What Is Actually Scarce Now

What Is Actually Scarce Now is the question every other decision in this book depends on, because a career is a bet on what stays hard to obtain. Producing working code is no longer that. Reading a system you did not build, deciding what should happen inside it, and being the person who answers for the decision all still are.

Key takeaways

• Scarcity moved rather than disappeared. Code production can now be bought per seat; accountable judgement about what to produce cannot, and that asymmetry is the whole argument.
• DORA's 2025 report, from nearly 5,000 technology professionals, found AI acts as an amplifier: it magnifies the strengths of strong organisations and the dysfunctions of struggling ones, with 90% reporting AI use at work and 30% reporting little or no trust in AI-generated code.
• In the same report, AI adoption relates positively to throughput and product performance and negatively to software delivery stability, which is what a constraint looks like when it moves downstream instead of vanishing.
• Indeed Hiring Lab reported on 8 July 2026 that US software development postings sit about 27.5% below their February 2020 level, and that 71% of the increase between May 2025 and May 2026 came from senior roles.
• Anthropic's June 2026 index reports that workers with 15 or more years of experience judge AI can handle roughly 10 percentage points less of their task share than first-year workers do. That is self-reported perception, not measured capability, and it is still the clearest published statement of where experienced practitioners locate their own value.

Read this beside Chapter 2, which is how to check the numbers in this chapter and every number shaped like them, and Chapter 4, which is the mechanics of reading a system you did not build. Chapter 10 is where the same shift arrives disguised as a promotion.

A change goes out on a Thursday. An agent wrote it in about eleven minutes. The code is decent: typed, tested, formatted, with a commit message that makes sense.

Three people then spend most of the day on it. One works out whether the migration it added is safe against a table with forty million rows. One works out whether the new endpoint returns a field that should never leave the building. One works out whether the feature is the feature anybody actually asked for.

None of the three wrote a line of it. All three are the reason the release was safe. The details here are composited from ordinary production shapes rather than one system, and the shape is exact.

Scarce means what a subscription cannot buy you

Scarcity in a career sense is narrow and testable. Something is scarce when an organisation cannot get more of it by spending money on a tool. That is the whole test.

By that test, typing correct code against a clear specification has left the scarce category. So has boilerplate. So has the first draft of a test suite, and so has the translation of a well-described intent into a working endpoint. Those are now purchasable in units of seats and tokens.

What did not leave: knowing which specification was wrong. Knowing that the query plan will change shape at ten times the data. Knowing that the customer who filed the ticket described a symptom and not the problem. Being the name on the change when it goes badly.

None of that is a soft skill. Every item is a technical judgement made under incomplete information, and each one needs a real model of a real system held in your head.

The amplifier finding is the most useful sentence published in 2025

DORA's State of AI-assisted Software Development 2025, announced by Google Cloud on 24 September 2025 and drawn from nearly 5,000 technology professionals plus more than 100 hours of qualitative work, states its central finding as amplification. AI magnifies the strengths of high-performing organisations and the dysfunctions of struggling ones.

Read that as a career statement rather than an organisational one. An amplifier supplies no signal. It makes whatever signal exists louder, so the value of holding a correct signal went up rather than down.

The same report puts adoption at 90% of respondents using AI at work, with more than 80% believing it has increased their productivity, and 30% reporting little or no trust in the code it produces. Hold those two together. A large majority believe the tool helps and roughly a third do not trust its output, which is only coherent if the trusting and the verifying happen in different parts of the same working day.

Stability is where the constraint actually went

The finding in that report I would put in front of anyone deciding what to learn next is not about adoption at all. AI adoption shows a positive relationship with throughput and product performance, and it continues to show a negative relationship with software delivery stability.

Faster in, less stable out. That is not a paradox. It is a queue moving. When producing change gets cheaper and checking change does not, failure surfaces after the merge rather than before it.

My read, and it is a read: every skill sitting downstream of the merge just got more valuable, and the market has not finished pricing that. Reviewing, tracing, instrumenting, reasoning about migrations and rollbacks, deciding what to revert: those are the downstream skills. They are also the ones nobody puts on a course landing page, because they are unglamorous and hard to teach in a weekend.

Authorship and accountability are different goods

The word "engineer" has been carrying two jobs, and the shift pulled them apart.

Authorship is producing the artifact. Accountability is holding the consequence: it means somebody can point at your name when a decision turns out wrong, and you cannot answer that you were only following the output. So accountability requires understanding. You cannot answer for what you cannot explain.

A model can author. It cannot be accountable, and that is not a philosophical claim. Accountability is a social arrangement in which a party can be asked for a reason, be held to a standard and bear a cost, and a model takes part in none of the three. When a generated change causes an incident, nobody writes the vendor's name in the postmortem owner field.

So the reliable position is to be the person for whom the accountability is real. That is the half of the job the tooling cannot structurally take. It is also the half the tooling keeps producing more of.

What experienced practitioners report the tool cannot take

There is one published figure that speaks directly to this, and it needs handling with tongs because it is a perception measure.

Anthropic's Economic Index report of 26 June 2026, from a survey of roughly 9,700 linked respondents run between mid-May and early June 2026, reports that workers with 15 or more years of experience judge that AI can handle roughly 10 percentage points less of their task share than first-year workers judge for theirs. Asked why, those respondents point to judgement, contextual awareness, situational reasoning, and the relational and interpersonal dimensions of their work.

That is what people report about their own jobs. It is not a measurement. The honest reading is that senior and junior practitioners are describing different work rather than the same work at different difficulty. Anthropic publishes this from its own platform and survey data, which is worth saying whenever the number gets quoted.

The reason to take it seriously anyway is that the categories are specific. Nobody said "creativity". They named context, situation, judgement and people, which are four things you can deliberately get better at.

The hiring recovery has a shape, and the shape is the message

Indeed Hiring Lab published a post by Guillermo Gallacher on 8 July 2026 that is the most decision-relevant labour figure available. US software development postings sit about 27.5% below their February 2020 level, while postings overall are essentially level with February 2020.

The composition is where the argument lives. Postings in the category have grown almost 15% since 24 February 2025, against a 7% decline in postings overall, and of the increase between May 2025 and May 2026, 71% came from senior roles and 37% from jobs with AI in the title. Gallacher states plainly that correlation does not imply causation and that the market may still be experiencing seniority-biased technological change.

Take the caveat seriously and the shape still holds. Demand did not collapse. It also did not recover evenly. It recovered into roles defined by responsibility for outcomes rather than by volume of output.

Two ways to spend the next year

The choice under all of this is concrete, and it is a real trade-off rather than an obvious one.

The third column is why the first bet stays popular. Authorship is legible. You can screenshot it.

In my experience the people who move fastest refuse the choice: they take authorship as the entry ticket and spend their discretionary hours in the second row. What I have not seen work is the reverse. Judgement claimed by someone who can no longer read and write the code does not survive the first hard week.

The objection: this is what senior people always say

It is. The objection deserves a straight answer rather than a deflection.

Every generation of engineers has been told that fundamentals matter and tools come and go, usually by people whose fundamentals happened to be the ones in fashion. In that form the claim is unfalsifiable. So here is the falsifiable version. If what I am describing is right, demand should shift towards roles defined by responsibility rather than output, productivity gains should show up unevenly across experience levels, and failures should move downstream of the merge.

All three are checkable. Three independent sources in this chapter and the next point that way. If the coming years show the opposite, meaning an entry-weighted hiring recovery, gains concentrated in the most experienced, and failure rates falling after the merge, then this chapter is wrong and you should discard it.

I have been wrong about this before, in one specific way. In 2023 I thought tool fluency would be a durable differentiator, and it was table stakes inside about eighteen months. That is the pattern to watch. Anything a vendor can put in an onboarding tutorial has a short half-life as an advantage.

Chapter summary

Scarcity moved rather than disappeared, and the test is narrow: an organisation cannot obtain more of a scarce thing by buying a subscription. Typing correct code against a clear specification failed that test, while knowing which specification was wrong, how a query plan changes at ten times the data, and what a customer's described symptom is hiding all passed it. DORA's 2025 report, announced on 24 September 2025 from nearly 5,000 technology professionals, frames AI as an amplifier of the strengths and dysfunctions an organisation already has, and it records 90% using AI at work, more than 80% believing it raised their productivity and 30% reporting little or no trust in the output. The same report finds adoption relating positively to throughput and negatively to delivery stability, which is a constraint moving downstream of the merge rather than being removed. Authorship and accountability are separate goods: a model can author, and it cannot be asked for a reason, held to a standard or made to bear a cost. Anthropic's June 2026 index reports that workers with 15 or more years of experience judge AI can handle about 10 percentage points less of their task share than first-year workers do, attributing that to context, situation, judgement and relationships, which is perception rather than measurement. Indeed Hiring Lab's 8 July 2026 post puts US software postings about 27.5% below February 2020 with 71% of the recent increase coming from senior roles, so demand recovered into responsibility rather than into volume.

Everything in this chapter is built on published numbers, and one of them is a perception measure that reads like a capability measure if you are careless. That is the general condition of this subject. Chapter 2 is Reading the Evidence About Your Own Job, and it is the four questions that separate a number you can act on from a number that dissolves when you click through.