Approval gates: where humans should stay in the loop
"Human in the loop" often means a person clicking approve on everything, which quickly becomes clicking approve on nothing in particular. Good approval design is narrower and stronger.
Gate the boundary, not the thinking
Let agents research, draft and plan freely. Put gates where actions leave the sandbox:
- Spending or moving money
- Sending messages to customers or the public
- Deploying or changing production systems
- Deleting or exporting data
Policy, not model judgment
An agent shouldn't decide on its own that a payment is "probably fine." Use explicit rules, such as allowlists, spending limits and named approvers, checked by code outside the model.
Make approvals cheap to give and easy to read
A good approval request says what will happen, why, what evidence supports it, and what happens if it's declined. If an approver can't understand it in thirty seconds, the gate will turn into a rubber stamp.
Record every approval
An approval that isn't recorded didn't happen, as far as anyone reviewing it later can tell. Each yes or no belongs in the decision record, next to the evidence it was based on.