AI writes code like a keen intern, so now I mark its homework.
I run eight production systems, and AI agents write a fair share of the code. They're quick, eager and sometimes confident about things they haven't checked. Here's how I keep them honest: one mechanism a post, drawn on a single notebook page, twice a week.
An AI agent can disable its own monitoring if the checks live in its process: here is how an external supervisor with two hard timeouts actually works.
When an AI agent holds more access than its task needs, a breach costs more than it has to. Here is how scoping credentials to each task limits the damage.
