“Human in the loop” is a phrase. This is a mechanism.
Every vendor promises humans stay in the loop, and the promise usually means an approval button. Here is what governance actually looks like when agents operate a delivery system, stated as the artifacts you can inspect rather than the reassurance you are asked to accept.
A published decision-rights table
Every decision the system touches, from admitting an increment to committing a quarter, belongs to a named human, published where everyone can read it. Not a role category, a name. When a skill arrives with a drafted next action, the table says whose action it is.
Batched write-backs, applied on approval
The skills propose changes; they do not make them. Every change to your system of record lands as a reviewed batch, applied only after the accountable human approves it, attributable to the engagement that produced it. Nothing in your systems changes silently.
Monitors that write nothing
The skills that watch your portfolio hold the strictest rule in the system: read everything, write nothing, ever. What they produce is evidence and drafted actions for humans to send.
A ratified threshold register
The monitors judge the portfolio against thresholds your leadership ratified, not defaults a tool shipped with. When the daily read says something needs attention, the standard it applied is one your organization chose.
Facilitate and propose; humans decide.
The design principle underneath all four artifacts is the one the whole system is built on: agents assemble the evidence, draft the next action, and stop. This is not caution slowing the system down; it is what makes the speed safe to use. An organization can let skills run its discovery, its ordering, and its planning events precisely because it can see, at every step, where the human authority sits and what changed with whose approval.