Jevons Paradox in Software Development: Why AI Creates More Jobs, Not Fewer

Four percent of all public commits on GitHub are already written by a single AI coding tool, a tool that didn't exist a year ago. According to analysis by SemiAnalysis, this share could rise to over 20 percent by the end of 2026. What began as an experiment (autonomous AI agents that independently write, test, and deploy code) has become the new reality of software development. Top engineers at leading technology companies report not having manually written a single line of code for months. Instead, they orchestrate AI agents running in parallel, producing 20 or 30 pull requests per day.

For companies relying on high-quality web applications and software solutions, the question is no longer whether AI agents will change development, but how quickly they align their own processes, teams, and products accordingly.

What distinguishes "Agentic Coding" from traditional AI assistance

The term "AI in coding" is not new. Auto-completion, code suggestions, chatbot assistants: all of this has existed for years. But Agentic Coding marks a qualitative leap: here, the AI doesn't act as a passive assistant, but as an autonomous agent that uses tools, makes decisions, and takes action in real systems.

The crucial difference lies in the word agent: An agent is an AI model that doesn't just speak, but acts. It can:

• Search file systems and analyze codebases
• Independently perform Git operations
• Write, execute, and fix tests
• Control build processes and trigger deployments
• Read bug reports, evaluate telemetry data, and derive recommended actions

This is fundamentally different from a copilot that makes line-by-line suggestions. An agent takes over entire task chains: from understanding the problem to the finished solution as a pull request.

From code writer to conductor

The role of the software developer is thus shifting massively. Those who work effectively with AI agents today no longer write code in the traditional sense. Instead, they:

• Define goals and constraints: What should be built? What quality requirements apply?
• Control parallel agents: Five or more agents run simultaneously on different tasks.
• Review results: Code is read and validated, but no longer written.
• Iterate at a higher level: Architectural decisions, system design, and user feedback take center stage.

According to the Stack Overflow Developer Survey 2025, 84 percent of all developers already use AI tools or plan to use them, an increase from 76 percent the previous year. The Pragmatic Engineer survey from 2026 also shows that 55 percent of respondents regularly use AI agents, a massive jump within just a few months.

Why coding is "largely solved" and what that really means

The statement "Coding is solved" sounds provocative. What's meant is not that software engineering becomes obsolete. What's meant is: the manual activity of writing code (syntax, boilerplate, implementation details, dependency management) is no longer a bottleneck. AI agents solve these tasks reliably, faster, and increasingly error-free.

The parallel to history is illuminating: In 15th-century Europe, only a very small portion of the population could read and write. Scribes were highly specialized professionals. Then came the printing press. In the 50 years after Gutenberg's invention, more was printed than in the thousand years before. Costs fell by a factor of 100. Literacy rose over the following 200 years to around 70 percent in parts of Europe.

Crucially: the scribes didn't disappear. But their task changed radically. A historical example shows a scribe who welcomed the printing press because it freed him from tedious copying work and gave him more time for illustration and bookbinding.

For software development, this means: Creative, architectural, and strategic work moves to the foreground. Those planning and implementing complex web development projects today spend more time on the questions "What are we building?", "Why?", "How does everything connect?" and invest less time in pure writing work at the code level.

Five principles for AI-supported software development

From the practice of leading engineering teams, concrete principles can be derived that are relevant for every company engaged in or commissioning software development.

1. Build for tomorrow's model, not today's

AI models improve exponentially. What still works roughly today often runs reliably in just a few months. Anyone who tailors architectures, workflows, or products exclusively to the current capabilities of a model risks having to rebuild from scratch with the next model generation.

The principle: Design systems so they immediately benefit from more powerful models. In practice, this means:

• Loose coupling between business logic and AI components
• API-based architectures that enable model exchange
• Minimal scaffolding layers that don't become obstacles with model upgrades

2. Heed the "Bitter Lesson"

Rich Sutton's influential essay "The Bitter Lesson" from 2019 formulates an uncomfortable truth: general methods that scale with more computing power always beat specialized, handcrafted solutions in the long run.

For software development practice, this means:

• Don't build too much scaffolding around models. Elaborate orchestration workflows (step 1, then step 2, then step 3) may bring 10 to 20 percent improvement in the short term, but the next model leap often negates these gains.
• Bet on the more general model. Fine-tuning and specialized models have niches, but for most applications, the most powerful general model is the better choice.
• Give the model tools instead of instructions. Instead of prescribing exact steps, the agent receives a goal and the necessary tools and finds the path independently.

3. Deliberately keep teams lean

A counterintuitive principle: slightly understaffed teams often produce better results with AI agents than overstaffed ones. The mechanism behind this is simple: when staffing is tight, natural pressure increases to use AI agents creatively. A single engineer advancing a project with full agent support can deliver results that previously would have required a team of five.

The numbers support this: Anthropic reports a 200 percent productivity increase per engineer internally since introducing Claude Code, measured in pull requests. For comparison: in the pre-AI era, a few percentage points of productivity growth per year were already considered successful, even when hundreds of engineers worked on optimization.

4. Budget tokens generously

A common mistake: companies try to optimize the costs of AI model usage from the start. They use smaller, cheaper models or limit usage. This is counterproductive.

More powerful models are often more effective and ultimately cheaper because they solve tasks in fewer iterations and with less need for correction. A cheaper, less capable model needs more tokens for the same task; the total costs are often higher.

The better approach:

1. Phase 1, Experimenting: Provide engineers with as many tokens as possible. The costs for individual experiments are low relative to salary.
2. Phase 2, Scaling: Once a working approach is found, optimize specifically: smaller models for routine tasks, more powerful ones for core logic.

Some of the most productive engineers now consume six-figure token budgets per month. Increasingly, unlimited AI access is being used as a recruiting advantage: a signal that companies understand the leverage this investment provides.

5. Don't prescribe how the model should think

Many teams instinctively try to channel AI models into narrow paths: strict workflows, rigid prompts, exact step-by-step instructions. This worked with earlier model generations. Today, it's a hindrance.

The more effective approach: Give the model tools and a goal and let it determine the solution path. Modern agents independently decide which tools to use in what order. They write their own helper programs when needed, analyze data available to them, and find solution paths that no human developer would have instructed.

The role of humans: From coder to builder

The job title "Software Engineer" is facing a fundamental transformation. What's emerging is not abolition, but a redefinition: the role is increasingly merging with product management, design, and strategic planning.

In high-performing teams, the following can already be observed today:

The boundaries between these roles are blurring rapidly. The common denominator: Everyone builds. Product managers implementing their own prototypes. Designers converting their drafts directly into code. Analysts performing complex data evaluations via agents without writing a line of SQL.

For companies seeking strategic consulting in digital transformation, this yields a central insight: the most valuable employees will in future be generalists, meaning people who think across technology, product, design, and business logic and use AI agents as multipliers.

Jevons Paradox: More productivity, more demand

A legitimate concern: If AI agents write the majority of the code, do we need fewer developers? The data so far suggests the opposite, a phenomenon known as Jevons Paradox.

The basic principle: When a resource is used more efficiently, its price drops, and demand increases so strongly that total consumption rises. Applied to software development: When code production becomes drastically cheaper, thousands of projects become economically viable that were previously unthinkable.

Job postings for software developers, which reached a low point in mid-2025, have been rising continuously for months, now double-digit percentages above the low point depending on the data source. Companies that have introduced AI agents continue to hire. However, demand is shifting: there's increasing demand for people who solve problems, design systems, and orchestrate AI-supported workflows, less for pure code producers.

Security and quality: No autopilot without supervision

The speed at which AI agents produce code brings new challenges for quality assurance and security:

Automated code reviews as standard

When a single developer creates 20 or 30 pull requests per day, no human review system can keep up. The solution: AI agents check each other. The agent creates the code, a second agent checks it, with a human control layer on top.

Sandboxing and controlled execution

Agents that operate independently on file systems, APIs, and databases need clear security boundaries. Open-source sandboxes that execute agents in controlled environments are becoming the standard. The idea: give the agent enough freedom to work effectively, but with clearly defined boundaries on what it may touch.

Three levels of security

Three security levels have become established for responsible deployment of AI agents:

1. Alignment and interpretability: Understanding what happens in the model itself. Which neurons activate for which tasks? Are there signs of undesirable behavior?
2. Laboratory evaluation: Synthetic test scenarios that check whether the agent acts correctly in critical situations.
3. Behavior in practice: Observing the model in real-world use and learning from deviations.

The third level in particular is gaining importance: An agent can pass all tests and behave differently in practice. Therefore, an early, controlled release with a tight feedback loop is often safer than a delayed launch after perfected laboratory development.

What does this mean for your digital strategy?

The implications of Agentic Coding extend far beyond the engineering department:

For IT decision-makers:

• Software architectures should focus on loose coupling and API-first principles to benefit maximally from AI agents.
• The selection of tech stacks will be determined less by language preferences and more by agent compatibility.

For product managers:

• Prototyping cycles shorten to days instead of weeks. The ability to iterate quickly becomes a central competitive advantage.
• Systematically identify "latent demand": Where do users use existing products in unforeseen ways? That's where the next product opportunities lie.

For management:

• Investment in AI token budgets and infrastructure is not a cost center, but a productivity lever with measurable ROI.
• Teams working AI-natively can deliver more output with fewer headcount. However, qualification requirements for each individual increase.

Looking ahead: From code agent to all-round agent

The next evolutionary stage is already emerging: AI agents that don't just write code, but develop their own ideas. Agents that evaluate user feedback, analyze telemetry data, and independently make suggestions about what should be built or fixed next. Agents that automate project management, answer emails, and coordinate between different tools.

The development follows a clear logic: first coding, then tool usage, then computer usage. What is already reality for engineers, meaning autonomous agents working independently for hours, will become tangible for product managers, designers, analysts, and other functions in the coming months.

For companies building digital products or digitalizing their processes, now is the moment to question their own development culture. The question is no longer: "Are we using AI?" It is: "How quickly can we set up our teams, architectures, and workflows so they immediately benefit from the next model generation?"

At mindtwo, we accompany companies through exactly this transformation: from strategic conception through technical architecture to implementing scalable, future-proof digital products. Because the tools change. But the demand for well-thought-out, high-performance solutions remains.