Supervise agents the way you grow junior engineers
No lead I respect hands a new hire production credentials on day one. Autonomy is granted in scopes and earned by track record: read the codebase before touching it, every change reviewed before spot-checks, one well-fenced service before the whole system. Yet I keep watching teams grant agents broad write access on day one — then, after the first incident, revoke everything and declare agents unready. Both moves are the same mistake. The supervision ladder we use for juniors is the right model for agents, rung for rung.
The rungs even have formal names now. Human-in-the-loop: the agent proposes, a human approves before anything consequential executes — the intern whose every commit gets reviewed. Human-on-the-loop: the agent executes autonomously while a human monitors and intervenes on exceptions — the mid-level engineer with a watchful lead. Fully autonomous: no human in the execution path — the senior you trust with the pager. Enterprise platforms encode the same ladder as three-tier permissions: read, simulate, execute. And the promotion criterion is the one good leads already use: graduated autonomy frameworks advance an agent only after it demonstrates a sustained success rate — the production literature cites thresholds like 99% over a defined period — within its current scope. A track record, not a vibe.
What decides the required rung is also familiar. The supervision-boundary literature converges on reversibility as the single most useful variable: irreversible actions — sending the email, executing the trade, deleting the data — demand the highest oversight, while drafts and analyses tolerate low supervision because undo is free. That's exactly the intuition behind letting an intern write all the design docs they want but not push to prod. Data sensitivity, action scope, and how far an error cascades before detection fill out the same calculus we run, mostly unconsciously, for people.
Supervision is a skill, not a checkbox
The ladder also imports the harder lesson: oversight has quality, and fake oversight fails silently. The regulatory phrase is "meaningful human involvement" — the reviewer needs the information, the time, and the authority to actually alter the outcome. A human who approves every agent recommendation without independent assessment isn't supervising, any more than a lead who LGTMs every PR unread is mentoring; both are single points of failure wearing a control's badge. And supervision imposes cognitive load, so it has to be designed: batched updates at natural breakpoints instead of a notification firehose, structured briefings before the agent hands back control, interrupt-and-redirect patterns that preserve state — the agent equivalents of a decent 1:1 cadence and a readable pull request. Skipping this design work produces the out-of-the-loop problem: the classic finding that a monitoring human's situational awareness quietly decays until they can no longer intervene competently. Leads who only skim eventually can't review; supervisors who only watch eventually can't act.
Where the analogy earns its limits
One honest break in the metaphor: a junior generalizes from feedback and keeps the lesson. An agent doesn't, unless you've explicitly built memory that persists corrections — and even then, a model upgrade or a prompt rewrite can reset behavior in ways no human career does. So an agent's track record predicts less than a person's: the trust you extended was earned by a specific configuration, not by "the agent." The practical fix is unromantic — re-run part of the probation whenever the configuration materially changes, and hold behavioral baselines (expected action and spend envelopes) that flag drift automatically.
But that caveat sharpens the thesis rather than undermining it. Scoped grants, earned promotion, reversibility-weighted oversight, supervision designed as real work — none of it was ever really about human nature. It was always risk engineering. We just learned it on people first.