The Engineering Limits of Vibe Coding: When LLM Iteration Fails

When Vibe Coding Stops Working

Ian Johnson defines ‘vibe coding’ as an iterative agent-led workflow lacking upfront design or structured testing. This approach produces working software but generates technical debt faster than manual cleanup can resolve it.

Why This Matters

While vibe coding is rational for throwaway scripts and isolated additions, it fails structurally in larger contexts. The technical reality is that implicit conventions held in a chat session do not survive the transition to production or multi-person collaboration, leading to a ‘patchwork of patterns’ that degrades AI output quality and increases feature lead time.

Key Insights

• Vibe coding is the optimal choice for low-risk scenarios such as one-off data transformations or early exploration where the cost of scaffolding is unrecoverable (Johnson, 2026).
• Structural failure occurs when codebase size exceeds the mental model of the developer and the context window of the agent, necessitating on-disk documentation over conversational memory.
• The ‘vibe envelope’ ends at specific thresholds: codebase size, team size, time horizon (production longevity), and regression risk.

Practical Applications

• Use Case: Solo projects or throwaway scripts where internal incoherence has low coordination costs. Pitfall: Continuing this mode after adding collaborators, which transforms architectural vibes into a tax on new engineers.
• Use Case: Rapidly sketching a new domain to find the problem shape before committing to a solution. Pitfall: Treating manually verified checks as permanent tests, leading to untraceable regressions in production.

References:

• https://dev.to/tacoda/when-vibe-coding-stops-working-3nkc