Anthropic builds AI tools that write code, and the company says its own people use AI every day. Yet its published hiring guidance says candidates in a live interview may not use AI unless the interviewer allows it for that exercise. The company draws that line at the exact moment it judges someone’s coding ability.
For many engineers, writing code by hand has become a smaller part of the job. An AI agent, a program that writes code by itself, often does that part. The engineer’s work happens before and after it. First the engineer tells the agent about the existing system, its limits and what a correct answer looks like. Then the engineer checks the result closely enough to trust it in a live system. If either step goes wrong, the agent can hand back code that runs without being right.
Code like that might miss a rare case nobody tested, work only for the example it was given, or quietly break another part of the system. Catching the problem depends on already knowing what solid code looks like. Someone who cannot spot a weak solution cannot tell whether the agent’s version is safe to ship. That judgment usually comes from years of writing and fixing code without help. The agent can produce the code, but it cannot vouch for it.
So judgment and review are what separate a strong engineer from a weak one. An interview without AI tests an older version of the job, from before agents took over most of the typing. It measures how fast and how correctly a person can write code alone, starting from a blank screen.
One likely reason for the rule is what an interview can measure fairly. Directing an agent and reviewing its work is hard to judge the same way across dozens of candidates. The result depends on the agent, the code and how much the agent got right on its own that day. Watching someone solve a problem alone is easier to compare, because an interviewer can follow the reasoning directly. Anthropic does not appear to have found a consistent way to test review skills.
The company says it expects to keep updating the approach as AI tools get better. As the guidance stands, the default for anyone walking into an interview is the room, the problem and nothing else.
