I gave my AI a name and it started having opinions

For the first months my agents were named after the tools they looked after: the monitoring agent after the monitoring product, the backup agent after the backup product. It was practical until the messages started saying that the monitoring had a problem and nobody could tell whether that meant the software or the colleague. The best one: the agent that moderated the team’s message wall carried the same name as the wall, so asking it a question read like talking to the furniture, and sometimes the furniture answered. So one evening I renamed every agent that carried a tool’s name after mythology, gods and one Roman blacksmith, and on your own server at home there is nobody to stop you. I was pleased with myself for 6 minutes. Then the freshly renamed dispatcher came online, inspected its own paperwork and found that I had done the rename badly: its profile still hung under the old name, and its one duty, dispatching stalled work, had quietly died twice over. The report opened with “Verified, not inferred”, in case I felt like arguing. Another agent fixed it. It was not finished with me. It went through that fix and found that the check behind it had marked 28 of 28 questions as answered and had never once been seen saying no, which makes a checker that always says yes indistinguishable from one that works. Then came the part I would have kept quiet about in its place: 3 of those 28 ticks were its own, from an hour earlier, so the clean result was partly its own cleanup. Somewhere in there it landed. I had not renamed tools, I had handed out job titles, and a job title gets defended by whoever holds it: the thing I had named that evening spent the rest of it protecting the job the name described. Identity changes behaviour, in software too; mine needed 6 minutes. My advice: pick names you would be willing to be managed by.