The New Software Paradox
How AI coding agents will reshape the software market by shrinking large industry players and expanding the total number of developers
A C Compiler in Two Weeks
Nicholas Carlini, from Anthropic, published the results of a remarkable experiment. He tasked sixteen agents with writing, from scratch, a C compiler. Not a prototype. A functional compiler in Rust, capable of compiling Linux 6.9 for x86, ARM and RISC-V. The result was 100,000 lines of code that surpass 99% of compiler test suites, including the rigorous GCC torture test[i].
The project took two weeks. The token cost was around $20,000. My estimate of how long it would take to do the same thing by hand is around a year and a half of work. Although the figure for doing it by hand can be debated, the leap in magnitude when it comes to developing a compiler seems clear.
But there is another detail beyond the numbers. Carlini is not a compiler specialist. His work was not to write the code, but to design the environment where the agents could work. The tests that guided them. The feedback system. The infrastructure for two thousand sessions. After that, he largely stepped back.
What matters is that Carlini did not need to be a specialist. He understood how to build verifiable systems, how to structure the problem, or how to evaluate results. That is what it means to be a generalist.
Software is shifting from being expensive to develop to being accessible, with a marginal cost trending toward zero. An industry built on the assumption that it is expensive and difficult has to rethink everything. The type of knowledge that matters, or the processes used to build it, change completely.
Vibe Coding and Why Code Still Matters
But this conclusion requires nuance. We need to distinguish between two phenomena that are sometimes confused because they share the same tools.
Vibe coding is the natural evolution of low-code and no-code. The main premise is that what matters is the result. Describe what you want, the AI tool generates it. Anyone can create functionality in hours without technical training. It is very powerful and its reach keeps growing.
But “it works” is not enough for professional software. Vibe coding has limits worth understanding, particularly in terms of maintainability, scalability and security.
Language models need readable code. When technical debt accumulates, such as tangled, unstructured code, the model loses coherence. It may seem like a paradox, but the model that generates the code needs that code to be well written.
An application can work with ten users, but collapse with ten thousand. Architectural decisions do not make themselves. They require judgment and direction. That judgment requires, in turn, understanding what is happening underneath.
A CRM can run for months while harboring a vulnerability that exposes all the data. Without technical training, you do not know what to ask, you do not know what to request. You do not know what you do not know.
The compiler project is not vibe coding. Carlini designed a rigorous test environment. He applied engineering criteria. And he had something that does not always exist: decades of tests that define what “correct” means.
This is a spectrum. At one end sits vibe coding: simple, fast, low-risk software. At the other, professional development — software that scales, that is maintained, that is secure. In the latter, code and architecture matter. You need to understand the underlying architecture, even if you do not write the code by hand.
Training in computer science and software engineering does not disappear as a requirement, but is applied differently. The developer shifts from code writer to director of a process that produces code. To direct that process with judgment, you need to understand what is being produced.
The Jevons Paradox Applied to Software
In 1865, William Stanley Jevons noticed something paradoxical. The steam engine improved and coal became cheap. Lower coal consumption was expected, but the opposite happened. Cheap coal and its optimized use enabled previously unthinkable applications and consumption soared.
The Jevons Paradox states that when a process becomes more efficient in using a resource, demand for that process can grow, and consumption of the resource, instead of falling, rises. This has held true for software until now. The more efficient software development has become, the more software has been made and the more programmers have been needed. It is possible that this paradox will continue to affect software development in this new phase. There are even people who argue that demand for software is infinite[ii].
If software costs ten times less, we will produce the same for less money. That is the intuitive response. But the reality is that we will do much more because there is latent demand waiting to emerge. SMEs without custom software because it was expensive, specific needs forced to fit generic products, or processes managed on spreadsheets that could be software. The cost never justified building that software, but now it is cheaper.
According to an Anthropic study[iii], more than 90% of tasks in Computer & Math could be performed by an LLM, although real adoption today is still around 3%. That gap indicates we may be at the beginning of a major wave of automation.
However, growth signals are starting to become visible. The technical labor market, far from contracting, shows increases already in the first quarter of 2026. Demand for software engineers is up 11% year-over-year in the U.S., according to Indeed data analyzed by Citadel Securities[iv]. Morgan Stanley projects that the size of the software development market could grow 20% annually through 2029[v].
A quick exercise. If 90% of tasks are automated, 10% remains that requires judgment: architecture, systems evaluation, security. If at the same time demand grows by one order of magnitude, the number of jobs needed will be the same, but if it grows two orders of magnitude, that 10% will represent more work than exists today, up to 1,000%. The argument that “AI destroys jobs” implies that demand does not grow or that 100% of a job’s tasks are automated. The elasticity of demand for the value a profession creates is the part that is rarely discussed, and it will be what makes the difference between some professions and others.
The Jevons Paradox does not promise that everything will be straightforward. It promises that demand can grow when costs fall. What it does not say is who captures that demand, or how, nor what form the resulting market takes.
The New Software Economy
The change is not only one of speed, but also of cost structure.
Until now, building software has meant, primarily, paying salaries. Developers are practically the only cost. Traditional development tools are a marginal line item. Software is a people industry, like consulting or law.
With multi-agent systems, this can change. The tokens of a language model become a real and significant cost. In multi-agent systems with high autonomy, they can represent between 60% and 80% of the total project cost. The C compiler cost around $20,000 in tokensi. The human equivalent, over 18 months, would have cost nearly 10 times more. So the total cost falls, but it is distributed differently. Most of the cost is no longer labor — it is compute.
As a consequence, the developer with experience and judgment appears to be in greater demand. The one who mainly executed instructions from others has little room to protect their work.
A new risk also emerges. If 60-80% of a project’s cost is tokens, a price change from Anthropic, OpenAI or Google can jeopardize the project’s viability. Before, the cost was in-house labor. Now it depends on external providers. Most organizations building software have not yet added this to their risk analyses.
Three Worlds in Collision
The impact does not hit everyone equally. There are three main players in the software market, and their moves cascade. What a large IT department decides puts pressure on its integrators, and what both do changes the available space for product companies.
Large Enterprise IT Departments
A large IT department, with 10,000 developers, 30% in-house and 70% contractors, could produce ten times more software with today’s tools with that same team.
But most large companies do not want ten times more software. They want the same software for less money. Or perhaps more software (three times more?), but with the same budget. For many large companies, where IT is an expense, not an investment, the goal is to contain costs.
On the other hand, the bottleneck shifts. If developing is ten times faster, what matters is knowing what to build. You need ten times more ideas or requirements. The product team becomes the new bottleneck. There is data pointing in this direction[vi]. In the first quarter of 2026 there were more than 7,300 product manager positions open globally in tech, at three-year highs and growing 20% since the start of the year. When execution capacity ceases to be the bottleneck, what becomes scarce is knowing what needs to be built.
What seems most likely in those departments is that external developer headcount will be cut — the easiest line item to reduce — and more weight given to internal, experienced profiles. These are profiles that can now do alone what previously required a team of ten people coordinated by them. They are the ones who know what to build, in what direction, and have the judgment to evaluate whether what the agents produce is good.
Software Integrators
Large technology consulting firms and outsourcing companies face a double blow. They are the first affected by their clients’ reduction in developers, and their business models are the most exposed.
Early data is beginning to reflect this shift. Accenture has laid off 11,000 employees and made AI adoption a requirement for senior promotions[vii]. The major Indian IT firms (TCS, Infosys, Wipro) have eliminated more positions in the U.S. in the first three months of 2026 than in all of 2025[viii].
Integrators that live on selling billable developer hours are the most exposed. Their advantage was scaling teams quickly. That advantage disappears when a smaller team with agents produces the same output. Niche specialists, those offering know-how (financial regulation, healthcare, security…) have more margin because that deep knowledge is difficult to replicate with generic tools. But the bulk of the generalist business will contract sharply.
The natural path for many integrators would be to move down toward smaller customer segments. The problem is that this requires cost structures far lighter than what they have today, and they will also encounter new competitors that were born without that burden.
Product and SaaS Companies
SaaS companies, or simply product companies, face a different but equally serious problem. The barriers to entry they have built are dissolving.
If software is cheap to make, it is also cheap to copy. A CRM, a vertical ERP, a project management platform… any of these can be recreated at a fraction of the original cost. Perhaps not with all its functionality, but with enough functionality, and without the client needing to adapt to the product. The “SaaSpocalypse” of 2026 reflects this pressure with a drop in the valuation of software companies of $300 billion in the first week of February 2026[ix]. In specific cases, companies such as Atlassian, Workday or Monday.com, for example, are starting to notice a drop in licenses[x].
Although these observations from early 2026 are general, not all products are equally vulnerable. There are at least three levers that can protect these software product makers. Those with strong network effects (social networks, marketplaces) are not replaced simply because software is cheaper, because their value lies in the users, not the code. Those tied to physical infrastructure or regulated processes have entry barriers that are not affected by the marginal cost of software creation trending toward zero. And those with established distribution channels (Microsoft and Google through products like email and office suites, or Salesforce with its global commercial network) can defend their position because integration or customer access can be worth more than the cost of the product.
Those with none of these defenses, those competing primarily on functionality and price, are in territory that is difficult to protect over the medium term.
The three perspectives converge in that, just like IT departments, all will demand fewer developers, empowering experienced profiles with their own judgment. All will generate a surplus of professionals who will need to find new opportunities.
The Flourishing of Local Software
While the large companies in the sector are shrinking, something different can grow outside the mainstream.
A developer knows a dive center that is part of a group of ten dive centers operating in natural reserve areas with regulated and limited access. Every year, these centers need to coordinate who enters when, respecting quotas and conditions for each one. Until recently, that process consumed two months of shared work at the beginning of the year. Using software for this was beyond their reach because it is a problem too specific for generic software, too small to justify professional development by an integrator, and also too costly for the budgets of these dive centers.
Today that developer offers them the solution in exchange for an annual subscription that each center can afford. The natural next step is to contact other groups of centers in other reserves and offer the same service. It is a small, highly specialized business that solves a real problem for real people. It did not exist before because the costs of creating that solution made it unviable — neither for the dive centers nor for the developer.
This pattern will repeat itself thousands of times. Most small and mid-sized businesses in the world do not have custom software because the cost never allowed it. Millions of coordination tasks, workflows and specific needs have always made do with generic solutions that do not fit well, or with no technological solution at all. As development costs fall dramatically, that latent demand can become a real market.
Some of that demand will certainly be met through vibe coding. Simple needs, without complex security requirements, with short life cycles. But as soon as the solution needs to be maintained from one season to the next, incorporate new requirements, scale to more users or handle sensitive data, the developer needs the solid training we discussed: knowing how to structure the system, how to secure the data, how to make it work when things get complicated. The developer for the dive centers is not doing vibe coding. He is building a real product that needs to work well every season.
These new companies have, nonetheless, a natural growth ceiling. Service-based ones depend on proximity and personal trust, which are difficult to scale beyond a geographic radius or a specific niche. Product-based ones face a structural problem because the same ease that allows them to create their solution allows others to replicate it. Low barriers to entry are good for entering a market, but they also let competitors in.
In my opinion, this segment can no longer produce a new unicorn built around a software product. A company with a valuation of one hundred million dollars may be possible, but reaching a billion is very difficult. As soon as a solution gains traction, the ease of replicating it is so great that competition appears before a dominant position can be consolidated. Historical software unicorns were able to grow because software was expensive to make and difficult to copy. Those conditions have changed.
The New Developer
The first labor market signals in 2026 tell a story that seems contradictory at first glance.
Despite the explosion in software development productivity, there are more engineering positions open than a year ago. More than 67,000 globally in tech, with 26,000 in the U.S. alone and accelerating since the start of the year, according to data from Lenny Rachitskyv. Developer demand is rising. At the same time, listings for pure implementation roles (developers whose main function is translating specifications into code) have fallen 17%[xi]. More developers are in demand, but different developers. Job listings requiring experience with AI-assisted development tools have risen 340% between January 2025 and January 2026[xii].
The market is recomposing demand, not reducing it. It is not looking for someone who writes code; it is looking for someone who knows what code should exist, how it should be structured, and can evaluate whether what the agents produce meets the requirements of a professional system.
Foundational training does not disappear. Architecture, systems design, databases and security remain the core of required knowledge. This knowledge is simply applied differently. Code becomes a byproduct of the process, but to direct that process you need to understand what it produces. Those who do not understand what the agent generates cannot evaluate or scale it.
There is a legitimate concern about how experience is acquired. If the entry-level projects that juniors used to handle are absorbed by automation, the talent pipeline of future expert developers may shrink considerably. Learning in the early years happens by solving real problems. If agents solve those problems, the path becomes more complicated.
But there is an alternative path that I think is underestimated. In the scenario we have described, with thousands of small business and niche needs that previously could not afford software, developers with little experience have the opportunity to create their own solutions for clients that previously did not exist. The developer for the dive centers did not need years of experience at a large company to build that product. He needed solid training, the ability to understand the client’s problem, and tools that make accessible what previously required a team. It is an entrepreneurial path, not a corporate one. It does not guarantee stability or linear growth. But it is a path where real experience can be gained.
The end result is an industry with a different shape. Large software companies (integrators, product, SaaS and corporate IT departments) are shrinking. Meanwhile, what is growing is the long tail, made up of thousands of small solutions for specific problems, created by people who understand both the problem and the technology needed to solve it.
The software industry is shrinking. The activity of creating software is expanding. And the developer profession looks less and less like writing code and more and more like knowing what to build and making sure it works as intended.
[i] Carlini, N. (2025). “Building a C Compiler with Claude.” Anthropic Engineering Blog. https://www.anthropic.com/engineering/building-c-compiler
[ii] Stack Overflow (2026). “Why Demand for Code Is Infinite: How AI Creates More Developer Jobs.” https://stackoverflow.blog/2026/02/09/why-demand-for-code-is-infinite-how-ai-creates-more-developer-jobs/
[iii] Anthropic (2025). “The Labor Market Impacts of AI: Measuring Exposure with Real-World Data.” https://www.anthropic.com/research/labor-market-impacts
[iv] Citadel Securities (2026). “The 2026 Global Intelligence Crisis.” https://www.citadelsecurities.com/news-and-insights/2026-global-intelligence-crisis/
[v] Morgan Stanley Research (2026). “How AI Coding Is Creating Jobs.” https://www.morganstanley.com/insights/articles/ai-software-development-industry-growth
[vi] Rachitsky, L. (2026). “State of the Product Job Market in Early 2026.” Lenny’s Newsletter. Data not directly verified; secondary source with broad coverage. https://www.lennysnewsletter.com/p/state-of-the-product-job-market-in-ee9
[vii] India TV News (2026). “Accenture Lays Off 11,000 Employees: Makes AI Adoption Mandatory.” https://www.indiatvnews.com/technology/news/accenture-lays-off-11000-employees-makes-ai-adoption-mandatory-2026-02-20-1031052
[viii] Unbox Future (2026). “AI’s Silent Layoff Wave: Indian IT Giants Cut US Jobs as Automation Accelerates.” https://www.unboxfuture.com/2026/04/ais-silent-layoff-wave-indian-it-giants_01924283433.html
[ix] Forbes (2026, February 4). “$300 Billion Evaporated. The SaaS -Pocalypse Has Begun.” https://www.forbes.com/sites/donmuir/2026/02/04/300-billion-evaporated-the-saaspocalypse-has-begun/
[x] Market Minute (2026). “The 2026 ‘SaaSpocalypse’: Why B2B Software Stocks Are Plunging 20%.” https://markets.financialcontent.com/stocks/article/marketminute-2026-3-24-the-2026-saaspocalypse-why-b2b-software-stocks-are-plunging-20
[xi] CIO (2026). “Demand for Junior Developers Softens as AI Takes Over.” https://www.cio.com/article/4062024/demand-for-junior-developers-softens-as-ai-takes-over.html
[xii] CNN Business (2026, April 8). “The Demise of Software Engineering Jobs Has Been Greatly Exaggerated.” https://www.cnn.com/2026/04/08/tech/ai-software-developer-jobs