Humans hold the judgment layer

Three decisions stay with people: what to build, what ships, and what is true. Everything below that line gets drafted and executed by agents, at whatever volume can actually be reviewed. The real work is figuring out which decisions need human judgment at all. Most things that feel like judgment are mechanical checks wearing a suit.

Decision rights

The dividing line is reversibility, not difficulty. If being wrong needs an apology rather than a later commit, the call stays with a person. If being wrong needs a fix and a note, an agent can make it and a person can check it. Hard-but-reversible goes to the machine. Easy-but-irreversible stays human.

Memory and context

Human attention is the scarcest resource in the system. Information arrives in a form a reviewer can read, so their effort goes on the decision and not on reconstructing what happened. Change summaries are computed automatically and annotated by a person, not written from scratch. Decision history is findable: corrections append, nothing gets silently rewritten, so yesterday's call cannot be quietly treated differently today.

Execution

The advantage is AI volume below the review line, and it only works if the truly mechanical checks run first: schema, references, required fields, format. The machine runs first and the human sees only what passed. Reverse that order and you spend expensive reviewer time on cheap checks, and the reviewer stops reading.

Governance

The approval gate sits at merge, the last irreversible step. Two ways judgment leaks out: approving so many diffs a day that signing replaces thinking, and re-verifying what the automated checks already checked. Both look like diligence. Both are the judgment layer quietly switching itself off.

"Human in the loop" means nothing without four specifics: which person, at which step, deciding what, with what in front of them. A company that cannot answer those four does not have a judgment layer. It has a phrase.

Observability

Review time per unit of output should fall while the rejection rate holds. A gate that never rejects is not a gate. If review velocity is climbing and the submissions have not changed, review has become a formality, and formalities are where the irreversible mistake gets through.