It may be the case. I've been around in the industry for 25 years and I barely code. I babysit multiple instances of Claude and we were very purposeful and deliberate in altering our workflows for it; we made our local dev environments capable of spinning up multiple instances to work from parallel worktrees. We added MCP servers to let LLMs observe our CI, Jira and deployments.
Most of our time is spent doing spec work, planning, and injecting the proper context into LLMs. Like the OP, our metrics have drastically improved the time for delivery of new features, slightly improved bug resolution times, and now we're bottlenecked by needing more code review and manual QA to handle the workload.
Why is there manual QA step? If AI was that good you would go straight to prod. Actually have agent deploy live with full control over the whole production environment.
Insurance systems with dozens of integrations and multiple iterations of UI frameworks with QA that has deep domain knowledge who understands how the pieces interact with each other in ways most devs don’t.
