From Coding to Commanding: How AI Is Rewriting the Developer’s Role

By Shangyu Zhao

In February 2025, Andrej Karpathy, a Slovak-Canadian computer researcher, described a way of building software that sounded careless. A developer writes what they want in plain English, accepts whatever the AI produces, and in his words, “forgets that the code even exists” (Karpathy, 2025a). He called this process “vibe coding”, and within days, the idea had spread far beyond the circles that usually follow machine learning research.

The appeal was obvious. In 2025, Collins Dictionary named vibe coding its Word of the Year (Collins Dictionary, 2025). Y Combinator reported that 25% of startups in its Winter 2025 cohort had codebases that were 95% AI-generated. The approach was not just being adopted by developers, but by designers, product managers, and students with no programming background, drawn by the possibility of turning ideas directly into working software. 

However, the problems appeared just as quickly. A 2025 analysis by CodeRabbit covering hundreds of open-source GitHub repositories found that AI co-authored code contained around 1.7 times more major defects than human-written equivalents, with security vulnerabilities appearing at nearly three times the rate (CodeRabbit, 2025). The issue was not that the systems failed to produce working code. It was that developers accepting outputs without scrutiny had no reliable way of knowing when something had gone wrong.

This concern was not limited to researchers. At IBM’s TechXchange conference, Neel Sundaresan argued that the shift away from careful development had gone too far: “It’s not about vibe coding; it’s about literal programming. It’s not about accessibility, it’s about security” (IBM, 2025). The speed that made the approach attractive was the same feature that made it difficult to trust.

One year later, Karpathy moved on, and in February 2026 he introduced a different idea: agentic engineering. Instead of stepping back entirely, the developer remains responsible for directing and evaluating systems that can act on their own. The work is no longer writing each line, but managing a process that unfolds across many steps. As TRM Labs put it, the shift is toward “orchestrating systems that write code” (TRM Labs, 2025).

Claude Code, developed by Anthropic, offers a concrete example. The tool runs directly in the terminal, reading entire codebases, editing files across directories, running tests, and resolving conflicts without waiting for instructions at each step. Between October 2025 and January 2026, the longest sessions nearly doubled in duration, as users became more willing to extend autonomy over time (Anthropic, 2026). Even so, experienced users grant full autonomy in only 40% of sessions. In the majority of cases, they remain actively involved, stepping in when the system moves in an unintended direction.

This detail matters. The most capable tools, used by the most technically fluent users still require human judgment most of the time. The role of the developer has not disappeared; it has shifted. Writing code is no longer the central skill. Evaluating what has been produced, identifying where it has failed silently, and making decisions about how a system should behave now matter more to users.

For students outside computer science, this shift is particularly significant. The barrier to building software has dropped sharply. A student in almost any discipline can now ask a system to analyse data, automate a task, or build a simple application. Whether the result is useful depends less on technical knowledge than on judgment: knowing when to trust the output and when to question it.

Karpathy compared the current moment to the arrival of personal computing in the 1970s (Karpathy, 2025b). If the comparison is accurate, the challenge is not simply learning to code. It is learning to think clearly about what machines are asked to do, and when their answers should be trusted.

References

Anthropic (2026) Measuring AI agent autonomy in practice. https://www.anthropic.com/research/measuring-agent-autonomy  

CodeRabbit (2025) State of AI vs Human Code Generation Report. https://www.coderabbit.ai/blog/state-of-ai-vs-human-code-generation-report  

Collins Dictionary (2025) Word of the Year 2025. 

https://www.collinsdictionary.com/woty  

IBM (2025) Beyond code: The new role of the developer, IBM Think, 7 October. https://www.ibm.com/think/news/new-role-developer-techxchange-2025  

Karpathy, A. (2025a) ‘There’s a new kind of coding I call “vibe coding”…’, X, 2 February. https://x.com/karpathy/status/1886192184808149383  

Karpathy, A. (2025b) Year in Review 2025. https://karpathy.bearblog.dev/year-in-review-2025/  

TRM Labs (2025) Vibe engineering in the age of vibe coding. https://www.trmlabs.com/resources/blog/vibe-engineering-in-the-age-of-vibe-coding