Maintainability sensors for coding agents

· Source: Martin Fowler · Field: Technology & Digital — Software Development & Engineering, Artificial Intelligence & Machine Learning · Depth: Advanced, extended

Summary

This article details an experiment using "maintainability sensors" to improve AI coding agent outputs, focusing on an internal analytics dashboard rebuilt with AI (Cursor, Claude Code, OpenCode). The author implemented various computational and inferential sensors across the development lifecycle, including during coding sessions, CI pipelines, and scheduled reviews. Key computational sensors like ESLint effectively addressed AI shortcomings in code structure and complexity, especially when paired with custom self-correction guidance. Dependency-cruiser enforced layered module architectures. While raw coupling data proved noisy, LLM-driven modularity reviews, using "Modularity Skills," successfully identified significant design issues like duplicate code and inconsistent patterns. Mutation testing emerged as a critical tool for evaluating AI-generated test suite effectiveness, uncovering assertion gaps despite high coverage. The findings suggest computational sensors are strong for local issues, but inferential sensors are essential for semantic interpretation of broader architectural concerns.

Key takeaway

For AI Engineers integrating coding agents into development workflows, you should prioritize implementing a robust sensor system to manage code quality and maintainability. Configure static analysis tools like ESLint with custom guidance for AI self-correction, and use dependency-cruiser to enforce architectural rules. Crucially, employ mutation testing to validate the effectiveness of AI-generated test suites, as high coverage alone can create a false sense of security regarding regression detection. This approach will significantly improve agent output quality and streamline your code review process.

Key insights

Maintainability sensors, combining computational and inferential tools, significantly enhance AI coding agent output quality and human review efficiency.

Principles

Method

Implement sensors across the development lifecycle: during coding, in CI, and on a schedule. Use custom lint formatters for self-correction guidance and LLMs for semantic interpretation of complex metrics.

In practice

Topics

Code references

Best for: AI Engineer, Machine Learning Engineer, MLOps Engineer

Related on AIssential

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Martin Fowler.