AMA | Episode 1 - The AI Agent Takeover: Separating Hype from Reality (Featuring AG2)

In this AMA, we kick off the AI Agent Takeover series by tackling one of the most important questions in the space: what’s hype and what’s real when it comes to AI agents? Hosted by our Marketing Lead Joules Barragan and joined by Sahara AI CEO & Co-founder Sean Ren and Chi Wang, founder at AG2 and researcher at Google DeepMind, we dive into the future of open-source agent frameworks, the role of decentralized AI in ownership and attribution, and the challenges of building truly autonomous systems. Whether you’re an early-stage developer or a researcher shaping the frontier, this session explores the breakthroughs, the roadblocks, and the opportunities shaping the agentic AI ecosystem. 

Link: https://x.com/i/spaces/1eaKbWYmEOYGX

Transcript

Joules: Hello everybody. My name is Joules with Sahara AI. I will be your host. Today we've got an exciting AMA lined up featuring two incredible minds in AI. Our very own Sean Ren, CEO and co-founder of Sahara AI. You've all heard him speak many times before. Always great to have him back. We also have our special guest, Chi Wang. Now, Chi, this is your first time on our AMAs.

Chi: Welcome.

Joules: For those of you who don't know, he is founder of AG2, formerly known as AutoGen, the open-source agent OS to support agentic AI and its parent open-source project FLAML, a fast library for AutoML and tuning. He has 15 plus years of research experience in computer science from Google DeepMind, Microsoft Research, Meta, UIUC, and Tsinghua. Did I pronounce that right, Chi?

Chi: Tsinghua. 

Joules: And he's received multiple awards in his field. So thank you for joining us today, Chi.

Chi: Thank you very much. Glad to be here.

Joules: Awesome. If you're listening, if you have any questions throughout this AMA, just drop them in the comments below and we'll get to them at the very end. Oh, and thank you, Sean. I didn't really let you speak. I kind of blasted over that.

Sean: I'm not speaking. Go ahead.

Chi: Sorry. Thank you.

Joules: Chi. AG2 started as a way to simplify AI agent development. I'm curious, what problem did you see in the AI landscape that really made you say this needs to exist?

Chi: Yeah, thanks for the question. I was working on AutoML and hyperparameter tuning as you just introduced. Before I worked on the AutoGen for AI agents, when I worked on AutoML and saw the power of language models, I tried to apply the AutoML idea to tune the inference parameters for these language models and to see how big the difference is when we change the way we use the models for the applications. It turned out there's a big difference and that's the starting point, especially how the model is used in the bigger system. So that's kind of a brand new question for these new types of models because compared to the previous generation of machinery models, they not only do prediction, but they can do open generation in open space. They can even become the new brain, the new kind of control plane for how the system works rather than just become a small component in one system. So that opens up a very big space for design. And still, I saw no good framework. When I started about two years ago to work on the problem, I saw no good solution, no good operating system that centers around these new types of models and to empower developers to build our source application on top of that. So we do need a new type of operating system that's optimized for these new types of models and empower all diverse applications running on top of that. So that was the missing piece I saw and that was my starting point.

Joules: That's really awesome. Sean, you've also been in the AI sector for what, it's been about 10 years now, right? And now you're at the intersection of Web3 and AI, which is amazing. What do you think is the most exciting breakthrough that's happening in AI agent technology today?

Sean: Yeah, I think maybe even more general than just AI agents. In this new era, what we've seen is just like the last era for the internet where people can start posting their own content, user-generated content on the internet and a really efficient way to browse what each other are producing and understand each other's ideas, communicating instantly with each other without any barriers of the physical and time limitations anymore. And I think today when we look at this new era of AI and AI agents, I feel like it really lowered barriers for people to further create interesting applications and content because think about a coding agent, which is looking very promising for laymen and non-tech people to really create applications, front end and all these user interactive applications within just like hours, rather than days or weeks today. And I think that further lowered the barriers for us to communicate our creative ideas and sort of even create monetizable applications in a very efficient way. So what I'm foreseeing here is that because of the much easier and sort of efficient and streamlined ways for people to innovate and produce ideas, it's also leading to the new challenge of how to protect people's creative ideas and ownership of these applications and whatever that is monetizable down the road. So the problem of basically copyright protections, traceable ownerships, attributions becomes a really emerging problem. That's why we started Sahara AI to tackle this spectrum of problems, which I think it's in parallel to improving the competence and ability of the agents.

Joules: Excellent. Really, really great points. You touched on some good points around lowering barriers and open access to a lot of this AI tooling and all of this leading to increased innovation. And a big part of that is this open-source movement and Chi. I know AG2 has been really leading in this space and growing a massive open-source movement. You guys have over 20,000 plus builders from Google, IBM, Meta, and lots of top universities. What's been one of the most surprising ways open-source collaboration has shaped AG2's development?

Chi: Yeah, that's a very good question. So I started working on open-source projects six years ago, starting from the AutoML library FLAML, and AutoGen is actually developed inside FLAML since the beginning. So every single line of code was written in open source since the beginning as a sub-package inside FLAML and only after several months, we moved it out to a standalone repo on GitHub and still keep it. Open-source since the beginning has been not just closed development, it has been open development since the beginning. And together with, for example, Professor Chung Yu was my long-term collaborator in Flaml and also we two the main authors of the AutoGen and when it evolved to a certain form, then more users joined from out of the community and all of the world. 

And I think the most surprising thing is that when we just built the AutoGen framework to do this, we target like very strong capable agents. But we quickly realized that there was a big wall to climb, right? So it's not easy. Even though these memory models have shown very strong promises, there are all sorts of problems to solve. So instead of just building towards that mission by ourselves, I quickly realized instead we could build the fundamental framework first, build a common infrastructure to make it easier for others to build on top of it and try out all their ideas so that we can potentially reach the goal much quicker. And that turns out to be the most effective way that the product has evolved. So a lot of the ideas were kind of done actually by the community and we absorb the ideas and try to iterate on them, refine them and add to the framework to make it even stronger. So one example is this group chat conversation pattern that allows multiple agents to converse, talk to each other in the same group, share the same context, and solve problems together. It is actually initially proposed from the open-source collaboration and as an experimental feature to show what this framework can potentially allow you to do. It's not the only thing you can do, but it was added as an example. But it turned out to be a very popular feature that's used and also developed further and refined to become much, much more feature rich. But it turned out to be a very useful pattern that's continued to be developed and used by the users and there are many such examples. And so the open-source contribution is really the reason that the community is very active and also the library itself got very interesting development along the way. Not only that, but the fact that there are so many people actually trying out their ideas and sometimes much earlier than some mainstream research or engineering teams to do. It's also amazing. 

So there are many, for example, early advanced exploration for quite advanced features similar to sort of the deep research which we saw today, like the commit has been doing at least one year earlier, the open-source community. And they have achieved very, very strong results. So these types of early exploration, very advanced features also fundamentally changed my way of learning from new ideas and doing work. So now the open-source community is my kind of main source of inspiration. So it's not just like occasionally I learn from them, but I constantly learn from them. Yeah. And also they sometimes achieve very high performance on the challenging benchmark, like top performance for example, on the suite bench for software engineering and so on and so forth. Yeah, so that's the kind of big learning I learned from this journey.

Joules: Yeah, that's amazing. Open-source development never ceases to surprise me, especially in this space. Web3 has been all about open source for so long and I'm deep in web 3. I've been here since 2015 and it's amazing to see the AI community developing so much in open source as well. Sean, I am curious. So open-source AI and decentralized AI, kind of two sides of the same coin. Both aim to put power in the hands of the community, but they are different approaches and they do touch on different things. How do you see these two approaches complementing or challenging each other? Especially in the lens of addressing the ownership and copyright concerns that you had mentioned earlier.

Sean: Yeah, I think that's a great question. Actually, let me break down this question a little bit more because I think decentralized AI has been a pretty overloaded and maybe vague term to many of you. And even with open-source AI, it's been overloaded by different companies. For example, Llama, would you say Llama is an open-source model? They might be understanding it as open weights, but definitely not entirely open source about the recipe and all of the process. Right. So I think just going back to Joules' question, I think they can coexist and they can complement each other. If I put on my researcher hat, I love open-source AI because that's basically what drives science advancement and open science. Everyone knows how exactly DeepSeek got created from zero to one with all of the details of the recipe and not just like the weights, but potentially we also hope to see all of the data that's being used in training the model. 

I think another great example is the OMO model that comes from an AI tool at an institute of artificial intelligence where they actually release all of the data sets that were used to create the models in the pre-training stage and post-training stage. And that enables all of the researchers to number one, build on top of those existing data and model and recipe to investigate all of the interesting phenomena that can help them inspire new intellectual ideas for the next iterations of improving the model. And number two, they can also build all sorts of interesting specialized models for every vertical domain and use cases and try to benefit the applications powered by those models. So I think open-source AI is the foundation for science. Without it, I think just simply put, the researchers and PhD students will suffer a lot from really catching up with the current advancements in the industry. 

On the other hand, I think decentralized AI, my definition of it is really just to give the ownership, give the control back to the owner of the data sets and models. If you imagine in the centralized AI paradigm today, developers, data service providers are all hired by the AI companies to do part of their jobs and all of the final products, the outcomes will be controlled by the company and monetized by the company. The revenues and users through the application gateways all flow back to the company. And the company controls how those money and benefits should be distributed across different parties. And apparently we've seen a very unfair distribution today where the model developers live a pretty decent life, they get a really high salary. Look at the packages given by Open AI and all these big AI model stuff. But then the vast amounts of people who contribute to the data, to contribute feedback to the model, to the chat models, they get nothing in return for whatever that is fundamentally useful for improving the model. 

So I think there's a sort of a set bias distribution of these monetary outcomes back to the contributor. And that's what decentralized AI is trying to disrupt. The methodology there is to go from the very beginning, like when things got created, when data sets got created, we know exactly who the co-owners and shareholders are. And this trace will basically propagate along the way into the owners of the model and the owners of the agents and owners of the application. And then you can use this trace to do revenue sharing all the way back to the upstream contributor. And that's like the hope of decentralized AI. And because of giving back all of the power to the creators and contributors, it's a more sustainable economics. Right. Because everyone takes their own part of the share. Now, I think the technical challenges which go way beyond this question is how do you do that distribution in an autonomous way, in a fair way, in an algorithm-driven way? And I think that's a lot of research questions including Sahara are working on today.

Joules: Thank you, Sean. Very well said. Chi, did you have anything you wanted to add to that before we move on?

Chi: No, that's more.

Joules: Awesome. I'd like to shift the focus a little bit. I know we are running lower on time. I'd like to talk a little bit about the real-world application of some of these AI agents in the wild. We've seen some, you know what's been called solopreneurs using AG2 agents to automate everything from emails, calendars, YouTube transcriptions, web scraping. You mentioned a few additional examples earlier. What's one case that just really surprised you and made you think like, oh wow, that's actually really cool?

Chi: Yeah. So one, there are a lot of them. But just to give you one example, also most closely maybe to this community, I saw an example of autonomous trading you build with AG2 and they use the agent to manage your web3 assets and to do automated trading all the time. Even when you go to sleep, the agent doesn't sleep and it's like 24 hours, seven days a week and managing all the assets and doing automatic sell and buys and to increase the portfolio. That was quite a surprising use case to me.

Joules: Yeah, I've definitely seen a lot more AI in DeFi. It's pretty wild. Sean, did you want anything, did you want to add anything to that?

Sean: Yeah, I think just to add on what to observe. Autonomous trading bot or agents have been a big thing in Web3 even though I will counter that they are not quite there or they're not quite usable yet. They're more like a prototype or MVP concept that, oh, we do see. This agent can be successful by swapping a hundred dollars of bitcoins to a hundred of sol for example. That can be done successfully. But allowing them to take a vague sort of user intent. Say I have $10,000 I want to spend on certain types of meme coins and just find the best sort of way of putting these $10,000 with a period of time that I want to do the investment. Like if I put such an instruction, it's still very vague and high level. Would the trading agent be able to take that and execute and have a highly plausible outcome? We're very far away from that. I think we are at least a year or something from there. 

So that actually leads to my point that I think we already seen huge progress made by transitioning from the last generation of AI that is very rule-based, very programmatic to today the more agentic AIs that are able to understand people's natural language comments that are probably under specified in many cases, but they are still able to take common sense actions and plausible actions. And I think that's a huge jump. I believe all of us are very excited about how this could again lower the barrier for us to do things right. For example, I'm not super familiar with all kinds of wallets out there. I'm probably familiar with a couple of them, I'm probably familiar with a couple of the decks, but can you give me an agent that I can just basically operate and execute on all of the decks on the market, things like that. I think it really helps mass adoption down the road.

Joules: Yeah, you bring up some really, really good points and I have some follow-up questions for both of you actually. So obviously we're seeing these agents get a lot better at automating tasks, at being autonomous, but full autonomy is obviously still a really, really big challenge. From your guys' perspective, what's the biggest roadblock to making AI agents truly autonomous today? Like what exactly are we waiting for? You said for these DeFi agents, for example, Sean, we're still probably like a year or a little more out before we, we actually see that.

Sean: I can let Chi go first.

Chi: Okay. So if you think about it, it's not hard to keep the agents running autonomously. It's hard to get them doing the right thing autonomously and with self-correction within a reasonable time. So, because you can, if you just simply want them to run, yeah, you can just give them instruction and they can keep running, but they will at some point do something wrong. When they start to do something wrong, they might go more and more wrong. So that's a big difference from an autonomous human who can, we all actually also make mistakes. We can't guarantee that we're always on the right track, but sometimes we will get reminded, we will realize, okay, we're doing something stupid and we should change our behavior. And I think that's one big missing capability for agents. If we can do that, then I think we will make a very big progress towards making them solve the task autonomously. But if you also take one step further to think about whether humans are really autonomous, it's often also not the case. If you ask some co-worker to work on something for you, maybe they can automatically do something, but it's not always what you wanted in the beginning. And you, you still need to give some more instruction, more feedback for them to improve. 

And so, even when humans have, right now humans have a higher level of autonomy than agents, they are also not fully autonomous. So depending on what your definition of full autonomy is, the bottleneck can be different. So once we reach the human-level kind of autonomy for agents, we probably also want to ask the question, can we, can we do better? Humans actually can do better because when someone works with you for a longer time, they actually get to know more, know your kind of habits better and can actually become more and more autonomous. Eventually you can get most of the things done without your close following up. Agents probably also need that capability. When we solve the current bottleneck, we want them to be able to improve over time and become more autonomous over time. Not just handle everything by themselves in the beginning, but over time they require less interference from the human.

Sean: Yeah. Just to, just to quickly add on what you said, I think I very much agree with what he said and I think we are moving from a programmatic AI that we need to be very specific about what AI should be doing step by step into a more like a goal-driven AI that they can take a high-level objective and try to make that happen, even though we are pretty far away from that end of the spectrum. And I think another big progress we made is this agentic capabilities, including using tools. This is not something we've seen in the last generation of AI that is able to use your browser, your database, use Microsoft Office and all sort of like accessing Twitter and do all kinds of compositional tasks in order to achieve a goal that's been set. So I think we are at the beginning of this whole new advancement and I think everyone is very excited about it. But I don't want to make people overly sort of optimistic about how far we are from there.

Joules: Yeah, you both bring up some really good points. It definitely is an interesting thing to think about. I really liked especially Chi, what you said about even humans and co-workers, like whether or not they're fully autonomous. And I think that was really eye-opening and is a good way of thinking about things. I've never really looked at it that way, so thank you. I think of doubling down on that part. Right. As agents do start to get better and they reshape the way we work. Right. Some people think that they'll just be assistants that help us and others are really afraid that they're going to eventually just replace a lot of people's jobs. I'm curious for both of you, where do you think we're actually headed?

Joules: Do you want to take it first, Chi?

Chi: Oh, I can say it quickly. Oh yeah, go ahead. Yes, go ahead. Okay. So, yeah, it's a really hard question. I think it is very likely they will change the job distributions. There are many ways of change we can imagine. But one thing, one trend I speculate is AI agents will enable individuals or small teams to achieve more. So we probably will see more and more small teams doing amazing things. So that can be one type of change of job distributions that can happen in the future. And another point I think is I'd like to encourage the thinking of what kind of new jobs can we get, can we build in the future? Right. Are there some new jobs that will request humans to solve problems that AI cannot? And if we focus on thinking about that problem, maybe there'll be a lot of opportunities to create.

Sean: Yeah, I think we already brainstormed a new type of job, which is like the moderator of agents because if today they are really having challenges of staying on track with the goal, then we need humans to keep an eye on them. And because there are so many agents going to be running autonomously or semi-autonomously online, we need human moderators to find a way to a good UX UI to moderate them in order for them to not get into trouble. Right. And I think that leads to my way of answering the question. I think really depending on the type of the jobs, we humans are still the decision-makers of the entire sort of agent-human society. And I think the trend will be like we're trying to push them to work on some boring tasks that we don't want to work on. And then we're trying to create some new interesting tasks for us to collaborate with agents. And that's probably going to be going for a while until some other disruptive changes happen. In that case, I think it looks like the agents will be sort of like a system role. Even though I'm trying to disentangle copilot assistance with autonomous running. Right. Because you can have autonomous running agents that are serving as your assistants or copilots, they don't really conflict with each other. So I think that's the trend. We will have boring jobs being replaced as much as possible, but we've seen a lot of boring, just unique physical human interactions with the real world that's really hard to be replaced today by virtual agents. Yeah.

Joules: Excellent. And as we approach this future, I'm curious, what's one thing that every AI developer, whether they're working on AG2, Sahara AI, any other AI project, what's one thing that every AI developer should keep in mind when they're designing AI and AI agents specifically?

Chi: Yeah, for me I think one important aspect to pay attention to is always remember these agents can make mistakes, they can get undesirable results, they can fail. And so if you think in the beginning, you consider that possibility and try to be fault-tolerant and/or add guardrails, you can increase your chance of success.

Sean: Yeah, for me I think it's again to echo the sort of the values of Sahara AI. Right. We need to give back the power and ownership to the contributors. Right. When you're building the agents, don't forget about who contributes to the building blocks, like the data set owners, the compute providers for the agents' training and all these recipes of the agents. They should all be part of the continuously monetizing parties with you and sort of work with you to improve the next version of the agent.

Joules: Excellent, great responses. I do want to make sure that we save a little bit of time for Q and A from the audience. We do have a few questions, so let me pull some of those up. One question we have: What's the steepest learning curve for Devs new to agentic AI? Do you have any tips for getting started?

Chi: So for me, when I think about agent API, I was thinking a few steps ahead and being forward-looking. So that makes me think about this AI agentic software in a fundamentally different way from traditional software. And that could be a big paradigm shift from people who used to think of traditional software in terms of functional programming or object programming. And to many people, it's maybe an open question like what exactly is the good way to reason about them and be kind of future-proof. Right. So because when we learn some skill we want to be able to kind of keep using that skill for an extended period of time instead of just getting obsolete quickly. But that's one challenge because the AI technology is evolving so fast these days and something you learn this month might quickly get not relevant next month and et cetera. So I tend to think in the most kind of forward-looking way to think about agents not just as an application that has strong capability and a natural interface to communicate, but also as part of a new type of software architecture. 

So in the future, we potentially can build software in the agent armchair programming way. And that's probably the most kind of wild thing. But once you kind of get used to that thinking paradigm, you actually find that the difficult software we traditionally need to build becomes much easier to build with this new type of thinking. Because now we can delegate a lot of things to the agents and we just need to think about what roles do I need to solve my problem and how do we get them together, working together. It's very unconventional, but I found it's very beneficial after I kind of got used to that way of thinking. It's helped me to simplify a lot of the kind of design process or reasoning process. Of course that requires some good framework to handle all the low-level details for you. And that's what AG2 and other frameworks can help you do. So you can leverage this type of higher-level programming paradigm more statically. You will find it a lot easier to reason about these agents and agent AI systems.

Sean: Yeah, I can quickly add that I think to me building an agent is not like the traditional type of model buildings for AIs or programming for software where you can use pretty consistent or even like one type of programming languages to do the whole thing. For an agent, you need to use different tools, coding different APIs, data, and interacting with different databases. It's almost like you are creating a very complex plan that has multiple digital axes that interact with each other and the error propagation and error accumulation is much more complex than before. So I think that requires you to really think about agent development as a system rather than just a typical computer program perspective and have all of the risk assessment and mitigations in mind. Yeah.

Joules: Thank you. Very insightful, both of you. We have a few questions on SIWA as well. Just a heads up everybody, we will have more AMAs on SIWA specifically on Discord later this week. So keep an eye out for that. We have time for one more question from the audience. So, Sean and Chi, where do you see the biggest opportunity for startups in the AI space right now? Is it infra tooling, vertical specific agents? What?

Chi: Yeah, I think vertical agents are a good space for startups in general that require maybe the least kind of barrier of entry. Your main, kind of the main element required to build vertical agents is to have some deep domain knowledge about something you deeply care about and have good knowledge about. But everyone, I think more or less has that. And today I think the most successful AI agents I've seen so far are still those that leverage deeply about some domain expertise in a clever way. I mean, the foundational models are already very powerful and they have some strong fundamental capabilities, but there are some missing gaps to make them really useful for a particular application. And there are tons of opportunities for bridging that gap and they are just so diverse. So the type of the amount of things you could do is very large. And also everyone, because everyone has a unique kind of experience and background. And so if you understand some domains more deeply than others, then you're likely to build a very unique AI agent that can solve one problem much better than others. So yeah, so that's just in general, it's a good way to think about it. But of course you have special skills in other things like infra or tooling, then just leverage your strengths to do the best you can. So I think, yeah, so that's my kind of a brief answer to that question.

Sean: Right. To add on what Chi said, I think there are some foundational tooling still missing in the ecosystem. For example, we are missing a good sort of agent evaluation sort of environments and leaderboards. Right. Like let's say when we talk about trading agents for DeFi, there's no consensus like what I mean, I feel like we are still in a very early stage of developing those agents, but let's say in a few months, once there are some matured agents out there, how do we know which one is performing better on which use cases? Right. We do need infrastructure such as a sort of battlefield for these agents on trading use cases and be able to understand their strengths and weaknesses on different aspects.

Chi: Right.

Sean: We need that infrastructure tooling and I think there's an opportunity of building that because this doesn't necessarily get picked up by those major Web2 agents or foundational model companies.

Chi: Excellent.

Joules: Thank you both. We are at time. I want to make sure that we have time. If you have any final thoughts, anything that you want to shout out Sean, Chi, now's the time to do it.

Sean: I would say again, just trying to emphasize my initial message. I think equally important to improving the competence of these agents and improving the agent capability, we also need to care about sort of people's contributions and tracking and tracing those people's contributions in the process of building better agents, especially these vertical agents that have very tangible monetization opportunities. So we, including Sahara AI, really want to grow this ecosystem and really like and make these directions to be paid more attention to and creating the infrastructure and tooling to help people monetize in a fair way.

Chi: That's interesting. So actually that's one thing I really care about. The open-source contributors have been a big part of the AG2 growth and any ideas that help these contributors to kind of get rewarded, get credited and keep kind of contributing and also getting more and more people joining. Yeah, if there are any good ideas that can help with solving that problem, I'd like to try them out. Probably Sahara can provide some good solutions for me. I'm super interested in your progress and learning from your practice.

Joules: Awesome. Thank you both. Big thanks Chi Wang from AG2 and Google DeepMind Sean from Sahara AI. If today's conversation sparked any new ideas, be sure to check out AG2's latest updates and be sure to follow Sahara AI to stay ahead of the decentralized AI movement. Thank you everybody for showing up. I hope you have an amazing day.

Sean: Thank you for having us. Thank you everyone.

Chi: Thank you very much. I appreciate it.