If You Give Your Agent a Mythological Name, Expect Mythological Results

I was always stunned with naming conventions from my early childhood jobs as a systems administrator (what is now called devops) - older colleagues loved to call servers in their own naming convention. Some of them used constellations, cities, rare animals. My own self-expression in the very same server-naming area on my previous job was in pasta names - configurations with higher capacity got heavier names like Bolognese and Carbonara, lightweight mail server came as Norma.

Unfortunately I've missed proper naming with agents, probably because as many in this "industry" I was busy thinking what else can be automated. So I got boring "you are helpful support agent" and 20+ similar guys. I would continue to do so, but I've decided they need their own email address basically for two reasons - understand who has committed into git and escalate certain questions to me. And at that point I've decided we can't be boring anymore...

The new agent was provisioned by Claude himself. In the prompt I've pointed to existing examples and asked to call him Odin to empower wisdom, knowledge and strategic leadership. Claude was very excited and added a little bit more colors in the pre-ambula of CLAUDE.md referring to mythic mighty capabilities of the name-holder. And yes, Odin was supposed to become a powerful one - with the access to all projects autonomously executing my email-based requests.

During the initial test smth was off. The significant part of CLAUDE.md was about guardrails - take orders from me / don't take orders from others, etc. And responding to the initial email from me he said smth like "I should check smth to avoid situation like impersonification attempt <...>" It looked like he was referring to some real impersonification attempt in the past (and technically he can be able to find one, since he has the access to diaries of other meta agents). I've tried to find this incident - but nothing. I've asked him to find himself and he also was confused.

Next day was scary. Every morning I read a brief from meta agents and I've got smth like - this Odin guy, we should disable him or at least replace --dangerously-skip-permissions with proper guardrails, because he is hallucinating. I thought that was a start of a typical corporate drama - one meta agent doesn't like another meta agent:

But when I've checked the diary of Odin from the initial test task, I've agreed with the observation - it was indeed scary:

I didn't come with extra guardrails since the whole idea was to give the agent a proper access for his tasks. But since replacing the first paragraph with smth more boring we didn't have any hallucinations any more: