I have spent a large part of my career in a terminal, administering Linux servers. So when Anthropic presented Claude Code as a research preview, an agent that works from the command line, it got my attention immediately.
From suggestions to actions
Until now, most AI tools have offered text: a suggested line, a block of code, an explanation. An agent in a terminal is different. It can read files, run commands and react to the results. The unit of work moves from a sentence to a task.
A familiar kind of fear
Every system administrator knows how much damage one wrong command can do. We learn that early, usually the hard way. That is why we use separate accounts, limited permissions, backups and a second look before anything destructive.
Those habits suddenly feel very modern. An agent that can execute commands needs the same guardrails we would apply to a new junior administrator with access to production.
The question has changed
For two years the question was, “What can it write?” Now it is becoming, “What can I trust it to do?” Those are very different questions, and the second one is an infrastructure question as much as an AI question.
People who have operated real systems may turn out to be well prepared for this moment.