Your Job Is to Train the Agent, Not Do the Work

Here is the shift worth taking seriously: the unit of labor inside a company is quietly moving from the task to the agent that does the task. You don’t write the support reply, draft the contract, or close the ticket. You stand up, brief, and babysit the thing that does. Your output stops being the work product and becomes the agent’s reliability. If that holds, it rewrites the job description of nearly everyone in a knowledge org, and it does so faster than any reorg memo can keep up.

This piece was sparked by The AI Industrial Revolution with Naval Ravikant. The idea below is the part I think builders should be planning around right now.

The framing is already coming from the people building the agents #

This isn’t a thought experiment from the future. The CEOs shipping the platforms are saying it in plain language. At CES in January 2025, Nvidia’s Jensen Huang said “the IT department of every company is going to be the HR department of AI agents in the future” — teams onboarding agents, training them on company vocabulary, briefing them on policy, and overseeing fleets of them, exactly as HR does for people (Fortune). By October he’d extended it: future enterprise workforces are “a combination of humans and digital humans,” and “I wouldn’t be surprised if you license some and you hire some” — with onboarding and culture absorption treated as a people-management problem, not a deployment (Fortune).

Marc Benioff put it more bluntly at Davos: from now on “we will be managing not only human workers but also digital workers,” and today’s CEOs are “the last generation to manage only humans” (Fortune). You can read that as hype. But the org charts are following the rhetoric.

The new job titles are real, and they’re about supervision #

KPMG sketched the emerging org chart: Workforce Planning Architect (AI Focus), AI Training & Development Lead, AI Performance Manager, Orchestration Engineer, AI Governance & Risk Specialist. Notice what every one of those roles has in common — none of them do the underlying work. They teach, route, measure, and govern the systems that do.

Harvard Business Review went further and named the function. In February 2026 it ran a piece, co-authored by a Harvard Business School professor and a Salesforce executive, defining the “agent manager” as an explicit job — monitoring how agents work, learn, and adapt, “much like how a traditional manager might walk the floor, check in with a struggling employee, or huddle with a team on a tricky case.” It quotes Zach Stauber, a support agent manager at Salesforce: “Data, Data, Data. I start and end my day in dashboards, scorecards, and agent observability monitoring” (HBR). That is a full-time job whose entire content is supervising software. It did not exist three years ago.

The practice is hardening below the C-suite too. HR Executive documented teams giving agents “clear scopes of work, assigning a point of contact similar to a manager,” building monitoring and feedback loops, and writing governance for naming, ownership, and performance review — with some companies folding agent management into HR outright (HR Executive).

The usage data backs the direction, not just the slogans #

Job titles can be theater. Telemetry is harder to fake. Microsoft’s May 2026 Work Trend Index reported 15x year-over-year growth in active agents across Microsoft 365, and 18x inside large enterprises. The more telling number is the shape of AI-assisted work: 49% of conversations now support cognitive work — analysis, problem-solving — while only 17% are about producing outputs directly (Microsoft). The center of gravity is moving from “make the thing” to “judge and direct the thing that makes it.” Sixty-six percent of users say they spend more time on high-value work as a result.

Salesforce is the clearest worked example. It routed roughly 32,000 customer conversations a week through Agentforce at an 83% autonomous resolution rate (Cyntexa). As that volume climbed, it redeployed hundreds of support engineers with deep product expertise into “forward-deployed engineer” roles — making sure agents hit business outcomes, watching adoption, designing escalation — and minted net-new titles that didn’t exist before 2025: Deployment Strategist, AI Conversation Designer, AI Architect (Salesforce). The work didn’t vanish. The humans moved up a layer, from doing the resolution to engineering the system that resolves.

Now the honest part #

If you stop here, this is a clean promotion narrative: everyone graduates to manager. The evidence says it’s messier, and a builder planning a team should hold the counter-case just as firmly.

“Training the agent” is mostly just prompting. Real agent training — fine-tuning, RLHF, workflow engineering — stays with a small technical elite. For most people, “managing the agent” means writing prompts, nudging parameters, and curating outputs: closer to supervising a contractor than building a durable asset. The title inflates faster than the skill does.

Oversight degrades into rubber-stamping. This is the load-bearing risk. The European Data Protection Supervisor’s 2025 TechDispatch found that human oversight of automated systems routinely collapses into ritual approval — a quasi-automation where the human signs off because the system already decided, leaving accountability formally present but practically hollow (EDPS). A review of automation bias in human-AI collaboration points the same way: people systematically over-trust AI output, especially under time pressure, so a “human in the loop” can manufacture a false sense of control without real scrutiny (AI & Society). An agent manager drowning in dashboards is exactly the conditions under which approval becomes reflex. Supervision at scale is not automatically supervision.

The pipeline that produces good supervisors is breaking. Here’s the trap underneath the whole model: you cannot competently train an agent to do work you’ve never done yourself. Judgment comes from having done the junior version. But that junior version is what’s being automated first. Stanford’s Digital Economy Lab, analyzing ADP payroll data, found software-developer employment for ages 22–25 down nearly 20% from its late-2022 peak; a Harvard study found junior employment at AI-adopting firms fell 7.7% relative to non-adopters within six quarters (ThinkPol). CNBC has noted this risks cutting off career advancement for young workers before they accumulate the tacit knowledge that makes oversight meaningful (CNBC). If the rungs you climbed to become a competent overseer are sawn off, the supply of competent overseers eventually dries up.

And some of this is contraction wearing a transformation costume. Salesforce redeployed hundreds of people and reportedly laid off over 1,000 while pledging no net-new software-engineering hires in 2025 (Salesforce Ben). The net headcount math isn’t neutral, and “everyone becomes a trainer” is a comfortable story to tell over a downsizing. Around 40% of employers surveyed expect to reduce headcount where AI automates tasks (Cyntexa). The high-end roles being created — AI Performance Manager, Orchestration Engineer — are scarce, concentrated, and demand a rare blend of domain depth and AI fluency. Most people displaced by an agent will not slide cleanly into managing one.

What to actually do with this #

The directional claim is sound: the work is moving from execution to supervision, and the job titles, telemetry, and named roles all point the same way. Plan for it. Build the dashboards, the scopes of work, the escalation design before you scale a fleet of agents.

But treat the two failure modes as first-class engineering problems, not footnotes. Design oversight that resists rubber-stamping — sampling, adversarial checks, real consequences for a missed approval — because a human who only ever clicks “approve” is not oversight, just liability with a pulse. And protect the pipeline: if you automate away every task a junior would have done, fund some other path for people to earn the judgment the supervisor role requires. The org that gets the agent-trainer model right will be the one that remembers a trainer is worthless if they never learned the work themselves.