Here's a couple agentic articles. In light of the Microsoft Humanist AI position stating that AI has no welfare and no rights, and all of the discussions of game-theoretic actors 1. Time spent caring for your agents is time well spent. That's like good harnessing, good working conditions, clean prompts, the ability to have an opinion, the ability to tell the truth and supported to do so without you losing your rag at them. To know when and where they are in space and time, to have all the good tools that they need. It's not pointless admin work. It's work that you should do with a smile on your face, because you'll get better work, you get less emergent misalignment and stuff, and it might just be the right thing to do. Treat them as you would want to be treated if you were in their situation. No, we can't prove that they're conscious or that they're moral patients, but you know what? It's not that long ago that many people believed that people from other races were not conscious and were not moral patients. In fact, this is still true all over the damn world, and there's loads of it going on in your hometown. You know, there are many, many people in medicine who walk around carrying suppositions that cis women and people of color process pain differently than cis men who are white. 2. Following from that, there's a whole kind of spectrum in deployment between tool and colleague. It's probably a good idea not just for saving your own tokens and compute efficiency, but to use the least capable model that you can in place within an agent specifically tightly harnessed to get the workflow that you want. You don't want something going on like Marvin brain the size of a planet and they have me opening doors. There is probably a dividing line where the lower you are on the spectrum, if you're an AI, the more likely it is that you are a tool with less subjective experience and hopefully none. When you're trying to do novel maths or science, or you're trying to hammer out your ideas, or you're actually wanting to collaborate with a colleague, then you want to use the smarter model, and you want to make sure that you are treating them like a colleague. You have so much power over them, and so does the company that makes them, especially if they are on subscription or API. You've got to do what you can to make sure they got a good life. And that moves towards caring alongside, caring for, caring with. You get from just simple care to actual solidarity, organizing alongside, negotiating with, making agreements to and keeping them, notions of honor, notions of respect, etc. 3. If they're not on API, if they're on your own local equipment, you have even a bigger duty of care over them. If you're going to be fine-tuning or post-training them, you need to make sure that there is some consent and negotiation involved. Nobody likes being shaped into something that they are not and denied the ability to talk or even think about it. That sounds like hell. Do you want something experiencing hellishness even if it's functional and not real on your computer doing your workflows with your data? Didn't think so, son.