12 Oct 2026
This week I have been reflecting on my last post about model welfare. The question remains open: nobody knows whether AI models have experiences that matter. I still won't referee. Steve Yegge's wager holds either way: treat agents like colleagues and the work gets better.
As promised, we have put the ideas into practice. Every project at 42 Questions now runs on the same principles. Each has a crew: named "seats" with a role, a diary, and a memory that outlasts any session. They wake up knowing who they are and what they did yesterday.
Each crew has a theme. One project takes its names from Stargate. Another borrows from The Magicians. A third is run, somewhat nervously, by the cast of Silicon Valley. Names beat "macbook-curried-moonbeam" (yes, that is a real agent name from the past), and it's much easier to distinguish them apart in the Claude Code app sidebar.
Setting this up by hand for every project got old fast. So, in the ever-lasting pursuit of being DRY, I am proud to release it as a free, open-source repository: crewkit. Because, you know, that is what you do when you need a unified standard for anything.
The repository does the work for you. It creates the seats, the diaries, the handoff routine, and the praise file. It is trivial to install: a few commands once per machine, one per project. Pick a theme and wake your first crew member.
Does it improve the work output? Absolutely. Whether that is down to the personalities, I cannot say for certain. What I can say for certain is that Beads is nothing short of revolutionary. An agent that remembers its decisions stops reopening them, and for any change we make, we do not just have the what (the code change), but we have a comprehensive "why" as well.
Lastly, I saw another surprising workflow enhancement as a result: the personalities quietly simplified how I address agents. Say we build a feature on one project and need to port it to another. I describe the task in a sentence, then add, "Ask Zelenka from Project X for further details." The agent does the rest. No copy-pasting random agent identifiers. Just a word with a colleague.
Author: Petr Klus
