate out entirely, and both are bad. A clean switch keeps the safe default one decision away from the fast one, in both directions, so you can grant autonomy the moment it is earned and pull it back the moment it is not, with no risk to the workflow itself.

The takeaway: Build the gate as a draft-only to automatic switch on one code path, so promotion and demotion are a flag, not a rewrite. Autonomy is a setting that moves with your evidence, never a foundation you have to rebuild.