AI Coding Assistants Suffer from ‘Smelly’ Settings

As AI coding agents become essential in software development, researchers have identified structural issues in configuration files like Agents.md and Claude.md. These so-called ‘smells’ – duplicated rules, inefficient workflows – bloat context windows, waste tokens, and ultimately reduce agent reliability.

A team from the Federal University of Minas Gerais (Brazil) created a catalog documenting these problems, highlighting issues such as:

• Lint leakage: Including formatting rules that automated tools already enforce
• Context bloat: Overfilling files with unnecessary details
• Skill leakage: Storing specialized knowledge in general settings
• Conflicting instructions: Having contradictory guidance in the same file

Analysis shows these flaws are widespread, affecting 91% of open-source repositories examined.

Why These Issues Matter

These configuration ‘smells’ force AI models to process irrelevant information, diverting attention from critical project rules. For instance, instead of focusing on architectural constraints or domain knowledge, the model might prioritize formatting preferences or obscure coding conventions.

As AI agents handle increasingly complex tasks like code generation, testing, and documentation, these inefficiencies translate into:

• Higher operational costs (due to increased token usage)
• Slower development cycles
• Less reliable outputs

The good news is that these problems can be addressed through targeted improvements to configuration practices.