FIG. 04 How an agent earns the next rung

Nobody hands a new hire the keys on day one, and nobody should hand them to an agent either. Autonomy is granted per function, against cases your own team wrote, and a regression sends it back down.

L1

Assisted

The agent drafts. A person decides and sends. Every run is traced from here on, so the record starts before the trust does.

the person does the work

gate to enter

none · this is where every function starts

L2

Reviewed

The agent acts, and a person approves each action before it lands. Most of the work, a fraction of the time.

the person approves each

gate to enter

accuracy clears the bar your team set, on your team's cases

L3

Supervised

It runs on a schedule. A person samples the output and handles the exceptions it flags.

the person samples

gate to enter

sustained pass across runs · no policy breach

L4

Autonomous

It runs the function. People set the policy, watch the dashboard and hold the kill switch. Durable authority for a run with nobody in the session is the piece still in build, so today an unattended run is bounded by the tools declared for it.

the person sets policy

gate to enter

evals hold at volume · a regression demotes it

Where this honestly is. Tracing and correction capture run in production today, so the cases accumulate from your first week. The replay harness and threshold gating are in build. The order is deliberate: capture is the part you cannot backfill, and an eval set grown from six months of real corrections is worth more than one written in a week before there was anything to correct.

One function can be autonomous while the one beside it is still assisted. You never hand off what you cannot measure.