Addy Osmani works on the Claude Code team and writes code with AI every day. He recently said something very heartbreaking: AI agents do too much work for him, and he rarely needs to use critical thinking.
He divided this process into three stages.
The first stage is called cognitive offloading, handing over the task. This is normal, this is how calculators and search engines come about.
The second stage is called obedience, where you begin to directly accept the answers given by the AI without verifying them.
The third stage is called cognitive surrender and complete trust in the AI.
The general idea of his original words was: Skills you no longer practice will disappear, and it may be faster than you think.
These words come from the mouth of a person who makes AI programming tools, and the weight is different.
People who write code should be familiar with the second stage. The agent changed twenty files, and the test was all green. Just glance at the diff and it is ready. It's okay once or twice, but after a year, your understanding of this code base may no longer be as good as that of the agent.
Osmani's own open source set of agent skill packages specifically includes specifications such as /spec in the process, forcing people to think clearly about what they want before taking action. It can be seen as a level he left for himself that he had to use his brain.
The tool doesn't decide for you where to stop.