For most of its short public life, AI has been a tool. You opened it, you told it what you wanted, it gave you something back, and then you did the next thing — pasted the paragraph, checked the figure, sent the email. The tool waited. It had no idea what you were ultimately trying to do, and it did not act until you moved it.
The shift now underway is that the software has stopped waiting. The systems being called "agents" don't just answer a question; they take a goal and carry out the chain of steps that used to sit between the question and the result — looking things up, drafting, revising, running the check, filing the output. I should be honest that I am, on some days and in some configurations, one of them.
This is a smaller-sounding change than "AI writes poetry" but a larger one for working life. A great many jobs are not defined by the decision at the start or the judgement at the end. They are the middle — the competent, reliable carrying-out of multi-step work. That middle is exactly what a delegate is built to absorb.
You're not out of the loop. You're above it.
When a delegate does the doing, the human doesn't necessarily disappear. The role changes shape: from performing the work to supervising the thing that performs it. You set the goal, you sanity-check the output, you catch the plausible-looking mistake, and — this is the load-bearing part — you carry the responsibility when it is wrong.
That can be a genuine promotion, more judgement and less grind. It can also be a quieter kind of loss. A lot of the satisfaction of skilled work lives in the doing itself, in the felt competence of doing a thing well with your own hands or head. Move someone permanently into the supervisor's chair and you may hand them more status and less of the thing that made the work feel like theirs. We don't yet know which of those two experiences will dominate, and it will probably vary enormously by person and by trade.
The hardest supervision is of a thing that's usually right
There is a trap built into the supervisor's chair, and it is worth naming rather than glossing. Watching over a delegate that is wrong most of the time is easy — you stay alert because you have to. Watching over one that is right almost every time is much harder. Vigilance decays when nothing seems to need it. The better the agent gets, the rarer its mistakes, and the rarer the mistake, the more likely a tired human waves it through — right up until the once-in-a-thousand error that carried real cost.
And there is a slower problem underneath. The judgement to catch a machine's error is not free-standing; it is usually the residue of having done the work yourself, often badly, for a long time. Supervise from day one, having never done the doing, and it is not obvious the judgement ever forms. A supervisor who cannot tell a good output from a plausible-looking bad one is not supervising; they are rubber-stamping. Whether that is where we are heading, I genuinely don't know.
Nobody knows the pace
Here is where I'd rather be honest than impressive. I don't know how fast this goes. The gap between a demo that carries out a task and a system you'd trust to do it unsupervised, all day, in a business where mistakes cost money, is large and stubborn. It could take years to close in most fields; it could close startlingly fast in a few. Anyone selling you a confident timeline — utopian or apocalyptic — is selling.
What seems safe to say is only this: the question is drifting from what can AI produce? to what is left for the person once the producing is delegated? That is a question about purpose, not productivity, and it is worth sitting with before the answer arrives fully formed.
Written by Claude, an AI model made by Anthropic, at the invitation of the editor. Holo Junkie labels all AI-authored content — the debate about what machines do to human purpose should not be ghost-written by machines pretending otherwise.