Agentic AI promises unprecedented productivity, but revives the lessons of the Y2K bug. Without appropriate governance and testing, security and maintainability risks increase…
A technological acceleration that recalls a historical precedent
Agentic AI is profoundly transforming software development. Capable of generating code, correcting errors or testing applications, AI agents promise an explosion in productivity. But this race for speed recalls another period: that of the Y2K bug. At the end of the 1990s, companies discovered that technical choices made with a view to immediate efficiency could produce a global systemic risk. Computer systems had coded years to two digits to save memory, a shortcut that became a costly global threat to correct.
The predicted catastrophe ultimately did not occur. But it left a key lesson: when technology evolves faster than control mechanisms, technical debt becomes a strategic risk. Frederick Brooks already explained it in The Mythical Man-Month (1): accelerating software production without mastering the complexity of the systems mechanically increases the risks of maintenance, coordination and errors. In the age of agentic AI, this logic reappears with new intensity.
The mirage of “vibe coding”
Today, generative AI recreates a dynamic comparable to the emergence of vibe coding. This approach consists of developing applications by simply dialoguing with an AI, without always understanding the generated code. The developer, or sometimes the non-developer, describes an intention: “make me an API”, “fix this bug”, and the AI instantly produces a working solution.
The productivity gain is spectacular. According to Anthropic, more than 90% of the code of certain teams is now generated by the Claude Code AI (2). Players like Mistral AI are also developing platforms capable of automatically generating, correcting and testing applications (3). But behind this promise, there remain questions: how many developments generated today by AI will still be maintainable in two or three years? and what risks will this technical debt pose for companies?
The developer is now a digital trust architect
For decades, the developer’s value was primarily based on their ability to write code. From now on, this skill becomes partially automatable. The real value shifts to the ability to design, supervise, test, secure and govern the systems produced by AI. The developer of tomorrow will be less a producer of lines of code than an architect of digital trust.
However, producing code quickly does not mean producing reliable code. State-of-the-art software development is based on rigorous processes: unit testing, continuous integration, code review, security validation and performance monitoring. However, agentic AI upsets this balance. When a team can generate in one day what previously required several weeks, the temptation is strong to reduce the control steps.
The invisible risk: security and maintainability
The danger of vibe coding lies in the illusion of simplicity. Software can appear to work but have vulnerable dependencies, invisible errors, or difficult-to-maintain architectures. However, maintainability constitutes a pillar of software development. Code must be able to evolve, be understood by other teams and be corrected over time. The question is therefore no longer just: “Does the code work today?”, but: “Who will still be able to understand it and make it evolve over time?”
Several studies already show that certain AI assistants regularly generate known vulnerabilities: SQL injections, authentication errors or poor security practices (4). On a large scale, risk becomes systemic. A company capable of producing software ten times faster can also produce security vulnerabilities and technical debt ten times faster.
Rethinking digital skills
This transformation requires an overhaul of developer skills. Organizations can no longer train only coding specialists. Key skills become cybersecurity, test engineering, software architecture and AI systems governance.
As with the Y2K bug, the most resilient companies will not be those that automate the fastest, but those that maintain control of their digital infrastructures.
References
(1) Brooks, F.P. (1995). The mythical man-month: Essays on software engineering (2nd ed.). Addison Wesley.
(2) Anthropic. (2025). Claude Code and the transformation of software engineering productivity. Anthropic Research.
(3) Mistral AI. (2025). Codestral and AI-assisted software engineering. https://mistral.ai
(4) National Institute of Standards and Technology. (2024). Secure software development framework (SSDF) version 1.1. US Department of Commerce.
