Why AI Needs Multiple Personalities

Key Takeaways

Okay, let's talk about coding agents. So, you you wrote about this idea that AI agents should take on multiple different personas, personalities. What what's going on here? Yeah, so essentially this kind of trend that we wrote about in this morning's newsletter is a new way that developers are using coding agents to get better performance out of them. Um and so I think maybe this is best explained with an example. And so, you know, in the past somebody might have given a coding agent a pretty high-level task like, you know, make a new login window for my app or something along those lines. Um now instead of doing that, people are basically spinning up, you know, multiple agents to take on different kind of personas that might, you know, pop up in every stage of this process. So, first they might ask one agent to act like a product manager um and create a doc that kind of outlines what they want this new feature to do. Then maybe they'll they'll tell another agent to act as a as a spec agent. So, basically turn that product doc into more specific instructions that really lays out kind of what technical work needs to be done. Then you might ask another agent to be like a tasks agent which turns that into specific coding tasks. Then finally a coding agent do those tasks. And then lastly a review agent to review kind of all of those different coding tasks that the previous agent just did. So, it's kind of all about like breaking up this, you know, larger kind of higher-level task into smaller ones and then assigning different agents each part of that task and telling them, oh, pretend to act like this type of role or person that would be doing this sort of task. Now, this is kind of interesting because I know that as it relates to chatbots, I mean, we've done a little bit of a report well, you've done you did all the reporting. I just I just talked about it here. But you know, I I recall you having written about chatbots taking different and this is not different personas, this is different personalities. So, you know, maybe you have an answer that's a little snarkier, an answer that's a little happier, maybe one that's a little more introspective. I mean, it is is the idea here not so much that that it's more like segmented by roles or is it that the coding agents might actually have different ways of of giving you the product or maybe it's I don't know, snarky or something like that. Yeah, so I think there's like kind of two different like related concepts here. So, I think first there's this idea of um AI chatbots or just, you know, AI tools that people talk with that have different personalities. And so I think part of that is kind of uh you know, for consumers for instance who might have a certain preference for a type of personality they want their AI chatbot to have. So, maybe I want, you know, ChatGPT to be more concise or more funny or more kind of straightforward with me. Um and then that's a little bit different from, you know, what we're talking about here with coding agents where it's less like personality and it's more like pretend to be a product manager, pretend to be a, you know, uh senior engineer. Um there is like some overlap though. So, we saw for instance XAI whenever they first came out with their Grok 4.2 model, um a lot of developers pointed out that whenever you looked behind the scenes, there were kind of these four different agents that were running all at the same time. And so, you know, one of them would kind of be more of like this contrarian agent, which some might say is kind of a personality trait, but basically its main point was to kind of push back on anything that the other agents said. So, you know, question things that the other agents wanted to do or bring up maybe alternative paths that the other agents hadn't thought of and that kind of, you know, made the the end product better. And so I think all of this is like very related and yeah, definitely very interconnected here. So, what do you think this development and this trend says about the usability of AI overall? Because on one hand, I see the argument that, hey, I I need to structure out my thinking. I would need a team with five different people. It's easier for me to think about what each different agent would do. On the other hand, you could make the argument that it's taking more instruction. I, you know, I'm having to do more of the thinking. I thought the agent was supposed to do all this for me. Yeah, I mean, it's a really tough question. I think there's like two camps which which you kind of laid out there. I do think that, you know, on one hand it feels a little bit silly to, you know, be in a world where people were talking about superintelligence and AI that can help us colonize Mars and, you know, educate every child on on Earth and then you're meanwhile you're having to tell like an agent like, oh, pretend to be like a paranoid, you know, engineering manager that questions everything that it's been told or something. Like that just feels a little silly. Um I think at the same time though, uh some people would kind of point out that, you know, if you were working with humans, you'd have to do the same thing, right? Like my, you know, my editor can't send me one sentence over Slack like, do this story. Like they I think it, you know, there's a lot of back and forth that happens where I'm talking through like a story idea with my editor and breaking things down. So, I think that process is something that you've been have to do with with humans. Um but I think a lot of researchers are saying that this is kind of a, you know, band-aid solution for now, but in the future they expect the models to be able to do this sort of thing by itself. So, instead of the human having to be like, okay, here's this you know, there's this broader task like let me break it down for the model into steps 1 2 3, the model itself will be able to do that kind of division of tasks and then assign each of those tasks to the agent without the human having to, you know, interfere too much. So, researchers are saying that that's in the future, but this is kind of, you know, a band-aid to deal with where the tech is right now. Great. Well, Stephanie, I want to thank you for coming on. That is Stephanie Palazzolo, our AI reporter here at The Information.