Can LLMs Generate Enterprise Quality Code? — Prasenjit Sarkar, Sonar
TL;DR · AI Summary
While LLMs achieve high functional pass rates (e.g., Gemini 3.1 Pro at 84.17%), Sonar’s evaluation of 4,444 Java tasks reveals critical maintainability and security flaws—614 bugs per million lines, verbose code, and high cyclomatic complexity.
Key Takeaways
- Gemini 3.1 Pro achieves 84.17% pass rate on SWE Bench but generates verbose code
- Sonar’s framework analyzing 4,444 Java tasks found LLM-generated code has 614 bu
- Current LLMs overlook engineering discipline; enterprise-grade code requires hum
Outline
Jump quickly between sections.
Developers widely adopt AI agents for coding, yet question the maintainability, security, and readability of generated output.
LLMs score >80% on benchmarks like SWE Bench but ignore critical dimensions such as security, architecture, and engineering discipline.
Sonar analyzed 4,444 Java tasks and found LLM-generated code suffers from high bug density and technical debt.
Despite 84.17% functional pass rate, it produces verbose code (307K lines), high cyclomatic complexity (234), and 614 bugs per million lines.
Human review combined with static analysis tools like SonarQube is essential to ensure LLM output meets engineering standards.
Mindmap
See how the topics connect at a glance.
查看大纲文本（无障碍 / 无 JS 友好）
- LLM能否生成企业级代码？
- 现状：AI代理普及
- 55%开发者日常使用
- 人类仍需审查
- 评估缺口
- 仅关注功能通过率
- 忽略安全/架构/可维护性
- Sonar实证研究
- 4,444 Java任务
- Gemini 3.1 Pro：高bug密度
Highlights
Key sentences worth saving and sharing.
55% of developers now regularly use AI agents for coding, but humans still review the generated code.
Gemini 3.1 Pro scores 84.17% on SWE Bench but generates 307K lines of code with cyclomatic complexity 234 and 614 bugs per million lines.
LLM evaluations often focus only on functional correctness, ignoring security, architecture, and maintainability—key enterprise criteria.