Nielsen Norman Group (NN/g) researchers have identified a critical failure pattern in the current wave of agentic AI adoption: the rapid emergence of Big Balls of Mud. In a recent study, the UX consultancy observed that non-technical users, dubbed vibe architects, are building complex agent systems that become unmaintainable messes. The core issue is not technical incompetence, but the removal of traditional scarcity constraints. Because AI makes code generation nearly free, users overbuild without architectural discipline, creating systems that work until they suddenly don't.

The Return of Code-and-Fix Culture

The term Big Ball of Mud originates from a 1997 paper by Brian Foote and Joseph Yoder, describing software patched together for short-term functionality rather than long-term durability. NN/g found that vibe architects are unknowingly retreading the history of software engineering, moving from agile planning back to a chaotic code-and-fix mode. One advanced participant noted that while AI makes building easy, it leads to overengineering where flashy tools deliver little additional value. Without the friction of high development costs, there is no forcing function to prioritize simplicity or maintainability, leading to systems that are expensive to diagnose and impossible to reconstruct.

Five Patterns of Agentic Decay

The study identified five specific patterns where these systems degrade. It begins with Throwaway Code, where users create untested rules.md files that quickly become essential dependencies. This is followed by Piecemeal Growth, where simple widgets evolve into complex dashboards without a release cycle, often growing beyond the user's comprehension. As systems break, users enter a Keep It Working phase, patching issues with prompts rather than fixing root causes. This leads to Sweeping It Under the Rug, where data is dumped into unorganized folders that only the AI can navigate. Finally, the system reaches a point of forced Reconstruction, often triggered when the user outgrows the tool or loses the mental model of how it functions.

The Illusion of Competence

The danger lies in the AI's ability to compensate for poor architecture. Participants relied on the AI to locate files and explain their own systems, creating a dependency where the user no longer understands where information lives or how the agent uses it. One participant lost 45 minutes fixing instructions because the AI retrieved the wrong articles from a folder called to be synthesized, highlighting the limits of this compensation. The study warns that while AI lowers the barrier to entry, it also hides the accumulating technical debt. Users feel productive because the system responds, but they are increasingly fragile and dependent on the AI's ability to navigate the mess it helped create.

Key Takeaways

  • Separate Context Types: Maintain distinct global, local, and ambient contexts to prevent architectural blur. If you cannot identify which context type needs an update, your system is failing.
  • Audit Global Context: Review global instructions regularly. If the global context is too large to read in one sitting, local specifics have likely leaked in, degrading performance.
  • Name Your Changes: When prompting the AI, specify whether a change applies to global or local context to retain control over the system's logic.
  • Reconstruct Before Collapse: The goal is not to avoid mud, but to know when the AI can no longer navigate it. Periodic reflection on what you would rebuild is more critical than tidy folders.

The Bottom Line

The era of vibe coding is just spaghetti code with a better autocomplete. Until builders learn to treat context libraries as architectural artifacts rather than disposable prompts, agentic AI will remain a house of cards waiting for the first strong wind.