AI coding is getting cheaper. So why aren't we shipping faster?

Atlassian recently shared an interesting finding: AI usage among engineering teams increased by 65%, but overall developer velocity didn't. Coding agents can generate in minutes what once took hours, yet writing code is only one part of shipping software.

The hidden cost often comes later: rework. When requirements are vague or constraints are missing, AI can make assumptions and start implementing them at speed. What looks like faster development can quickly become another cycle of reviewing, correcting, and regenerating.

That's why the next big optimisation in AI development may happen before coding starts.

It's also the thinking behind Spec-Driven Development. At MySpec, we help turn an initial idea into structured requirements, solution design, and implementation tasks before an AI coding agent starts building. The goal isn't to slow AI down, but to give it clearer context so there's less to undo later.

AI-generated code will keep getting faster. But maybe the real bottleneck isn't coding anymore. It's clarity.

For those building with AI coding agents, where do you spend more time today: generating code or fixing what gets generated?

Source: atlassian.com/blog/company-news/ai-sdlc

MySpec is during beta, feel free to try: myspec.dev