A developer who has followed the rapid spread of AI coding agents is calling for a slower, more disciplined approach to using them in production software. In an essay published on his website, Mario Zechner argued that the systems make it easy to generate projects and large amounts of code, but can also amplify weak design choices and recurring defects.
Zechner distinguished between experimental personal projects and software that people depend on. He described coding agents as enjoyable tools for building ideas that might otherwise remain unfinished, or for exploring an unfamiliar technology stack. In those settings, maintainability may be a secondary concern and the consequences of failure can be limited. His concern is the transfer of the same high-speed workflow into production codebases.
The essay’s central argument is about scale rather than the existence of errors. Human programmers also introduce defects, duplicated logic and awkward abstractions, Zechner wrote. But an individual developer can add only so much code in a day and may eventually respond to the resulting pain by repairing it. An orchestrated group of agents can produce far more code without experiencing that feedback, allowing individually minor problems to accumulate quickly.
According to Zechner, attempts to steer agents with instruction files, memory systems and documented lessons can reduce particular categories of mistakes. Those controls still depend on a person first noticing the problem. If design decisions, reviews and implementation are delegated together, he argued, teams can lose an accurate understanding of the system while automated tests generated in the same process may offer false reassurance.
The essay also criticized workflows centered on coordinating large fleets of agents and maximizing the amount of code produced. Under that model, Zechner contended, developers can become detached from how features are designed and implemented.
Zechner presented these conclusions as an experience-based view, not a measured industry study. He referred to anecdotal reports from peers and to visible quality problems in software, while acknowledging that outside observers generally cannot see companies’ internal development practices. That distinction matters: the essay does not establish that coding agents caused any specific outage or broad decline in reliability.
The practical warning is therefore narrower than a rejection of AI-assisted programming. Zechner’s case is that output volume should not replace engineering judgment. Human review, deliberate architecture and direct testing remain essential when software affects real users or valuable data. The speed offered by agents, in his account, is useful only when teams retain responsibility for understanding what is being built and for controlling the complexity that the tools can create.
