Money Maker

I walk out of the room and the machine keeps working.

That sentence has sold more courses, more funnels and more Telegram groups than almost anything else on the internet. So I am going to earn it backwards: mechanism first, money last. If the plumbing does not convince you, the numbers at the end should not either.

The planning stage turns messages, requirements and open questions into instructions precise enough for an unattended system to execute. The useful output is a specification: what to build, what constraints apply, and how the result will be checked.

The intended property is durable progress: a workflow can resume from recorded state, detect failures and produce evidence for acceptance without requiring an uninterrupted interactive session. Duration alone is not a quality metric.

It is also, by now, boring. That is the part nobody mentions. The magic wears off and what is left is infrastructure.

The Job Question Is a Decoy

Almost everyone asks me the same thing: is AI going to take my job?

It is not a stupid question. Stanford's Digital Economy Lab has been tracking it against ADP payroll records across more than 730 occupations, and at the entry level the pattern is ugly. Workers aged 22 to 25 in the most AI-exposed occupations have been shedding employment at roughly 4% a year through April 2026, while the same age group in the least-exposed jobs kept growing around 2%. Controlling for firm-level shocks, the early-career decline in the exposed group lands near 16%. It is not arriving as mass layoffs, either — it shows up as openings that stopped appearing. (Canaries Dashboard, Stanford Digital Economy Lab) That is observational data, not a controlled experiment. It shows the floor moving under young workers in exposed jobs; it does not prove AI alone moved it.

So the fear is not irrational. It is just aimed at the wrong object.

“Will AI take my job?” is a question about wages, asked while machine execution is changing how software gets produced. A parallel question is who owns the systems, products, and distribution that turn that execution into useful work.

The economic question is how useful work becomes a sustainable product. Ownership, customer demand, operating costs and compensation shape the outcome; generating more artifacts does not itself create wealth.

Model access creates another implementation option; compare its complete cost and accepted results.

What Actually Runs While I Am Not There

Everything I have written about AI for the last few years has been one long progression, and it was never really about prompts. Simple agent use, then knowledge bases, then the wall you hit when a single markdown file stops scaling, then knowledge bases that maintain themselves, then graphs instead of similarity search, then specs, verification, orchestration, multi-agent review, cost control. Every layer existed to fix a failure the previous layer exposed.

Prompting is a weak moat and a worse operating model. What comes out of the planning stage is not prompts — it is decisions, constraints, priorities, acceptance criteria, corrections and open questions. The raw material of a specification. Human language is the highest-leverage programming language we have now, and the tax on being vague went up, not down.

Then seven things have to be true at once, or the whole thing produces confident garbage for twenty-four hours instead of work.

Written down like that it sounds bureaucratic. In practice it is the opposite — it is what buys the freedom to leave. Autonomy is not a permission you grant an agent; it is a property of the environment you build around it. And autonomy is not the absence of a human. It is the absence of unnecessary human blocking. The system should never ask me how to run a test suite. It should always ask me whether a product decision is worth making.

The standard should be concrete: outputs must satisfy stated requirements, pass relevant verification and expose uncertainty. A technical brief should cite its evidence; code should have appropriate checks; a change should identify its risks and acceptance conditions. An academic-sounding label is not a substitute for those criteria.

The Part Where the Skeptics Are Right

The single most important number in this argument cuts against me, so let me put it first.

METR ran a randomized controlled trial with sixteen experienced open-source developers across 246 real tasks in repositories they had worked on for an average of five years. With AI tools allowed, they were 19% slower. Afterwards, they estimated AI had made them 20% faster. (METR, July 2025)

Believed: twenty percent faster. Measured: nineteen percent slower. That gap should end every conversation about AI productivity that runs on feelings — including mine.

The honest caveats: those results concern early-2025 tools and mature repositories familiar to the participants; they do not rank every possible workload. METR have since redesigned the experiment. But the finding that survives all of that is the perception gap, and it is the reason my entire system is built on gates that do not care how fast I feel. If a workflow feels impressive and loses against the baseline, it dies. I have written about velocity as a drug precisely because I am susceptible to it.

Two more objections, all of them fair:

A prepared demo does not establish reliability on an existing codebase. Existing systems can have contradictory history, undocumented conventions, fragile integrations and years of architectural sediment. Dropping an agent in there without a knowledge substrate or verification gates is not innovation, it is negligence with a progress spinner. There is also a long list of things these models are still surprisingly bad at, and pretending otherwise is how people get hurt.

Solo-founder wealth stories are survivorship bias. Behind a successful product may be failed experiments, distribution built earlier, capital, timing, and luck. Cheaper production does not remove market risk; it can make an unsuccessful experiment less expensive.

That is the distinction: the workflow should lower the effort required to produce an accepted result. Whether it does so must be measured against the relevant baseline; artifact volume alone does not answer the question.

Persistent execution infrastructure

Persistent execution can run on a local server or a cloud host. Choose the environment around availability, isolation, recovery, data handling and cost requirements rather than treating cloud placement as proof of autonomy.

This is not an infrastructure cleanup. It is a change of category.

A workstation configured to sleep interrupts execution; a persistent host can continue while its operator is away. Either host can be local or remote. Durable state, bounded credentials, monitoring and recovery determine whether a long-running workflow is dependable.

METR’s task horizon measures reference-human task duration at a stated model success probability. It is not the elapsed time an agent runs unattended. The January 2026 methodology update also reports broad uncertainty and sensitivity to task composition.

Hosting decisions therefore need measurements of actual execution time, interruption tolerance and required availability. A rising human-equivalent task horizon alone cannot establish a need for continuous cloud execution.

Always-on operation also means an always-on blast radius. A persistent system needs least-privilege credentials, budgets, quotas, audit trails, isolation, and a kill switch. Continuous model use can accumulate substantial cost, so cost control belongs in the architecture alongside correctness and access control.

Nobody Should Be Managing Agents

The end state is not a dashboard where I assign tasks to six digital characters and watch them talk. That interface is a transitional mistake. If I am still deciding which agent does what, I am still the orchestration layer — I have just given the bottleneck a nicer UI.

An Automated Software Factory is an architectural direction: persistent knowledge, state, agents, workflows, orchestration, policies, evaluations, integrations, and applications around them. Humans express intent through familiar interfaces; the system turns that intent into candidate specifications and verified artifacts. Agents become implementation details within a workflow whose decisions remain accountable.

A factory does not fix bad judgment — it manufactures the consequences of it faster. Which is why the human stays exactly where the leverage is: direction, design, taste, risk, accountability. Everything below that line is machinery, and machinery should not need a manager.

And none of this is specific to software. Anything whose output can be specified and evaluated fits the same architecture — research, operations, analysis, planning, documentation. "Any domain" does not mean every result is automatically correct. It means the same production system adapts without dragging humans back into every low-level loop.

What Appreciates When Execution Gets Cheap

Now the part I actually care about, and the reason this article exists.

Revenue per employee can describe one aspect of company economics, while annual recurring revenue is a run-rate measure with its own assumptions. Neither reveals profit, cash flow, ownership allocation or individual compensation.

Lower production costs may change company economics, but employee outcomes depend on ownership, profitability and compensation. A small organization with high revenue is not evidence that every engineer has realized personal wealth.

Meanwhile the picks-and-shovels trade is already crowded and already priced. Estimates for 2026 hyperscaler infrastructure spending cluster between $600B and $690B, up from roughly $388B the year before, and the binding constraint has quietly moved from chips to electricity — Morgan Stanley projects US data-center demand near 74 GW by 2028 against a shortfall around 49 GW (Morgan Stanley, 2026). Everyone can rent a pick now. Owning one distinguishes nobody.

Here are five strategic hypotheses about where value moves as execution becomes cheaper. Each depends on assumptions that can fail.

Own the system, do not rent out your hours. Expertise still pays, but income tied entirely to your personal execution has a ceiling and now has machine competition underneath it. AI is not coming for your identity. It is coming for your pricing power. AI can broaden the implementation tasks an individual engineer can undertake — the question is whether that engineer owns the output or invoices for it.

Equity in what machinery produces is different from access to the machinery. Model access is widely available; products, licenses, recurring customer relationships, and ownership of useful output have different economics. None of those assets becomes valuable without demand.

Distribution appreciates as production inflates. When everyone can generate code, copy, design and prototypes, supply explodes. Attention does not. Trust does not. A channel to the right buyer does not. When the software business model itself gets repriced, the audience is the part that does not get commoditized.

A maintained knowledge base can be difficult to copy. Its value comes from decisions, relationships, evaluation results, failure patterns, and outcomes that improve future work. That is a stronger foundation than an undifferentiated folder of documents.

Compute and energy are real, but they are the consensus trade. Capital-intensive, brutally competitive, and now rate-limited by grid interconnects measured in years. Knowing where to dig still beats owning a shovel.

The scarce asset may be the operating system around the model: knowledge, orchestration, evaluation, integration, distribution, and judgment. Those capabilities determine whether rented intelligence becomes useful, durable output.

Thought Is Getting a Supply Chain

There are documented industrial deployments, with narrower tasks than a general-purpose household robot. Figure reported in November 2025 that its Figure 02 deployment at BMW used ten-hour weekday shifts for sheet-metal handling. That is a dated vendor account of one task and environment, not proof that unrestricted physical automation is solved or a ranking of all robotics deployments.

I have written the speculative version of this twice, as the last interface and the last machine, and mapped the decade in Decade Zero. The unglamorous version is the one I believe: the distance between a precise human intention and a finished thing in the world keeps collapsing, and every layer of that collapse gets bought by whoever owns the machinery on the right side of it.

The Window

None of this is guaranteed. Models could plateau, infrastructure economics could change, and products could fail to find customers. These are architectural and market possibilities, not a report of private investments or a promise of returns.

The opportunity is to build a workflow whose accepted results justify its operating costs. Evaluate reliability, recovery, customer value and ownership explicitly; neither more generated work nor greater autonomy guarantees a financial outcome.

Start with a bounded workflow and a measurable acceptance standard. Keep the evidence needed to decide whether additional autonomy improves the result.

The engineering opportunity is to build systems that keep producing verified work without continuous human supervision.