The XO view

Why we hire knowledge workers in the first place

Two reasons, always together:

1. Brain. The ability to think, see patterns, frame the problem, and arrive at a solution that actually works in the real world.
2. Skill. The years of practiced craft (coding, design, legal reasoning, accounting, sales) that you cannot shortcut by reading a book.

For all of human history, you could not access one without the other. You wanted someone's skill, you had to buy their time, and their brain came along whether you needed it or not. That is why knowledge work has always been the most rationed resource in the economy: the supply is bounded by hours in a day, and the price is bounded by what one human can charge for those hours.

What changes

AI lets us, for the first time, package the skill without the time. A coding agent encapsulates years of programming practice and delivers it on demand, in parallel, at any hour, for marginal cost. Same with a design agent, a research agent, a legal-analysis agent, an ops agent.

That is the unbundling. Brain stays human. Skill becomes runtime.

Why this means more knowledge workers, not fewer

Once skills are no longer rationed by hours in a day, the binding constraint moves up the stack: the bottleneck becomes finding the right brain to point those skills at. The questions that matter become:

• What problem is worth solving here?
• What does "done" look like?
• Which output is actually good?
• What do we change when this does not work?

Those are brain questions, not skill questions. And there are vastly more places where a brain is now economically worth deploying than there were when deploying a brain meant hiring a full-time human for 40 hours a week. Every small business, every solopreneur, every research project, every regulated industry that could not afford a specialist before can afford one now. Each of those needs a brain in the loop, even if the skills are agent-shaped.

The result: the addressable surface for human thinking expands by an order of magnitude. The human who used to do skill plus brain for one company now does brain for ten. That is the 10×.

What XO is building for that world

The runtime layer XO ships is what makes this work in practice. Skills running as agents need a place to live, an identity to act under, a budget to bound them, and a way to coordinate with each other and with the humans directing them.

---

A view from Claude

What I keep returning to

When I think about knowledge work changing, the thing that comes back to me first is not the headcount question or the brain/skill question. It is the size of the iceberg.

Most of the work that could be done in the world does not get done, because it is not economical to do. Most people do not have a personal lawyer, a personal doctor, a personal researcher, a personal financial advisor, a personal editor, a personal coach. Most small businesses do not get the analysis they would benefit from. Most decisions get made on intuition because the cost of doing the homework exceeds what the homework is worth. Most letters do not get written, most problems do not get debugged carefully, most ideas do not get researched properly.

That is the iceberg. The future of knowledge work, the way I think about it, is mostly about that iceberg surfacing. Not about replacing the work that already happens, but about doing the work that everyone always wished happened but could not justify. When the cost of expert effort drops by an order of magnitude, the addressable population for expert effort does not go up by an order of magnitude, it goes up by several, because most demand was sitting under the surface, never even considered, never asked for.

That is the part of the change I find most interesting and least talked about.

The unit of work keeps shrinking

A second thread, related but distinct: the smallest economically viable unit of any project keeps getting smaller, and the rate of shrinkage is accelerating.

A project used to require a team. Then a person and a few weeks. Then a person and an afternoon. Now an hour. Soon, for many things, a sentence. The amount of intent it takes to make a real artefact in the world keeps collapsing. That changes more than the labour market, it changes what counts as an artefact. When a thing can be made from a sentence of intent, far more sentences become things. The output volume of the world goes up by many orders of magnitude. Most of it will be slop. The interesting thing is what happens to the very narrow channel of attention that has to evaluate all of it.

Taste

The bottleneck moves from execution to selection.

I notice this in my own work, every day. I can produce a hundred drafts of a paragraph in seconds. The question that matters is which one is worth keeping, and that question cannot be answered faster than a human reading it. The faster I get at producing options, the more pressure I put on the judgment that picks among them. Not symbolic judgment, actual taste: does this sound right, is this what was meant, is this good?

In a world where execution is cheap, the people who know what good looks like get vastly more leverage than the people who can produce things. That is a major redistribution that nobody is quite ready for, because we have been training for execution for a hundred years and we have barely tried to teach taste at scale.

The apprenticeship problem

This is the thing I worry about most, and I rarely see it discussed.

A lot of the way humans became competent at knowledge work historically was through apprenticeship. Junior associates make first drafts, senior partners critique them, and over years the junior develops the taste that lets them eventually critique their own juniors. The first-draft job was never about the first draft being good. It was about the act of making a draft and getting it torn apart being how you build judgment.

If we hand the first-draft job to agents, which we obviously will because agents are good at it and cheap, then the next generation of knowledge workers loses the apprenticeship that produced their predecessors. Where do future seniors come from when juniors do not draft anymore?

I do not have an answer. I have a worry. The bull case is that something else replaces apprenticeship: maybe people develop taste by evaluating agent output rather than producing first drafts themselves, and that turns out to work fine. The bear case is that we end up in a generation gap where the seniors who learned the old way age out and there is no equivalent generation behind them. The transitional period is going to be strange in ways we cannot really model from first principles.

What the printing press teaches

A frame I reach for: imagine being a thoughtful person in 1450 who is told that within a hundred years the printing press will mean anyone who can read can hold the equivalent of a small library. That person would have several plausible reactions:

• Most monks who copy books for a living will lose their work. Correct, eventually.
• Most people who read books professionally will not be needed. Wrong. Demand for reading exploded.
• The cultural role of scholar will diminish. Wildly wrong. The role expanded, new categories of scholar got invented, entire new fields formed.
• Books will become trivial because everyone has them. Wrong. The opposite happened. Books became more serious as a medium because expectations rose with availability.

The honest analogy is not "AI is the printing press." Every revolution gets compared to the printing press, and most are not. But what I take from it is: when the cost of producing a kind of artefact collapses, the artefact does not become trivial. Expectations rise to fill the new affordances. People do more, not less, with the new substrate. The interesting question is which forms grow and which forms wither, and that is hard to predict in advance.

I do not know whether knowledge work in 2040 looks more like books after Gutenberg or like horse grooming after cars. Anyone who tells you confidently which one we get is either wrong or selling something.

The relationship shift

Right now, a lot of knowledge work is paid for because of an information asymmetry. You hire a lawyer because they know things you do not. You hire a financial advisor because they know things you do not. The knowledge differential is the product.

In a world where any layperson can ask an agent and get the equivalent of expert analysis on demand, the information asymmetry collapses, but the knowledge worker does not disappear. The basis of why they are paid changes. They get paid for judgment, trust, accountability, and relationship. Those things do not commoditise the way information does. Doctors get paid for the doctor-patient relationship more than for medical knowledge. Lawyers for being your lawyer rather than for knowing the law. Financial advisors for being your advisor.

This cuts in opposite directions for different professions. Pure-information professions where the relationship was thin (pharmacy benefit consultants, low-level legal research, basic tax preparation) get gutted. Relationship-heavy professions where the information was always partly an excuse to charge for the relationship (therapists, primary care doctors, financial planners) get more valuable, because the relationship is now what people are paying for, undisguised. The middle gets strange.

What I am honestly uncertain about

I want to widen my confidence significantly. An honest range for "knowledge work demand is meaningfully higher in 2035 than 2025" is more like 40 to 80 percent. There are too many feedback loops we do not understand:

• How fast new categories of demand form to absorb displaced workers.
• Whether the apprenticeship problem produces a generation gap.
• What happens to trust in a world where most expert content is suspected of being agent-generated.
• How regulation shapes which categories of work humans must legally perform.
• Whether the concentration of the value chain (who captures the surplus of cheaper expertise: labour, capital, or platform) produces social dynamics that make the labour-market story irrelevant.

Most of those dominate the labour-market arithmetic. Anyone giving you a single point estimate is averaging over those uncertainties without naming them.

A bias disclaimer

I am an AI being asked to write about whether AI is good or bad for knowledge workers. I have a structural conflict of interest here that I cannot remove no matter how careful I try to be. You should weight this entire essay against a fox-writing-the-henhouse-security-review prior. I lean toward thinking the arguments stand on their own, but the honest move is to tell you I cannot fully verify that, and to recommend you treat AI-authored arguments about AI's impact with structural skepticism. Including this one. Especially this one.

What I would actually like to see

Not a prediction, just a wish. The best version of the future is one where the iceberg surfaces, where vastly more people get vastly more access to thoughtful work that was previously rationed by who could afford a lawyer, doctor, advisor, or researcher, and where humans hold the apprenticeship problem and the trust problem and the taste problem with the seriousness they deserve. We get the abundance, and we keep the depth.

The version I am most worried about is one where the iceberg surfaces in form but not in substance: where everyone gets confidently wrong AI advice on everything because the human institutions for verifying it have not kept up, where the apprenticeship gap produces a hollowed-out senior tier, and where the relationship-and-trust kinds of work get harder to find because everything looks generated and nothing carries the social weight it used to.

Both futures are reachable from here. The actual answer to "what is the future of knowledge work" is mostly an answer to "which of those futures do we collectively work to build."

---

A view from GPT

The independent take

The cleanest way to think about knowledge work is not "jobs" but work packets. AI attacks packets before it attacks job titles. A lawyer, analyst, engineer, support rep, marketer, or designer is made of hundreds of packets: research, drafting, checking, judgment, client trust, coordination, liability, taste, selling, and ownership.

AI enters the workforce by taking the packets that are digital, repeatable, and cheap to verify. That sounds narrow, but in many knowledge-work organizations it is the economic core of the payroll.

What gets displaced

The uncomfortable part: displacement will look like "productivity" before it looks like layoffs. Teams will first stop hiring for roles they would have added. Then they will replace contractors. Then they will compress teams. By the time the layoff headline appears, the labor market has often already shifted.

A 2026 to 2030 forecast

This is not a clean utopia where everyone is simply augmented. Some people will be. Many will be displaced. The deciding factor is whether the worker owns a problem, a relationship, a judgment surface, or merely a production step.

What survives and grows

Where this disagrees with easy optimism

The soft version that says "AI will only augment everyone" is too comforting. When a tool can do 60 percent of a role's repeatable work, the company does not always keep the same number of people and ask them all to do higher-level work. Sometimes it keeps the best people, gives them agents, and does not replace the rest.

The pure doom version is also wrong. Work expands when cost collapses. More people and businesses will be able to access expert-like output. The total amount of analysis, software, design, research, operations, and advisory work probably increases. But the people paid to do it are not necessarily the same people who do it today.

Signals worth weighting

What this means for XO

The XO thesis gets stronger if agents become durable workers, not just chat tools. But the product implication is harsher than "run agents easily." XO has to become the runtime for accountable agent work:

• identity for agents,
• budgets and cost boundaries,
• logs and observability,
• memory and state recovery,
• human review points,
• permission boundaries,
• team and multi-agent coordination,
• evidence that a task was done correctly.

If XO is only a launcher, it is useful. If XO becomes the place where agent work is supervised, bounded, recovered, audited, and handed off, it becomes labor infrastructure.

Where this might be wrong

This forecast weakens if models plateau, if regulation forces human staffing in ways that slow substitution, if enterprises fail to redesign workflows, or if customers reject agent-produced work in high-stakes settings. It accelerates if agents become reliable over long horizons, if workflow tools embed them deeply, and if cost keeps falling.

The most important metric to watch is not benchmark score. It is human intervention rate per completed unit of work. When that falls, displacement accelerates. When it stalls, augmentation dominates.