AI Coding Agents: Real Utility Alongside Real Risk
Two threads today pull in opposite directions on AI coding. A team vibecoded a fuzzer in a weekend and used it to find a real division-by-zero bug in FFmpeg, a genuinely impressive result that would have taken a skilled engineer days of setup. Separately, a decompilation project finished an N64 game in 84 days by leaning heavily on LLMs for pattern recognition work, cutting what would have been a multi-year project.
The counterpoint: the FFmpeg thread immediately drew pushback arguing that AI-generated code quietly introduces bugs faster than it finds them, and that the net quality impact on codebases is negative when you account for subtle, hard-to-find regressions. The Claude Code thread has older engineers reporting that it re-ignited their passion for building, while others found it deeply unfulfilling to watch an AI grind through their backlog.
The pattern is not that AI coding is good or bad. It is that the variance is high. Expert-guided use on well-defined tasks produces real results. Unsupervised generation in large codebases produces risk that compounds quietly.
So what?
Use AI coding tools as force multipliers on well-scoped tasks with clear correctness criteria, not as autonomous agents on complex production systems. The FFmpeg result shows the right model: generate a specialized tool, let it run, then review what it found. That is very different from letting it touch your core application logic unsupervised.