A software engineer spent 18 months testing AI coding assistants like Copilot, Claude Code, and GPT 5.2, documenting their evolution. The tools have improved in generating idiomatic code and bug hunting, but still struggle with complex frontend development and can degrade a developer's mental model of the codebase.

The author argues we're in a Cambrian explosion of programming, where AI is shifting software engineering from pure coding to a team sport reliant on human oversight. The key is treating AI as a pair-programming buddy to be shaped and guided, not a replacement for human judgment.

Moving forward, the industry needs better tooling to help engineers maintain system understanding and train on AI failure modes, similar to aviation's pilot training. The core lesson: AI excels at known problems and debugging, but product quality still demands robust human skills and scaffolding around the tools.