The uncomfortable question for any autonomous system is "what can it do without asking?" We think the honest answer has to be a dial, not a slogan — so the pipeline runs in three modes. Auto asks once, at publication. Semi inserts two sign-off points, one after planning and one before publishing. Manual asks at every decision. In every mode, publishing waits for a human's explicit confirmation: a button press, never something inferred from conversation.
Paused is a feature
One default we refuse to relax: ad creative always lands paused. Even in auto mode, a paid campaign arrives in the ad account switched off, waiting for a person to turn it on. Organic content is recoverable; ad spend isn't. The blast radius of a bad automated decision should be proportional to how reversible it is, and money is the least reversible thing we touch.
The hard part is waiting well
Approval gates sound like an afterthought — an if statement and a notification. The engineering is in what happens during the wait. A campaign might sit at a publish gate for an hour or a week. Meanwhile laptops close, the Runner restarts, updates land, and nothing about the run can afford to evaporate.
So the orchestrator runs as a durable state machine. The run pauses at the gate and waits; progress is checkpointed after each step. When someone finally presses Approve or Request changes — in the Slack thread, in Teams, or at the CLI — that exact paused run resumes where it stopped: same plan, same assets, same context, even across a restart.
A review never costs you the work that preceded it.
Without that property, gates create pressure to skip gates, and the whole governance story collapses into a checkbox.
Flag, annotate, keep moving
The gates themselves follow the same philosophy as our grounding checks: they flag and annotate rather than block. A brand-safety hit is logged with a severity and attached to the asset; the flagged asset waits for review while the rest of the campaign keeps producing. One questionable social post shouldn't hold forty finished assets hostage.
What lands in front of the reviewer is therefore a decision, not a mystery: the asset, its flags, its manifest, its lineage back to your sources. Approve, edit, or reject — with the evidence in front of you. In Slack that is a button on the thread; at the CLI, a few seconds at a prompt:
We're building this pre-launch because it can't be retrofitted. Autonomy earns trust in exactly one way: by making the moments it defers to you fast, informed, and impossible to lose.