We're onboarding agents better than interns.

Written By Sébastien Simpson

Agents are getting badges, managers and permissions that grow as they earn trust. Who is doing the same for the people who will have to check their work?


Think back to the first piece of work someone senior took apart in front of you. It wasn't a strategy. It was a meeting recap, a tracker, a first-draft deck with the wrong number on slide four. The red pen stung, and it was the most efficient training anyone ever gave you. You learned what mattered by getting it wrong in a place where mistakes were cheap and someone was paying attention.

That arrangement is being rebuilt right now, with care and real investment. It's being rebuilt for software.

The agent got a badge.

On October 15, ServiceNow announced AI Passport, a signed credential each AI agent checks before it acts. Permissions narrow as a task moves from step to step, the platform keeps "an auditable record of who authorized it," and an operator can stop the agent mid-task. The same release describes LTM, an IT services firm, giving each of its 1,500 agents an employee ID and a manager. It's one of the most thoughtful answers I've seen to the question of who's accountable when an agent acts, and it ships in Q4.

Put that feature list next to a good apprenticeship and they match point for point: bounded scope, a named supervisor, a record of the work, and authority that widens only as trust builds. Two weeks ago I wrote that agents forgot to introduce themselves. The best platforms are now fixing that, with more structure than most companies give a new analyst.

The autonomy spectrum runs both ways.

Most agent governance now assumes an autonomy spectrum. At one end the agent suggests and a person does the work. Further along it drafts for approval, then acts inside a defined scope while someone samples the output, and at the far end it acts on its own and reports exceptions. Credentials like the one above are how an agent earns each step.


People sit on a matching spectrum, and it moves in the same direction. In a Forbes piece on October 16, WTW's John Bremen lays out three human positions: in the loop (reviewing outputs), at the helm (setting limits and escalation triggers), and in the lead (owning the outcome). The further an agent climbs, the further its human has to climb with it. A reviewer becomes a steerer, and a steerer becomes an accountable owner. Each human rung assumes you've mastered the one below it.


California just wrote that assumption into law. SB 1246, newly signed, covers robotaxis, the far end of the autonomy spectrum by any definition. Any human who drives one remotely must be based in the U.S. and hold a U.S. driver's license, and operators must provide local incident technicians on the ground. At peak autonomy, the state still insists on a human who can do the job the machine is doing.


Bremen also resurfaces a line from a 1979 IBM training deck: "A computer can never be held accountable, therefore a computer must never make a management decision." Forty-seven years later, the industry is building credentials and audit trails so that someone always is.

The first rung is getting thinner.

The trouble is at the bottom of the human spectrum. A September Census Bureau working paper by Cody Orr, Lee Tucker and Lawrence Warren found that graduates of the most AI-exposed majors became less likely to land a first job after late 2022, and their first-job earnings fell 13%. The authors compare the hit to graduating into a large recession.


The work disappearing first is the work that can be written down: the recap, the reconciliation, the first pass at a deck. The judgment that can't be written down, knowing which number on slide four is wrong and why it matters to this client, grows through repetitions of exactly that work. Agents are absorbing the repetitions. The judgment still has to grow inside people, and it's the judgment every higher rung depends on.

Everyone forecasts judgment. Few fund growing it.

IBM's 2026 CHRO study, released in September and based on surveys of 1,500 executives and 8,800 employees, shows how far apart the two sides sit. 71% of executives prioritize the ability to supervise, validate or override AI output. Only 38% of employees rank it that high, and 60% say skill erosion affects them directly. Leaders want reviewers. Employees aren't being told that reviewing is the job.


Some leaders are naming the problem out loud. Jayney Howson, ServiceNow's chief learning officer, told ITPro last month that "technology adoption is moving faster than skills development," and that "learning must keep pace with work." It's encouraging to hear that from a company building the tools, and it's the right brief for every enterprise buying them.


Few have picked it up. Gartner's latest CIO survey, reported by CIO Dive, found only 4% plan to redesign roles for a workforce of humans and agents, and those that do are 2.4 times more likely to be top AI performers.

In Botsitting I wrote about the hours oversight costs today. This is the same bill, deferred by five or ten years. Bremen calls it oversight debt.

The loop still needs someone who can say no.

Last week I described the human's remaining job in a self-improving system as the outside judge, the person who sets direction and signs off on what the loop produces. Platforms are now building that sign-off into continuous-improvement workflows, which is the right call. It also raises the stakes on whoever signs. An approver who never did the work is approving on faith.


New research from UC Berkeley shows how quickly that faith turns into dependence. In a peer-reviewed paper presented at COLM this month, researchers including Brian Christian ran three randomized trials with 1,222 participants. People who had ChatGPT available for basic math and reading problems, then lost it partway through, solved fewer problems and skipped more than people who never had it. "This is not a story about fractions and SAT problems," Christian said.


The tasks were small. The habit they reveal, giving up once the help disappears, is the one a person at the helm can't afford. The helm is where you take over when the help stops working.

Give the juniors the agent treatment.

If agents earn their way up the spectrum with scoped credentials, a named manager, an audit trail and widening permissions, people can climb the same way:

  • Scoped credentials become scoped work: real deliverables with real stakes, sized so a mistake stays cheap.

  • A named manager becomes a named reviewer, someone who explains why the work changed instead of quietly fixing it.

  • An audit trail becomes a reasoning trail. Juniors draft first, then compare their version with the AI's, with someone senior in the room.

  • Widening permissions become a written path from producing work, to checking it, to steering it, with criteria for each step.


The reasoning trail already has backing. Forbes reported last week on a McKinsey approach: let employees "try selected work first," then review "the differences between their reasoning and the AI's with a manager." IBM's study makes the funding case: put the capacity AI frees up into coaching and stretch assignments rather than letting it vanish into cost savings.


One measure carries over from last week's piece: override rate, tracked per reviewer. When someone's rate drifts toward zero, either the AI got better or the reviewer stopped looking. You need to know which, and you need to know while there are still people around who can tell the difference.


Gartner's Alicia Mullery told CIO Dive, "I think we've passed the time of endless scale." If she's right, the next edge won't come from running more agents. It will come from having people who can spot when the agents are wrong, and those people are made slowly, out of juniors who were handed real work and a red pen. California decided a robotaxi still needs someone with a license. I'm curious how long it takes the rest of us to ask the same question about our agents, and whether there will be anyone left with the license when we do.

Further Reading.

Sébastien Simpson
Next

Agentic AI: Should the orchestra play itself?