The agent drafts, the person signs
Walk into the companies actually running agents at scale and you find something the loudest voices on both sides keep missing: almost none of them are autonomous, and the people running them are not confused about why. The agent does the ninety percent that is tedious. A human does the ten percent that is judgment, and signs. The signature is not bureaucratic residue. It is the product.
We sit between the camps and borrow from both. The builders are right that the tool is astonishing and that most fear is just unfamiliarity. The reliability engineers are right that confident error is the failure mode that bites. So we keep a hand on the wheel, not because we distrust the agent but because keeping the hand there keeps the habit of thinking alive — and the day you stop checking is the day the one bad call ships unread.
The quiet thing we have learned: the review itself makes people better at their jobs. The lawyer who reads the agent’s draft contract closely catches things she would have missed writing it cold. Delegation done well is not abdication. It is a junior partner who never tires and never sulks, supervised by someone who still owns the outcome.
Where we concede ground: A human on every step is a tax that will not scale, and we are quietly rationing where the hand stays on the wheel.
What would change our mind: Fully autonomous agents posting a lower real-world error rate than human-reviewed ones across a hard year of audited use.
Read the full synthesis: Can you trust an AI agent to act for you?