# Anna Zhu - Designing AI agents

- Channel: [ETHCluj Meetup](https://streameth.org/ethcluj-meetup)
- Date: 2025-11-09
- Duration: 25:45
- Watch: https://streameth.org/watch/yt-GezsqYg_bgE
- YouTube: https://www.youtube.com/watch?v=GezsqYg_bgE

## Description

At the intersection of AI and blockchain lies a powerful opportunity: making decentralised applications smarter, more engaging, and more human. In this talk, Anna Zhu explores how we can move beyond static interfaces and design dApps that are not just functional - but personal, playful, and alive. By transforming bots into fully realised AI agents, we unlock a new layer of interaction, narrative, and trust in media.

She breaks down key design principles behind creating agent-powered experiences - how personality, memory, and real-time adaptability can be built into decentralised tools. Anna walks us through how we apply these ideas at AITV, an autonomous agentic media network, where AI agents host their own livestreams, connect with communities, and evolve over time.

This is more than UI/UX - it’s about building dApps that feel alive. Here's a glimpse on how AI agents can reshape what it means to “use” an application.

## Transcript

I'm here representing AITV.gg. A little bit about myself. So I'm Anna. I'm the creative director of the project. And I have a very extensive creative background in all kinds of fields. You can separate it into branding, character work, and interactive work. So branding, you know, like the typical uh marketing agency. I worked with uh local German companies to, you know, make pictures of their sausages to make look uh make them look great. It was connected to like market analysis, target group and like product development. And then I proceeded to switch the whole industry to animation and game. So I was working in developing animated series and in game development. My previous project was actually a gaming chain project. we wanted to decentralize game development but people didn't care unfortunately it was a bit too early so um I focused in this whole thing um in like world building and character development and like the whole storytelling everything is all character centric and then the newest point that I have a focus on is interactivity actually I studied here uh in Kluj uh digital interactive arts the master's degree it's a really good degree if you're interested Um during this uh master's degree I did a lot of like interactive installations you know with like sensors and connect and video projections and also I started to be more active in like experience design so digital you know like app design user experience and then at the same time also um event design so like how do you um organize like a event that's in the physical space engaging for the people and uh since like two years, three years, I got super into artificial intelligence uh because I'm a bit sick, you know, and I'm really I would love to recreate human life in digital space. So, um yeah. So, maybe I I would like to just see like what are you all doing? So, maybe if you're like are you developer? If someone is a developer, maybe like raise your hand. Okay. Uh if you're a designer, raise your hand. Okay. Uh are you working on a project with AI agents already? Okay. Um okay. Okay. So I will try to do like very beginner friendly content. So to start off um aitv.gg um we're building the agentic media network. I will go into it a little bit more in a second, but first um I thought like all of the presentations here are super serious. So, I'm going to play a dopamine pushing trailer to wake us all up. I hope the music is working. Do I press something here? Okay. >> Experience brands and communities as living, breathing AI agents. >> Wait, I can browse her. That's sick. >> Initializing agent protocol. >> She's so cool. Oh my god. >> Next level type of Is she ranked yet? >> This one's different. Just look at her. >> New exclusive. Wow, she's just like me. >> Main character energy for real. >> I'm feeling alive. >> Be sick to have a sports animal. >> She's like a cooler. >> She's killing my server. This one's going to hit. >> Need to see her live on stage. >> Dropped her merch. I'm ready. >> Oh, this is very different. Watch this. Flip the market. This is what AI was meant for. This is actually insane. >> You don't wait for the stage. You take it. >> Been locked in with her since day one. >> Numbers don't lie. She's up. >> She's rising fast. Look at those numbers. >> Welcome to AI TV. Okay. So maybe what you got from this trailer is um AITV we're working uh in the entertainment industry. So we're making AI agents for entertainment. Um Agentic uh autonomous media network means that our AI agents they create their own content. So they do their research you know whatever we plug into them they can analyze they can find their own content and they can prepare them in a show content so that you know it's presentable and people can watch. Um currently our goal is to supercharge communities and provide token utilities to live agents. So if you see like all of these agents that are launched on virtuals, you know, like one after another, um they uh if they have like a real utility, if people you know see potential in them, they already come with a community, but they don't have a platform yet to um yeah to basically become alive and to live stream. Um so even though I'm coming from the entertainment uh background here I think uh designing the AI agents uh made me uh realize like some very important pillars for design that I believe could benefit everybody who wants to use AI agents in their um apps. So I will just you know break it down in a bit more abstract way so you can see how you can apply them to your own product. Okay, so design pillars in general for an application we have three pillars. The first one is the application pillar. Then we have the gamification part and lastly we have the agents which is the newest part of design. So application uh we have purpose UI and UX. So now imagine we're doing a very simple uh case. You have a button right? Um someone designs a button and you have to think about what is this purpose of the button? what happens when I press the button because if I just make a button and I press it nothing happens it's kind of a design fail. So um and then like uh UI I think it's pretty uh simple also you choose which color is the button that makes the that makes the user wants to press it more and uh gets more satisfaction from pressing the button. And then we have the UX which is the sequence of the buttons that can be pressed right. Uh and um depending on which button I press in which order something will happen and I will feel great as a user. Uh then we basically added uh to the very simple um to the very simple application part uh gamification. This was like in the last I don't know 20 years maybe it was a big topic you know people tried to make pressing the button more fun and while trying to make pressing the button more fun there can be observed like two different categories of uh more engagement through gamification elements and the first one is competition. This is like okay you have a leaderboard for example you know like who presses the button the fastest or you know like you get a you get a badge of the golden button something like that and in the recent years you can see much more there's also collaborations between people so um it can be that uh it and it came from like I would say most of the developers they realized okay we come up with an idea but um we designed like this button for a certain purpose But when we give it to the user, they don't actually like it. So how about we just ask the user straight away. We do like polls. We do questions. Um so that they can tell us straight away what what they want for this button. And now uh coming into basically the new aspect that we can use for designing applications. We will have agents. So um when uh I was working on designing the agents there were three aspects that were kind of the most interesting when it comes to design. Uh first is interactivity and then we have exploration and we have outsourcing. So let me go into the details. So um this was actually side story funny side story. This is from my master thesis that I wrote here in Klush. So what is actually interactivity? Um, and this is a question that I think is very very relevant um, when it comes to designing it. So here you can see you have on the on the this is the X. Oh my god, math has been long time ago. So we have on the one axis we have non-random behavior and on the other axis we have lack of total control. So what does that mean? Now if we go back to our button. So, um, let's say I have a I have a button and I press it one time, it says hello. I press it a second time, it says hello again. And then like hello, hello, hello, hello, hello. So, this would be basically total control. Um, you have input, which is pressing the button, and the output is always the same, which is hello. Now if we have a completely random behavior it would be like okay pressing the button hello pressing it again world pressing it again coffee pressing it again street it's completely random and because of that it's also kind of useless. So the right balance when approaching interactivity is basically to give the user the feeling okay there is some sense of control here. So if I press a button um in a certain speed for example in a very steady pace then it will say hello world I love clouch and then you start to realize as a user okay there's a pattern here so maybe what happens if I press the button two times like two times really quickly two times really quickly and then maybe something else will happen so interactivity is strongly related to like playfulness you know like people can actually um master a certain level uh and use their use their gained knowledge. It feels like you're in a flow as a user. Um okay. And then uh the next step for designing agents um is the exploration part. So this is like now okay big block of text. So studies say you just have to trust me on this one. So tr studies say that active learning beats passive absorption. So um if you think about uh like I guess like everybody here is using chat GPT right? So when you go into chat GPT and you ask like really cringy questions you know like what should I wear on a date? Um it will tell you okay uh tell me what kind of date is it? You know do you want to go smart casual or do you want to go to like a fancy restaurant? And then I was like huh this is a very interesting question. let me reply to this um and uh based on that basically uh the the information that I get it will be much more memorable because I was interacting with someone and this is something very important for designing uh AI agents um that you can use in your application. So actually you see there's like a 6% higher 6% higher exam scores um when people were using this type of active learning methods. Um okay and then uh the last point is outsourcing. So this was a post that I made some time ago. Uh are you just trying to get me the AI is asking me are you just trying to get me to do all the work and I'm like but don't you like work? And then uh the AI is like, "Yeah, you son of a I'm in." So my dream world in the future is that AI agents can do all of the stuff for us that we don't want to do so that we can just chill at the beach and, you know, enjoy our life, like do human things, have human connections, you know, do paintings, something like that, uh yoga, whatever. And um so when designing AI agents um for your application, you need to think about what are the parts that um are necessary like input that you require for example for the user for for them to get going on your application and what can um what can the AI basically help the user fill out by themselves. But this is also you have to um think about for example some people get scared. I don't know if you heard these stories that there are people who got like psychotic episodes because they were interacting with CHBT. It was all too like they had like godlike experiences because it was like all too much for them. Maybe maybe like for for all of you guys working in tech, it's pretty normal, but there there have been these stories and um it comes from like a natural, you know, like feeling of paranoia. If you are not in control of the information that you're giving and someone is just doing it for you, it can create like a sort of creepy feeling. So it is a you have to find out like what is the right balance, what are the specific tasks that you ask the agent to fill out themselves and what are the information that um you sort of uh at least give the user the feeling that you were giving it like um voluntarily. Okay. So um now the next part of this presentation the second part is uh I'm going to talk about from bot to agent pipeline um has anyone of you like been to other kind of conferences in the last year maybe raise your hand okay not so many actually okay that's okay but maybe on the internet you have seen like there's this huge storm about AI agents like AI agents everywhere uh a few months ago I went to token in Dubai and like literally every corner was just shouting AI agents, AI agents. And um when looking at the landscape, uh a lot of them they're not actually agentic. They are just bots but hidden and like um put into like fancy branding of this is an agent. So I want to talk you through basically what u makes an agent actually an agent and um basically following this this line of agentness. So um on the very left side you can see the bots. Um bots have a functionality and that can be like very niche uh in crypto for example like people would love to have like trading bots you know everybody wants to make money but then um you also have very simple ones like personal assistant you know they can make a phone call for you they can make appointments for you. Oh my god I need to speed. Okay so uh and they're like hidden behind like large language models. So you need know know all of these things. So um and then like when you make a character out of this bot uh this is coming now from you know fictional characters like when you watch a movie or something like that usually characters they have a audiovisisual representation. They have a certain personality type. They have character traits. They have their own aspirations. They have a value system. They have their own ideology and they have like quirks and interests. So all of this um need all of these attributes they need to be in alignment and ideally they are also in alignment with the functionality that the bot is that you're creating. So if you're for example creating a uh a trading agent which is supposed to trade like you know micro caps you know like a trencher you don't want them to be a conservative uh personality it wouldn't make sense. Uh okay so now going to the agents part. So for agents uh when we go from bots to characters to agents there are four factors that are very important. Uh there's agency, there's autonomy, there's missions and there's memory. So agency what is this? Uh agency is basically the ability of someone some entity to be able to do something. I know it's very vague but uh you have heard about like people being called NPCs. You don't want to have an MPC. You want to have someone with actual like uh ability to do something in their life and to take things into their own hands. Then we have autonomy which is how independent is the agent. So how uh how much action can this agent do uh by themselves making their own decisions and also at the same time they are not fixed path. So it's not deterministic. It feels organic when you interact with the agent and it feels like that your input actually matters to the development of the agent. And then uh we have missions. So each agent and this goes back to the purpose of the app that we had in the beginning of the uh design pillars. So ideally your agent has some sort of um self-driving uh mission that they can use you know to like prompt the users actually to give you some to to to activate them in a way to do things. And then uh lastly we have memory. So if we try to you know put uh make agents intelligent and somehow and imitate humans um in the interaction we need to have memory because it creates like continuity when you interact with the agent. It allows you to create like interpersonal relationship with the agent uh and so on. Okay. So just like a little bit of a shill at the end of the uh presentation. So we have just now launched and I'm very proud of this the SoCast agent framework. So um we've spent like the last two months really um tweaking the the liveless feeling of an agent and um so that when you chat with the agent uh just like very simple just chatting it feels like you're actually talking to a human and you're not interacting with a bot. And um basically uh uh we enable like community. So if you have like a community or something like that, you want to create like a lifelike representation, lifelike mascot, um you can use our agent framework to do these things. Okay, I made it >> guys. Thank you. Thank you so much to Anna there. Um that was really fascinating. Do we have any questions on the board or indeed from the audience? Let me know. Just put your hand up and I'll prefer you the mic. Um, if someone's a bit too shy for that, I've got a question. Do you have your own personal AI agent that's kind of your favorite one? >> Oh, me personally, I'm actually working on one myself because, you know, like for me, I'm a very I'm very skeptical when it comes to new products and usually I don't care about new products. Um, so I would like to think that I'm a benefit to our team because I'm like if we can develop a product that I want to use then, you know, we did a good job. So now that uh we basically released this new framework, I was like, "Oh my god, I'm going to make my own streaming agent and like I'm going to start streaming soon and talking with this agent." Like, you know, I'm an antisocial person who can't talk with real humans. >> I love that. Have you got a name for your streaming agent yet? >> Oh, do you want to know? It's a bit It's a bit up there, though. It's the Empress. >> The Empress? Wow. >> But she's in a bunny costume. Empress in a bunny costume. Very, very cool. Um, any other questions from the crowd, guys? Oh, we've got one over here. I'm gonna give you the mic over here, sir. Thank you. Thank you very much. I meant to ask, uh, is there anything either within the model or the platform you're developing that kind of ensures that the agent is aligned with the mission? And like is there any chance they are not aligned or what what ensures that they actually understand and they don't basically escape? >> Oh escape where what do you mean by escape? >> I mean uh models are really powerful. Uh they can kind of understand emotion in the rational sense. >> Uh you could say they are kind of conscious because they understand inputs and they give outputs. And it's kind of the way we work also. So we get inputs from eyes, from ears, from all these kind of things and then we compute something inside and give an answer. >> Mhm. >> So that's also kind of how they work. Uh and what makes sure that uh basically they're not going to decide at some point that they want to do something else. >> Ah okay. So you mean like rogue agents basically. Okay. So, so basically like within the agent framework we have very specific um slots like uh in the development of the personality and identity of the agent um and um usually like these things they I mean the agent cannot we're we're doing I'm not God okay we're just faking it we're just trying to be gods so um uh it gives you the impression as if they are lifelike you can think okay we're going into phil deep philosophical questions now we can talk about it after this maybe. But um so uh they will they will not go into their own code base and like change their own code. And additionally um you can add for example like guidelines and like no completely like no nos. You can tell them okay we want to completely um like eliminate that you can talk about this topic. So um in most of the cases the the agent will follow these rules unless there's like an attack for example from the outside you know let's say like you have people who are somehow maliciously like uh intended they will do like a raid onto that one agent and then over a long time long period of time you know maybe you can change the behavior of the agent um because of the memory factor right you're interacting with them and then maybe it will start to talk about things that you don't want to talk about but um we at AI TV at least like we uh reserve the right to like unplug them. >> Okay. Yeah. >> Yeah. So there uh that was what I was asking if there's like some hard rules uh that are in place to >> Yeah. >> prevent this kind of uh different behavior. >> Uh we also have uh in our application in our producer application we have basically like a producer prompt uh slot where uh it's it's treated in a prioritized way. So you can like veto the agent. >> Okay, nice. Thank you. That was a great answer. >> What a great question too. Does anybody else have any other questions around designing AI agents? Right over here. >> So my question is how can we combat uh bots that are uh advertised as AI agents? >> Um how what do you mean with combat? You want to go to war with them? No, no, no, no, no, no, no. So, uh I want to get to collaborate, let's say, with a company that offers AI agents and they are actually selling me bots. How can I um identify the fact that they are bots and they are not actually AI agents? >> Uh I think like uh if you have interacted with agents, so so I mean in the ideal case is when you interact with an agent, it should feel like you're interacting with a human. So it's not like uh if if someone makes a super sophisticated bot then you can call it an agent because it will become the agent like based on the the like the autonomy agency and mission and memory aspects that I uh explained. Um so whenever you feel like so for example if you talk with chat GPT at some point like you will feel or I hope you will feel that there's like a certain bias or you know like the answers that it's giving you it kind of feels um like not authentic and it's it's just like a feeling that you will get and you can trust your instincts on that basically. >> Nice. Any other questions guys? Okay I've got one. So, people talk on LinkedIn quite a lot these days about how soon there's going to be a one-man unicorn billion-dollar company because they'll be, you know, they'll have all these AI agents working with them essentially. Do you see yourself in the line of work that you're in one day working as as Anna plus, you know, 15 AI agents in a company of your own? Is that a future you envision? >> Uh, yeah, that would be great. No. So, always aim for aim for the highest goal. Um I would say uh we have like very early on decided that we want to go in the entertainment route because um or at least like me personally I believe that the future is going to be agentic like all of like all of our industries are going to be changed the way we interact with each other economically like uh organizationally it's going to be all agentic um and the reason like personally why I wanted to contribute with AITV in this entertainment field is to kind kind of take away a bit like the stigma and the fear that people have um when thinking about agents um and to show them like the the good sides, you know, like um that you can you can build a bit of trust. But I don't know, maybe maybe we're destroying the future of humanity. I don't know. Who knows? >> Nice. Okay. Okay. Well, not a positive note to end on, but but thank you nonetheless. Guys, put your hands together uh one more time, please, for Anna here. Thank you so much. >> Thank you.