How do I stop an AI agent making changes I didn't approve?

Why this matters

Free inside what you settled. A genuine crossing surfaces the choice — to you, first.

You've seen both failure modes. Approval-on-everything, where you become the slowest component in your own pipeline and the agent's speed is wasted. And open throttle, where you wake up to changes nobody authorised and spend the morning working out what moved.

Gates belong only on the load-bearing. Inside what you've settled, the agent should run free — that's the whole point of it. What makes that safe isn't surveillance; it's that your decisions and their boundaries are in the agent's reasoning before it acts, so a step that would cross one is visible as a crossing — a choice for you, not a surprise from it.

The way through

And whatever a run does, your last good decision is always intact: settled decisions are never edited and never overwritten, so the agent re-anchors on what you settled, not on its last mistake. That's the decision layer — your files are your version control's job; your decisions are ours.

With Unl you author the brief, hand it off, and the agent executes inside what you settled — full speed on the mechanical work, your decisions arriving in its reasoning before it acts on the parts that need them. And done stays yours to say.