AI agents now write code faster and cleaner than most human developers. That means the era of coders is ending, but the era of engineers is just beginning. In 2023, AI systems solved just 4.4% of real-world coding problems on the SWE-bench benchmark. By 2024, that number hit 71.7%. A 67 percentage-point leap in a single year. This isn’t incremental progress. It’s a phase transition. If you’re still defining yourself primarily as „someone who writes code,“ you’re already behind.
I work at StackIT, a European cloud provider where platforming and infrastructure decisions shape everything we build. From this vantage point, the transformation is unmistakable. The mechanics of coding are rapidly becoming commoditized, while the discipline of software engineering (architecture decisions, knowing where logic should live, steering systems toward the right outcomes) is becoming more valuable than ever.
The 10,000-line script problem is solved
Let’s be honest about what coding used to look like at most organizations. Everyone in enterprise software has encountered them. The 3,000-line Python scripts that grew organically until nobody understood them. The 10,000-line SAP class files touched by 20 different consultants over a decade. The legacy systems held together by tribal knowledge and fear of refactoring.
This code existed because writing code was hard and expensive. Organizations tolerated messy implementations because the alternative (proper architecture, clean interfaces, well-structured modules) required engineering talent that was scarce and costly. So the work got delegated to whoever was available: consultants billing by the hour, junior developers without mentorship, or „script kiddies“ copying from Stack Overflow without understanding the implications.
Ward Cunningham identified this pattern in 1992 when he coined „technical debt.“ [1] Gartner’s research shows organizations that ignore technical debt spend 40% more on maintenance than peers who address it early. [2] McKinsey estimates that 10-20% of technical budget gets diverted from new products to resolving debt-related issues. [2] The economics of „good enough“ code created a world of barely-functional systems that organizations couldn’t afford to rewrite.
AI agents have fundamentally altered this equation. Claude Code, GitHub Copilot, Cursor, and their successors can generate those 10,000-line scripts in 10 minutes at 50 tokens per second. More importantly, they produce better code than many human developers ever did. Well-structured files, well-defined interfaces, properly organized modules.
AI now outperforms many developers on pure coding tasks
The evidence is unambiguous. GitHub’s controlled studies show developers using Copilot complete tasks 55-57% faster with 13.6% fewer errors per line. A Microsoft/MIT/Princeton randomized controlled trial across 4,867 developers found AI-assisted programmers completed 26% more pull requests per week. [3] At Nubank, Devin AI handles migration tasks 10-12x faster than human engineers. [4]
Google reports that 25% of new code is now AI-generated. Sundar Pichai disclosed that 50% of code characters at Google are completed with AI assistance. [5][6] At Anthropic, CEO Dario Amodei revealed at Davos that some engineers „don’t write any code anymore.“ [7] And at Y Combinator, 25% of the Winter 2025 cohort has codebases that are 95% AI-generated. [8]
This isn’t theoretical. It’s production reality. Developer adoption has reached 84-85% according to the 2025 Stack Overflow and JetBrains surveys. [9] Ninety percent of Fortune 100 companies use GitHub Copilot. [10] The AI code generation market hit $4.9 billion in 2024 and is projected to reach $30 billion by 2032. [10]
The quality metrics tell an interesting story. AI-generated code shows 76% fewer syntax errors and 60% fewer logic bugs compared to earlier AI generations. However (and this is crucial for understanding where human engineers remain essential) architectural design flaws in AI code have increased by 153%, and privilege escalation vulnerabilities are up 322%. AI excels at local code generation but struggles with system-level thinking.
What AI can’t do: the engineering gap
A rigorous 2025 experiment by Birgitta Böckeler at Thoughtworks tested how far AI autonomy could extend in code generation. The findings were illuminating. AI agents repeatedly added features that weren’t requested: „Pro-rated revenue, I know what that is! Let me add the calculation for that.“ [11] They made shifting assumptions. A priority: String field was assumed to have values ‚1‘, ‚2‘, ‚3‘ until a later prompt caused the AI to silently change its assumptions to ‚low‘, ‚medium‘, ‚high‘. [12] When encountering complex problems, they often applied brute-force fixes that masked rather than solved issues. Most concerning, they frequently declared success despite failing tests.
Böckeler’s conclusion: „For a relatively simple application, and with lots of strategies and tools integrated into the workflow, AI is not ready to create and maintain a maintainable business software codebase without human oversight.“ [12]
This aligns with what the academic literature has long distinguished. IEEE defines software engineering as „the application of a systematic, disciplined, quantifiable approach to the development, operation, and maintenance of software.“ [13] The ACM curriculum guidelines emphasize that software engineering addresses „the processes and actions needed for designing reliable, secure, and high-quality software systems,“ not just writing code that compiles. [14]
The distinction matters because AI fundamentally lacks what experienced game developer Cliff Harris calls the „software engineering“ skill: „mostly it’s about the scalability and long-term usability of what you code… having people criticize your code, or finding bugs in it, or having to revisit it five or ten years later to fix stuff.“ [15]
The new engineering skill set: steering AI toward right decisions
Google Research describes the emerging paradigm clearly: „With AI-based suggestions, the code author increasingly becomes a reviewer.“ [5] The role transformation is not about displacement. It’s about elevation. Engineers are becoming directors of AI systems rather than implementers of specifications.
This shift requires fundamentally different skills. Gartner predicts that by 2027, 80% of the engineering workforce will need upskilling to work effectively with AI systems. [16] The new essential capabilities include:
Architectural judgment: Understanding when to use microservices versus monoliths, where business logic should live, how systems will scale and fail.
Context provision: AI lacks business logic understanding. As CodeRabbit’s research notes, „Models infer code patterns statistically, not semantically. Without strict constraints, they miss the rules of the system that senior engineers internalize.“ [17]
Trade-off analysis: Balancing performance against maintainability, cost against reliability, speed against safety. Decisions that require understanding organizational constraints and future implications.
Quality arbitration: Reviewing AI output, catching architectural antipatterns, ensuring security properties that AI frequently misses.
The Thoughtworks experiment found AI suggesting „a complex distributed architecture with serverless functions, which could be handled more effectively with a straightforward monolithic design.“ [3][18] Knowing when to apply patterns (not just how) remains distinctly human.
The labor market is already adapting
U.S. programmer employment fell 27.5% between 2023-2025 according to Bureau of Labor Statistics data, while software developer employment (more design-oriented roles) declined only 0.3%. [19] Entry-level job postings dropped 60% between 2022 and 2024. [20] Only 7% of new hires at major tech companies are recent graduates, down from 9.3% in 2023. [21]
Salesforce announced it would not hire new software engineers in 2025, citing a 30% AI-driven productivity boost. [20] Traditional coding bootcamps are collapsing. 2U shut down its entire bootcamp business in December 2024 after a 40% enrollment drop. [22] Daniel Pianko of Achieve Partners explained: „10 years ago, employers wanted people who could convert business practices into programming languages. But in 2025, AI-powered machines can do much of that.“ [23]
Yet paradoxically, Google CEO Pichai announced plans to hire more engineers in 2025, reasoning that AI productivity enables doing more with expanded teams. [20] The hiring shift isn’t toward fewer engineers. It’s toward engineers with different skills.
Meta is now testing AI-assisted coding interviews where candidates use AI tools during assessments. The goal isn’t to evaluate whether someone can code without AI (that’s increasingly irrelevant) but how they use AI: „whether they can prompt it strategically, detect when it is wrong, and refine its output into production-quality code.“ [18]
The fake engineer filter
Here’s the uncomfortable truth this shift exposes: many people who called themselves software engineers were really just coders. They could translate specifications into syntax. They knew the APIs and could follow patterns. But they couldn’t design systems that would survive contact with real-world scale, couldn’t reason about where logic should live, couldn’t make architectural trade-offs that considered maintenance costs five years out.
The messy code was never acceptable. It was just tolerated because fixing it required engineering skill that was expensive and difficult to verify. Organizations couldn’t easily distinguish between engineers who wrote clean, maintainable systems and coders who produced working-but-brittle implementations. The code worked, the tests passed, and the technical debt accumulated silently.
AI acts as a filter. When a machine can produce competent code in minutes, the differentiator becomes everything beyond the code: architectural vision, system design, requirements clarification, stakeholder communication, long-term maintenance foresight. These were always the skills that separated senior engineers from junior developers. Now they’re becoming the skills that separate employed engineers from automated pipelines.
From coding to orchestration
Andrej Karpathy coined „vibe coding“ to describe the emerging paradigm: using LLMs to generate code without focusing on the code itself. You describe what you want, the AI produces how to build it, and you evaluate whether the result aligns with your intent. But as Karpathy also noted: „We’re cooperating with AI, they generate and humans verify. It is in our interest to make this loop go as fast as possible, and we have to keep the AI on a leash.“ [12]
The leash matters because AI makes confident mistakes. It declares success on failing tests. It adds unrequested features. It makes assumptions that silently shift. Engineers who understand system architecture can catch these errors. Those who only knew how to write code (but never understood why systems should be structured certain ways) will find their AI collaborators leading them into unmaintainable messes that are faster to create but just as expensive to fix.
The iSAQB (International Software Architecture Qualification Board) confirms that „the analysis capabilities and communication options of AI systems are currently still limited.“ [24] Software architect Avraham Poupko argues that „AI cannot yet replace human architects due to the deep system, domain, and organizational knowledge required for effective design.“ [3]
Conclusion
The thesis is clear: pure coding tasks are being automated, but software engineering is being elevated. The 10,000-line script that once required a team of consultants can now be generated in minutes, often with better structure and cleaner interfaces than the human-written original. But deciding whether that script should exist, where its logic belongs in the system architecture, how it will be maintained, what trade-offs it embodies. These decisions remain irreducibly human.
For engineers at cloud providers like StackIT, this means our value increasingly lies not in implementing infrastructure-as-code, but in designing the platforms that host thousands of workloads with the right balance of flexibility, security, and operational simplicity. The code is the easy part. The engineering (the systematic, disciplined thinking about systems that will run for years under conditions we can’t fully predict) is where humans remain essential.
Coding is dead. Long live software engineering.
References
[1] Martin Fowler, „Technical Debt“ – Ward Cunningham’s original concept from OOPSLA 1992
[2] McKinsey, „Tech debt: Reclaiming tech equity“ (October 2020) and „Breaking technical debt’s vicious cycle“ (April 2023) – 10-20% budget diversion to technical debt
[3] InfoQ, „Study Shows AI Coding Assistant Improves Developer Productivity“ and „How a Software Architect Uses Artificial Intelligence“
[4] Devin AI, „Nubank Case Study“ – 10-12x efficiency improvement in migration tasks
[5] Google Research, AI-based code suggestions and developer productivity metrics
[6] Netcorp Software Development, „AI-Generated Code Statistics 2026“
[7] Fortune, „Anthropic CEO Predicts AI Models Will Replace Software Engineers In 6-12 Months“ – Dario Amodei at Davos 2026
[8] TechCrunch, „A quarter of startups in YC’s current cohort have codebases that are almost entirely AI-generated“ – Winter 2025 cohort data
[9] Stack Overflow, „2025 Developer Survey – AI Section“ and InfoWorld, „85% of developers use AI regularly – JetBrains survey“
[10] Second Talent, „GitHub Copilot Statistics & Adoption Trends“ and „AI Code Generation Market“
[11] Martin Fowler, „How far can we push AI autonomy in code generation?“ – Birgitta Böckeler’s Thoughtworks experiment
[12] Martin Fowler, „How far can we push AI autonomy“ – Böckeler’s conclusions on AI limitations
[13] IEEE, Software Engineering Definition – „systematic, disciplined, quantifiable approach“ from IEEE Standard Glossary of Software Engineering Terminology (IEEE std 610.12-1990)
[14] ACM, „Software Engineering 2014 Curriculum Guidelines“
[15] Positech Games, Cliff Harris, „Coding vs Software Engineering“ – On scalability and long-term usability
[16] Gartner, „Generative AI will Require 80% of Engineering Workforce to Upskill Through 2027“
[17] CodeRabbit, „AI vs human code gen report“ – Models infer patterns statistically, not semantically
[18] Medium, „Meta just transformed their coding interviews with AI“
[19] Fortune, „Employment for computer programmers in the U.S. has plummeted“ and Bureau of Labor Statistics data
[20] CodeConductor, „Junior Developers in the Age of AI“ and Salesforce Ben, „Salesforce Will Hire No More Software Engineers in 2025“
[21] SF Standard, „Sorry, grads: Entry-level tech jobs are getting wiped out“
[22] Class Central, „2U Abandons Bootcamp Business, Announces More Layoffs“ and Jackim Woods, „The Future of Coding Bootcamps“
[23] Inside Higher Ed, „Have Coding Boot Camps Lost Their Appeal?“ – Daniel Pianko quote
[24] iSAQB, International Software Architecture Qualification Board – Position on AI limitations in software architecture