What is the architecture behind agentic AI?

Room A – Panel 2 Agentic AI Architecture

REED ALBERGOTTI:

00:00:03:17 Let's get started. I'm Reed Albergotti. I am the Technology Editor with Semafor. So, we're here talking about agentic AI and I mean, let's start with you, Jacob. So, you're at PwC, early on in this whole, I guess the ChatGPT moment. There's a lot of talk about, like, the death of consulting. Like we're not going to need you people anymore.

00:00:27:15 That doesn't seem to be the case, though. Do you agree or disagree with that? And what are you kind of seeing out there in that world as you work with companies?

JACOB WILSON:

00:00:35:19 What an easy question to start with. So, I think, when ChatGPT was launching, I think even for us, like when we were trying to figure out, hey, is there value in this technology? What is the impact to our organization overall? So, at that point in time, we didn't know what we didn't know.

00:00:54:22 So we kind of squashed 200 people together, two locations, just got hands on keyboard and tried to start to figure it out. Right. What do we think the actual benefit or the risks are to even our business? Across advisory, tax and audits and all of those areas, I think quickly what we found is there's definitely a lot of immediate value.

00:01:14:08 Right. And obviously, you've seen the capabilities evolve over the past couple of years, like it keeps continuing to get better. I think, even for our own transformation, what we found is like when you go through a big scale top down, transformation like big T type of transformation, right? It still takes a lot of work to get these multi-agent systems to get you the ultimate business value.

00:01:37:23 And how to meet the objectives and KPIs that you're trying to accomplish? So even for us in our own journey, we've poured a lot of money into it. We've gone through a large-scale transformation. I think we're still finding it takes a lot of work. Right? And even with our clients. Right. There's finding the same thing where it's not just flipping on a switch and

00:01:53:17 magically, I've got a big, massive transformation across my organization where I'm driving all these efficiencies or new, top line revenue streams. So, I think to your point, there's a lot of value. It's definitely disrupting us. We're finding ways to use it to help change the way we deliver our services

00:02:12:0 and market, but also helping our clients still to figure that out in terms of what they need to do to kind of enable transformation within our organizations.

REED ALBERGOTTI:

00:02:20:10 Thanks. And Sudheesh, I mean, you're working in enterprise web agents. So yesterday, OpenAI launched their web browser. I've used Comet from Perplexity and Project Mariner stuff at Google. The idea behind all this stuff is you're going to be browsing the web, and these agents are kind of just doing all these tasks, these sort of annoying internet tasks that we all do all day.

00:02:45:18 For us, it's on the one hand, amazing. On the other hand, I find it just isn't quite there yet. And I'm wondering if you maybe you disagree with that, but if you do agree, why is it not there? What's it going to take to get there where we can just like tell these things to go do stuff for us?

SUDHEESH NAIR:

00:03:03:04 First of all, thank you for hosting us and thanks for joining us. We are not there yet. I am not going to disagree with that. And that is because currently the state-of-art models are really good when there is a clear definition of a goal that they need to course towards, and when there is a clear goal, it can find out there is one right answer, a bunch of wrong answers.

00:03:24:21 It can reinforce itself to learn and get there through a number of different techniques and that amount of data. The problem with browsing is that often it is a multi-variable equation with extremely ambiguous goal, with a sense of happiness that is derived through the journey and we are trying to automate that. So, the state-of-the-art, I would say what OpenAI announced as usual, they announced something.

00:03:51:23 What I think is going to get really better really fast. So, we should all assume that browser agents are going to get really better. So, if you were to take a step back, the primary intent of agentic AI is that is finally like the brain is getting the arms and legs, which means it can start doing things. If LLM is the brain, agentic AI is the arms and limbs.

00:04:11:10 So, there are lots of things to do. And one of them, we spend a lot of time on browsers, so it makes sense that we start automating it. What the state-of-the-art browser agents, what I call them agentic browsers are expecting a human to sit and observe and monitor and sort of oversee your browser being taken over.

00:04:31:21 Tiny Fish, the company that I've co-founded and be the CEO. We are trying to take it to a maybe three, four level out, which is imagine no humans involved at all. You set the goal and hundreds of thousands of reasoning agents talking to hundreds of thousands of action agents simultaneously working on 50,000, 100,000 websites and web pages at the same time trying to accomplish multiple goals and get it done.

00:04:57:18 So yes, it is not there, but that is precisely why it is the right time to go do these sorts of things. Because this is a time to build.

REED ALBERGOTTI:

00:05:05:01 But is it, are you waiting for the models themselves to get better, or is this just like, is there a lot of work here and sort of rolling up your sleeves? Is it bespoke, custom, ways to get this to work?

SUDHEESH NAIR:

00:05:17:19 No, models are not. We are not expecting models to get better at all. What we are trying to do is to build a system that uses the model as they are. And this apply to every agentic company out there. If you are expecting the model to do it, then you are not necessary. Every time they release something, you will be like holy shit, what is going to happen to my company?

00:05:35:10 That's not the point. The point is there is such a vast area of expertise that can be applied because this is about applied AI. Agentic AI is about doing things. So, in this case, navigation, first of all, understand the goal, translate that into skills and task. How do you deploy them in a deterministic way when the models thrive in a non-deterministic fashion.

00:05:57:11 So enterprise customers want actions and results that they can replicate across multiple sites. So there are so many opportunities there. And then infrastructure. How do you spin off hundreds of thousands of browsers that are heedlessly running on your command trying to get your goal done. How do you manage them? How do you observe them? What is observability? That's an opportunity area.

00:06:16:13 And then you flatten the websites. Then you get that data, like hotel booking. Everyone talks about this. The reason is the price of a hotel is hard to find. If you flatten the website and bring that data, that data itself is valuable. How do you use that to make the models better? So, building a mixture of expert model for navigation that is precisely built for that will be something that is derived because frontier models are getting better.

REED ALBERGOTTI:

00:06:39:15 Okay. Well, you mentioned infrastructure. We have someone from Oracle here. You're making a few investments right here. And infrastructure. How much do you think is the sort of limitation of the tokens, right, that we have right now? How much is that holding us back? And as you build this, these massive data centers, do you think that's going to change the capabilities?

JUN QIAN:

00:07:02:23 Yeah, I think first year, we are talking about and it's a year of the large language models. And the second year we are talking, oh, it's a year of the agent right now. It's a year, kind of the year of the agent applications. And to go back to your question, I think, is foundation models, keep getting better and better.

00:07:19:23 And, but I feel there is still that's a gap. Right. And when you're talking about the enterprise solutions. Right. And we clearly saw some successful applications like ChatGPT. Coding agent like ChatGPT and the higher I think in the enterprise solution, it's not just about the token, I think. Right. And one of the challenges where we are seeing is still the integration, right?

00:07:45:08 When your integration is with legacy systems, for example, authentication, right? When you connect to the backend database, when you connect to some systems which don't even have an API. And how do you conquer this integration hurdle? I think that's one of the challenges. Right. And we are facing right now, then that means how do we build all the infrastructures?

00:08:09:18 Right. Even catch up, the intelligence that existing in large language models, I feel that we still have a big gap there. And everybody knows MCP, that clears one direction. And however, I think, right. And still a lot of work, a lot of opportunities. Right. And for the middle layer and how do you connect all the legacy in the enterprise system?

00:08:29:21 How do you solve the authentication issues, observability issues, right and the (Unintel Phrase ___08:36) issues? I think that's one area where we are focusing. Yeah.

REED ALBERGOTTI:

00:08:39:05 I mean, Jed, you're, I think in Dataiku you have an interesting vantage point because you just see so many different industries and what they're, the models that they're using and what they're using it for. Are you seeing like what are the use cases and is like, can you tell if agentic, do you have any indications like agentic is actually catching on?

JED DOUGHERTY:

00:09:02:00 Sure. So, I think agentic architecture and figuring out agentic use cases ends up being a business problem and a human problem. Everybody's been doing a lot of analogies over the last couple days, but I like the Industrial Revolution as a good analogy. So, we had the steam engine created in the 1770s. Somebody could probably correct me on that.

00:09:26:07 And we had the full actualization of that steam engine concept. Not until, let's say, model T Ford the full assembly line. If we think of the relation between like LLMs and agentic pipelines as the steam engine and assembly line in 1903. So that's like 135 years or something like that. Which obviously nobody's going to wait for this time.

00:09:50:02 We need to do it maybe 20 times faster. So that would be like something like 2028 is the full understanding of how agentic architecture can actually create a lot of business value. And the human side of that is that it took somebody like Ford to see the full actualization and the full concept of what these things could do.

00:10:13:21 And today, understanding which business applications can have agentic systems applied to them — it does require, I think it was mentioned over here, a very precise understanding of how those business systems work today, with the humans being inside of them. And you look at that business system and then you identify which subcomponents of that system can be replaced today by agents or by LLMs.

00:10:41:12 But you need a very clear idea of like, what all of your business systems are in your organization. Probably you can help people understand that. But a lot of people, when we talk to them about how agentic workflows could replace some of the components of their existing business, they don't have a clear idea of how that works today.

00:11:02:04 There's like a lot of, like, oh, I sent it over to Sally, and then she does something and sends it to Robert. Until we understand at the executive level even or at a very clear programmatic level, what happens, we can't replace it with agents. So, understanding your business means understanding how to build agents.

JACOB WILSON:

00:11:19:06 Yeah. And if I might add on to what Jed is saying there, I think agentic transformation like it's really gotten organizations to think about what do my current processes even look like, right? I'm doing deeper analysis into the work that's being performed better, understanding where that work happens, whether that's onshore, off shore. And how can I really think about getting the efficiencies in the process through gen AI. So, I think it's creating businesses and organizations to kind of relook at everything, right?

00:11:49:02 From finance to HR to operations to a whole gamut of things. So, I think that's what's really been interesting this journey, right? It's causing you to kind of rethink and how you can apply the technology to accomplish the objectives you're trying to go through. But again, it is digging back into the bowels of really understanding the work where it's happening.

00:12:08:20 And how do you actually transform it with the latest technology with the LLMs, the LRMs and the multi-agent systems?

REED ALBERGOTTI:

00:12:15:23 When I talk to people who are doing stuff like what you're doing, right, helping businesses automate processes. When you really dig into it, like there's a lot of marketing hype around this stuff, as we all know. But like when you really dig into what they're doing, it's like 80%, right, other software and maybe 20% LLMs on top of like this more, I guess you call it more traditional software, maybe machine learning solution.

00:12:44:02 Right? Is that what you're saying, or do you think there's a world where LLMs actually does do everything?

JACOB WILSON:

00:12:50:11 Yeah. Let me maybe ground us in like a few examples. Finance is a great one. If you think about procure to pay, order to cash, report to record. Those are very prominent processes across any organization. To your point, there are capabilities coming online like within SAP and Oracle, as an example, right, where they're bringing more native generative AI capabilities into their stack to help solve some of those process issues in the finance function.

00:13:18:09 But we find is that, you might have multiple ERPs across multiple locations. Those are all on different versions. So rather than spending large amounts of money to consolidate all of these things, right, which is kind of the traditional methodology, how can I use agents on top to help simplify the process? Right.

00:13:40:04 And then connect all those systems together and then redefine what that process looks like, with agents instead of a person in finance having to go balanced between all these different systems, which might include a reboot and all these other things to actually do the procure to pay process. So I think, it depends,

00:13:59:10 I would say on some of these use cases, I do think we see agents coming in as like a just equivalent to as the ERP in terms of the capabilities it can bring to connect all these systems together and truly transform the process.

JED DOUGHERTY:

00:14:12:23 A quick follow-up on that. I think we see kind of two paradigms. You have like high-originality, low-autonomy agents. So, agents helping with relatively original problems. And those are like your assistants. And then you have low-originality problems where we give agents high autonomy. And those are like kind of your glue pipeline assistants and are agents and yeah, both are very viable built. They're very useful for businesses.

REED ALBERGOTTI:

00:14:38:13 But what's the consumer? What does that mean for a consumer? I'd love to hear your, because consumers can't hire PwC to come in and figure out how to automate their summer camp planning process or something. So…

SUDHEESH NAIR:

00:14:50:06 I will just say that, maybe look, every time you go to a conference and listen to a panel like this, you want to put everything that we say through a filter. First, we are all focused on our self-interest. I'm here to make sure that you all know Tiny Fish. Let's be real. I would love for you all to learn.

00:15:07:06 But most important that, right? Number two is we are our perspective. We are all coming from different perspectives that are limited. So, everything we say, having said that caveat, I feel like we don't understand the sense of urgency around this is extremely understated. So what? I have a higher sense of urgency. Maybe because I'm in the bubble.

00:15:29:08 I'll give you an example. We have these agents that can navigate internet like humans to. But like a superhuman skill. There is a butterfly effect that happened in the last few months. It's been building and something very minor, but very significant that I'll give you an example of the sort of things that you should anticipate, no matter what business is in.

00:15:47:07 We all know that Microsoft, Google, OpenAI, they're all going for each other's throats, right? And they're all trying to go at their business other than Nvidia, everybody's like, Nvidia is making bank. Everybody's like fighting. Right. And one of the reasons is Google's entire business is surveillance capitalism. You go to Google and say, oh, Google, where do I go?

00:16:07:13 And they will say, here you go. And people pay to put their names out there, their products out there. And that is going to be fundamentally disrupted if there is going to be an agent or an LLM that says, no, no, this is the answer. So, a CEO to GEO and all of a sudden everybody's like, oh my God, we got to make sure we pay a lot of money.

00:16:24:04 So that OpenAI answers will always be my product. This is not sustainable. We all know that we can't go from one page to survive to one answer to survive. There has to be a better outcome-driven web. So as part of that, you could see Cloudflare saying, oh my God, our business is going to be screwed because if they take all the content and then answering, then I won't get anything.

00:16:45:15 So I'm going to stand in between the content creator and say, you pay me, then I'll give you content. So they are trying to do that, and then ten days ago, some people may have noticed this. Google basically said, hey, we are going to restrict that search result to ten results as opposed to 100. This has huge implications because all retail companies use Google Shopping or Google Lens to see what others are selling, so I can set my price.

00:17:08:15 All of that broke because all of a sudden you can't really search for other people products through Google, Google fundamentally stopped it so that bots can do it. We got a bunch of calls from retail customers saying, if you can navigate like humans, can you find this? This is an example of the butterfly effect of all these massive things and the realignment happening.

00:17:27:20 This will affect everyone's business and you might think like you are safe for now, but when that change happens, it will be overnight. And the replication and the ramifications of that would be massive.

REED ALBERGOTTI:

00:17:39:01 But are you saying and maybe someone else has an opinion on this? Are you saying that right now? Yeah. I mean, to make this stuff really work, you can't just have the AI agent going out and sort of figuring out how to navigate the web. You need those websites or companies to kind of participate in this process.

00:17:57:14 And that's going to change. And I don't know if that's like I mean, Jun, you've been doing AI research work for a long for your whole career. So like you, I think you see the trajectory, like, do you think that that is going to change?

JUN QIAN:

00:18:12:02 Yeah. I think if you thinking loud. Right. And all the interface we have built today, including browser, all the software and in certain sets we built for humans. So if you're thinking loud, is the system we're building is really good for the agent in the future? If you think this way. Right. And because I am always thinking from the mantra from first principle.

00:18:38:07 Right. And why we built a computer? The reason we built a computer and we built the operating system. And the reason is, sorry. Yeah. I think the reason is we're trying to communicate and with a machine, and then we build the operating system, we build a computer language and we build an API.

00:18:59:13 Right. And then we build the browser. Right. All this. Right. So fundamentally, and you can think about is, we're trying to build all these communications. Right. Human and the machine. However, if we're thinking loud, if the future. Right. And if the agent is smart enough, do we need to have all these different layers. Can we have an agent talk to the agent directly?

00:19:25:15 I know we already have A to A, right, agent protocol. We have MCP. However, I think, at this moment, right. I think we should rethink how we build a software and how we even build operating systems, I think. Right. And that's why OpenAI and Perplexity, I would assume at Google, they are all building native browsers.

00:19:50:19 And the reason is that's the only way they can embody the whole large language models as your native experience. Right. And essentially, you can use this native agent as like a booking ticket, right? I think everyone's talking about booking ticket as an example as an agent. However, today we still don't have very successful system, right? Booking agent. I think we have to rethink all these things fundamentally.

00:20:18:10 How can we enable the larger language models natively right in the system allow or not even in our human interactions? So, I think we need to fundamentally change the interface. How human communicated with them.

REED ALBERGOTTI:

00:20:30:18 Do any of you disagree with that point that we have to rebuild the whole sort of web, I guess, if you will, to make these things work?

JED DOUGHERTY:

00:20:37:21 I think there's definitely a pessimistic view of white-collar work around the world today, where white-collar work is essentially taking information from one computer system, doing something mushy with it, and then putting it into another computer system. And if that's true about white collar work, then, yeah, we can have agents directly doing that mostly, like pretty quickly.

JUN QUIN:

00:21:01:21 Yeah. Maybe they don't need to speak English. Right? They can invent their own language and talk to each other.

REED ALBERGOTTI:

00:21:07:13 But that's going to take a while. Right. And then I think that’s…

SUDHEESH NAIR:

00:21:10:11 I will disagree and I won’t just say it.

REED ALBERGOTTI:

00:21:12:06 Which part?

SUDHEESH NAIR:

00:21:12:14 The question that you asked is a very important question. I don't want people to walk away with the idea that the long tale of these websites to transform. It'll take time. Don't think that. And I won't just say it. I'll give you an example. If you go to TinyFish, our website, TinyFish.ai, you will see a small, a video.

00:21:30:02 That's our main headline. It's an eight-room hotel in Japan in some mountains. They have eight rooms — no technology, an old-school stack reservation system. Our customer is Google Hotels, which is a travel Meta. Google Hotel wants to take every single hotel everywhere in the world and make it available so that when you search, it will show the name, address, availability and price.

00:21:51:11 Problem is, there are thousands and thousands of these small hotels where they don't have the API to constantly update the pricing. So, if you go there, you'll see the hotel name. And it'll say call property for pricing and inventory, and no one does that. They'll just skip and go to the next one and see the price and go to Expedia and book it.

00:22:07:21 Our agents are logging into this hotel multiple times a day, navigating their complex Japanese website, figuring out different combination. Monday check out, Friday check out. We got Queen room, King room, whatever. Finding the availability, finding the price and automatically updating without rate limiting so that the website doesn't get overwhelmed. All of a sudden, these hotels are showing up with availability and pricing.

00:22:34:05 Hotel did not change a thing. They are now part of the system. What we did and we did this similarly thing for 40,000 nail salons. Where we are the front end where you can ask questions like, hey, do you use organic products? Yeah, sure. We scrape it. And then do you have availability for this Friday? How much does it cost?

00:22:50:21 We are able to login live, do that. What did we do? Those 40,000 websites are now MCP compliant so to speak. That is our company. The reason our company is called TinyFish is because we want every small fish in the ocean to have the same capabilities.

REED ALBERGOTTI:

00:23:06:02 I'm not sure I fully understand that you, so to excuse and explain that in plain language. You got these hotels to, who's asking the questions here? Just kind of like…

SUDHEESH NAIR:

00:23:16:10 So if you are a consumer trying to figure out does this hotel have availability for this Friday and what is the price? That's a question that you cannot search and store, right. It has to be somewhat real-time. Most large properties will use API to keep updating Google and Expedia and others to do that. But these small hotels, they have, just like a WordPress website and they don't have it.

00:23:37:20 So what our agents are doing is hotel prices don't change every second, but they may change 3 or 4 times a day. And but to know the pricing, you have to execute a workflow. Three people checking on Monday. Exiting on Friday, I have this corporate discount, Queen room. How much will it cost?

REED ALBERGOTTI:

00:23:54:21 So it is a form of scraping. Like, basically you're proactively doing this?

JUN QIAN:

00:23:57:29 I think what he's saying is we are going to not building website anymore. We're going to just building the agent to star ways.

SUDHEESH NAIR:

00:24:09:14 Yeah. Instead of scraping. It is human-like navigation to bring the data. All of a sudden web pages are less important, websites are more important. And you should be able to ask, hey, who are you? What do you do? How much did this cost? Can you do this for me? And every one of these teams will answer that, and it will not be because every website changed. It's because companies like us will be front-ending them.

REED ALBERGOTTI:

00:24:31:11 So why doesn't that work for the entire, if you solve this problem, right, does that just scale and you can do that for the whole internet at this point?

SUDHEESH NAIR:

00:24:38:09 That is my point, that's what I'm saying. We shouldn't expect that to happen everything.

REED ALBERGOTTI:

00:24:44 Should you launch your own browser then? I mean.

SUDHEER NAIR:

00:24:43:06 So we have browsers that we are built on chromium that executes in cloud but less than browser for us what is important is fingerprinting, Captcha, solving, human-like navigation. So, that we have invested in the infrastructure as well. Absolutely, we did. Yeah.

REED ALBERGOTTI:

00:24:59:06 Okay. Okay.

JUN QIAN:

00:25:03:12 So think about, I don't know, five years, maybe we don't need the browser or website anymore. You just have your personal assistant. Right. Which goes to get all the information you need. Right. And we can do more interesting things, right? Instead of browser (Unintel Phrase ___25:19) gravity information or take actions like booking ticket. I think that's totally a waste of your time.

SUDHEESH NAIR:

00:25:23:20 Look at the Airbnb website. That's actually a good example of potentially what it could be like. Airbnb website is not a website, right? It's something where you log in, it is bespoke. It is basically about action. Imagine that, for a who are you? What do you do? How much does it cost? Do you have availability? It should be able to answer who, what, how and do.

JACOB WILSON:

00:25:44:15 I think along the lines of like agent-to-agent communication because that's kind of like the heart of what we're getting at. Right. It's like, we do see organizations, like, and food and beverage industry is a great example, right? Where you might be licensing your recipe, right, to produce something. Right.

00:26:07:12 And then now you've got agents communicating with that manufacturer or who owns the recipe to, for example, bottlers right on the other side or you have different things where now you've got agents communicating with other agents from like a central operation center out to different agents in the field on different devices.

00:26:27:04 So I think that whole protocol shift, right, is definitely kind of reshaping the thinking around things and how you manage the information flow between, like, you as a producer versus your distributors or as you're kind of manufacturing things, right. How do you manage central control, but then also communicate and manage things more directly in the field, or whether it's construction or building out the latest and greatest AI data centers.

REED ALBERGOTTI:

00:26:50:23 Got you. Yeah. So, one question I'd love to touch on before we kind of turn it over to the audience. Do you guys, do you think we're in a bubble right now? That's a big conversation. I think it ties to agents in a way.

JACOB WILSON:

00:27:05:08 So I got to take this one. So, let me give you an example. So just because we live in the physical world. So, I'm with, my daughter's back to school night, right? I'm just sitting in the room, we see a picture, like, of an old school stadium. And my wife and I are having a debate on which stadium that is on the wall.

00:27:21:23 So I turn around and take a picture of it, send it through ChatGPT. And I live in Denver. So just for context. So, it thinks it's the Oakland Football Stadium, right? And then it generates all this chain of thought reinforcing its perspective. No, that's Oakland. And we're like, no, that's Old Mile High stadium in Denver, right?

00:27:43:08 That is definitely not Oakland. So that's a good example of where if not prompted correctly. Right. And if you just assume it knows everything, you will not get the right result. And if you monitor the chain of thought, it will continue to reinforce itself. No, I am going down the right path. This is the right answer.

00:28:03:21 So, I think that it's still like a challenge, right? Like, it's great. You can do large-scale transformation with your organization, but it takes a lot of thoughtful approach in terms of how you instruct, guide the agents, the guardrails you're building in, the context layer that you're managing to make sure that they have the appropriate information and guidance to actually get to the right outcome.

00:28:24:01 So I think the bubble is, there's a lot of perception created that it's going to go solve all my problems. And it has a lot of great applicability, but it can also fall short on the simplest task if not done correctly.

JED DOUGHERTY:

00:28:38:10 And if anything, I think there's a gap in the market right now. They, we have tons of these individual little components floating around out there, but we don't have the agentic architecture in a widespread enough way to allow businesses to use them immediately effectively. So, I think that gap will be filled,

00:28:58:07 is being filled, but and that now will continue to push the capabilities and the value of these underlying components through the riff. One example as kind of the challenges of AI architecture that I think people are trying to figure out is, all right, let's say you have a bunch of agents even just in your business, you're not even talking across multiple businesses, but in your business, how are they relating to each other?

00:29:26:03 Is it a hierarchical relation similar to like HR? So, you have like a top agent and then the sub agents and then agents down below that. Or is it a network where they're all cross-referencing to each other. Is it HR or is it a social network and the way that you need to put rules in your organization and the way you need to manage these agents is very different between those two concepts.

00:29:50:02 Understanding the difference, understanding which one is more applicable to your problems and understanding how to deploy them in a way that's safe and governed ends up being really, really critical. And I think businesses are trying to figure out how to do that right now. And when they do, we're going to see a second explosion in this market.

REED ALBERGOTTI:

00:30:09:07 Any other thoughts on that topic?

SUDHEESH NAIR:

00:30:11:15 No, we are not in a bubble. That's not.

REED ALBERGOTTI:

00:30:17:12 Okay. We're not. I also wonder, like, if this, the way you talk about this, it sounds to me like when you think of agentic, you think of like, I think that's like, right where you could see massive unemployment or something, right? Because it's just doing things that people otherwise could do. Are we going to see that or do you think we're, do you have a more optimistic view?

SUDHEESH NAIR:

00:30:41:02 No. I am very optimistic. But, there is going to be a transformation and there will be job losses. I mean, look at our company, the kind of work that we are able to do with 26 people. Before this, I was a CEO of a company called Thought Spark. Before that, I was the president of Nutanix.

00:30:57:03 In eight years, we took the company public. We were like 6000 people. I think that we are probably as productive as a 300 or 400-person company with 20 plus people. And I think we are still not quite there yet in terms of efficiency. That has to account for. A couple of weeks ago on my LinkedIn, I posted a blog saying like, what the hell are we supposed to do as humans?

00:31:19:06 And there is so much to do. The problem is the thing that makes us human is what we are forgetting. What makes us human is going out. You all know this Bitter Pill, paper that was written on. You should search and read that, it's actually a good one. There's a big conversation around that. But the idea that, well, what the models are doing what we have learned to do, but it hasn't learned to do what we are, who we are.

00:31:45:01 I mean, the reason 400 years later, we still talk about Shakespeare is because he experienced life in a way that he is able to interpret in a unique way and LLMs can emulate that. But the problem we have is that we are not going out and feeling the wind in our face and see what it feels like, LLMs cannot do that.

00:32:02:08 That creates a productivity that will be fundamentally different. When I start speaking like that, people think I'm cuckoo. But the reality is, the more time we spend short circuiting the journey and getting to the result, which is what LLMs are doing, we are not experiencing the hardship of learning and researching in the library. All those things that we learn our brain to do is the reason why we will be better.

00:32:25:06 So those who are going to be mindful, those who are going to take that journey, those who are going to switch off how to experience like journalists who used to write versus the short circuit the whole process and put it through an LLM, they are going to find, like, what the hell are we supposed to do? So, I am very passionate about this. I wrote some stuff, but again, these are all opinions.

REED ALBERGOTTI:

00:32:43:05 Yeah, well, I would love to have it write for me. That would be great. Save me a lot of time. I want to open it up to audience questions.

AUDIENCE SPEAKER:

00:32:51:03 Can you talk a little bit about identity management and agents? Because they're kind of things on their own, entities on their own. Love to hear your thoughts on what's happening, where things are going.

JACOB WILSON:

00:33:00:07 Yeah, I can take that. I think it’s great as the ecosystem is today with all the agent frameworks and everything. One of the things we are finding more commonly fall flat is the security and the management around the agents. So that's a lot of where we've been helping customers to in terms of,

00:33:16:05 no matter where your agent sits, no matter what cloud provider using AI, like a Hyperscaler or SaaS provider, how do you get that consistent authorization around who can invoke the agents? But then also how do you flow the authentication, right from like a JWT for your authentication claims down to the agent to then flow to the back-end system?

00:33:42:10 I think things are improving there, but there are still a lot of gaps, in terms of how to do that. And then also, I think another trend, right, to is manage identities for agents. And what you're explicitly granting that manage identity for that agent to actually go back in the system to do. But then you still have to apply an AR back layer on top of that, because then if you're giving an agent direct access to a system, perform certain things, you don't want anybody invoking that agent. Right?

AUDIENCE SPEAKER:

00:34:04:11 I have automated a lot of my personal workflows of doing shopping, using some kind of an agentic workflow. And first time I use it, the immediate thought I had was, then why do we even have a front end? Why couldn't they just talk to the back end and then I see the world where people are probably making their web application easier for agents to navigate, but then maybe it could even be one step further, right?

00:34:30:00 Why even have a front end? So, I wanted to get your thoughts on, why do we have this intermediate step of having a vision model which can understand the web layout? Why not directly go to that final stage of not even having the front end?

JUN QIAN:

00:34:44:10 Yeah, yeah, I think I can take this one. I think it's just time. Right? And given that we have a lot of a legacy systems, legacy information, right in this work, however, if we start building agent, AI and native applications starting from today, right, every one of us and you can imagine in the next few years, like TinyFish, right?

00:35:05:22 We already have a lot of the native AI agent application in the world. And if you're starting today and build as a community, I think this is going to evolve quickly.

JED DOUGHERTY:

00:35:17:03 Quick follow-up on that. As a child of the 90s, I'm like very nostalgic for the internet as it used to be, like prior to social media, where like it was weird and there were a billion websites and this, I think, even more, makes the internet not weird and makes it like simple and bland and centralized. And I hope there's some pushback and always some place for a weird internet.

JUN QIAN:

00:35:47:16 Well, actually, it's another thing, right? And apps. Right. We're building the mobile apps. Right. And when mobile apps started, right. And we're building a lot of mobile apps. Right. And yeah, I think a lot of folks probably won’t be using website. Right. They're just using mobile apps. You can think about. Right. And we have website and mobile apps, agent applications. I think it's just the progressing of them.

AUDIENCE SPEAKER:

00:36:10:08 Hi, my question is around transformation. So, enterprises that might be feeling behind or overwhelmed about what to do to take advantage of this year of agents and age. Are you seeing or have any thoughts or guidance on things that are commonly missing, like there are too much experiments from the ground up and not enough strategy, or there's too much waiting for strategy?

00:36:34:11 Or should there, make sure at least have a task force and center of excellence or something else?

JACOB WILSON:

00:36:40:23 Yeah, I think, a couple of things. For sure, AI Center of Excellence is kind of a common theme in terms of how do I really start to get a centralized function, right? Really looking at and understanding all the opportunities and complexities, right, that come along with how do I transform my business with generative AI?

00:37:01:10 So you kind of need that centralized thought leadership to kind of help look at that across your organization. That's kind of a common theme, as well as how do you start to think about the skill sets that you need to build up to enable the agentic transformation, just like any other large-scale transformation?

00:37:20:10 Right. Like you need some type of centralized expertise, but then ultimately you get mature enough for it to where you can kind of federate that out and distribute those kind of technology aspects into the different business functions. But I think in terms of like the biggest thing what I was talking to earlier is,

00:37:34:05 in order for you to get the value of the agent transformation, you really have to understand your current process. So that comes in to really understanding what are my L1, L2, L3 type of processes in each of those functions. And then where does it actually really make sense to deploy agents. Because if I've got a thousand L3 processes, does it make sense to really put agents against all that?

00:37:55:10 Or where can I get the biggest bang for the buck, in terms of using the agents to automate a core grouping of those L3 processes like in a finance function. So, I think the more you can understand around the current processes and where the hours and the cost sit and those kind of level of details,

00:38:13:12 that gives you a much better idea in terms of where you can then start to focus and say, okay, maybe for that group of activities, we don't really need to send that to someone to sit there and manually review. Like we can hand that off to agents to go do that work, perform that action in the system.

00:38:29:05 But still, you want to also understand where you want the human in the loop, right. Because there's going to be certain things like on invoice approvals, for example. You don't want an agent approving $1 million invoice, like, you need to bubble that up for a human in the loop to review and provide final feedback on.

AUDIENCE SPEAKER:

00:38:43:01 One of the things that I'm trying to build right now at work is a platform for agentic workflows. And the biggest problem is around how we can include structured, clean and reliable data sets into the entire system. But a lot of the feedback and a lot of the interactions that we get are unstructured, but we want to make sure that

00:39:01:17 we include them into the general feedback loop. So, I'm curious on during the design phase of building these kinds of systems, how do you think about tackling those problems and making sure that you still have, you're not polluting the way that the agents act, I suppose?

JED DOUGHERTY:

00:39:19:10 Sure. I can take that since it's what Dataiku does. You absolutely, you need to be able to take your structured data. And just as you did from, who here who ever did machine learning back before AI was cool? All right. Like half the room.

00:39:37:10 So 80% of building machine learning model was data prep, data cleaning, getting the data in the right situation, building features. And to a large extent, like you don't have to train models anymore, but if you're asking your model to hit your underlying data, you still need to do that.

00:39:51:13 You still need to transform your raw data into understandable data sets with good metadata that then the model can run text to SQL on and provide good information back. So, your data prep and your data transformation does not go away with agentic AI. If anything, it's more important. So, you still need to invest in that. And ideally you invest in a system that makes it very easy to combine those underlying data sets that you've worked hard on with the agentic systems that you're building on top.

JACOB WILSON:

00:40:18:12 Yeah, I think maybe just to add on to Jed's point, because the unstructured piece is interesting, right? Because unstructured data can come in so many different formats. And to Jed’s point, like you've really got to understand because like when agents go to retrieve information, right, like you think about a hybrid search where it's doing, keyword, metadata filtering in addition to semantic search to enable the optimized context retrieval for the agents.

00:40:44:01 So, when you think about that, you don't just want to say, oh, hey, I've got a thousand documents, I'm just going to go embed all my documents and magically, like, we've got the bad stuff. And (Unintel Phrase ___40:56) pulling back from the hybrid search index, you really have to think about the information that you really want to prioritize for the agents to retrieve.

00:41:01:05 Think about the metadata that you can extract along with those documents. You can use Gen AI to help you kind of sift through that as well. But you still need the metadata along with the embeddings around the text to actually optimize how you want to get the most relevant information to the agents at that particular point in time?

JUN QIAN:

00:41:18:13 Yeah. I just want to add one practical point, just like you’re building any machine learning system, right. And when you build an agent, right, you want to build your evaluation data set, including unstructured data to star ways.

CHRISTINE:

00:41:28:00 Hey there, this is Christine from Eli Lilly. My question has to do with that earlier question about going from web to mobile to agents. What platforms, for example, mobile, are the most relevant to invest in knowing where tech is heading and how consumer behaviour may shift particularly in the next one to two years?

00:41:50:06 We've talked a lot about agents, but we haven't talked about access and thinking about consumer engagement in that way. Thank you.

JUN QIAN:

00:42:01:14 Yeah. At Starways, I think most recently and you probably all heard about OpenAI opened the ChatGPT. Right. You certainly, you can build your application there using ChatGPT as a platform. And I think there are many, many open source softwares — LanGraph, LangChain and you can build your agent and OpenAI, I think as one example and the response APIs and how the agent SDKs and when I think about.

00:42:26:20 Right. And OpenAI is building this as a platform, just like you can launch iPhone app in the app stores. Now you can build app on top of ChatGPT. I think, other, I would assume, right. And Google and Amazon they follow the same thing.

JED DOUGHERTY:

00:42:45:04 I have a fun anecdote around this. My girlfriend just built an avatar, an interactive avatar of herself, where I can, like, talk to it like this. And then she talks back to me with, it looks like her. It sounds like her. It's just doing its own thing. I think that the text-based conversational level is going to be evolving to appearing to be truly a human conversational level and faster than we might think.

00:43:13:01 I think we'll have human conversing, yeah, within the next year or two. Pretty widespread. So maybe that'll be the next.

JACOB WILSON:

00:43:19:08 I would double down on that, especially if you look at all the investments from like, OpenAI as an example. Right. It's a lot around real-time voice. Right. And how do I and especially like if you think about developers too, right. It's one thing for me to just talk and try to, like, converse with an agent to write code for an application.

00:43:38:21 And you can be much more fluid in that conversation. Right. And provide much more, real-time feedback and accelerate the development process that way, as an example, versus going to the keyboard and trying to type. But I think a lot of like the investments that you see going on in the industry right now is around real-time voice, the multimodality. So that way you're not just sitting at the keyboard typing.

REED ALBERGOTTI:

00:43:58:08 We have a few more minutes. I love that story. How are you all using? I'm curious, like in your personal lives outside of the enterprise space? What are the agents you built?

JUN QUIAN:

00:44:12:23 So, actually, both my kids are using ChatGPT quite a lot. One is in high school, once is in elementary school. So, I think they are mostly using ChatGPT like a brainstorming and even my daughter and she quickly learns. Right. And you have to prompt the engineering carefully. Right. And to generate reusable results. And for me, I'm also using some of the coding agents.

00:44:37:18 Right. And just as a hobby during the weekend and you can clearly see how the, for example, the Codex and Claude Code. Right. And they are getting more and more intelligent.

JED DOUGHERTY:

00:44:50:04 I’m really mean to mine, I have custom instructions to my ChatGPT that's like, you should be short, concise, more of a tool, provide all you need for information. You're not human. When asked for opinion, don't comment on the question. Just write the opinion. Never encourage me. Never be witty. Shut up. So, I’m mean to mine. I'll be killed in the revolution.

REED ALBERGOTTI:

00:45:11:11 Don't be nice. Don't be sycophantic. Right?

SUDHEESH NAIR:

00:45:16:08 Oh, no. Look, I think, I'm a start-up founder in the AI space. There is no personal life, it’s all work. I'll tell you this one thing that does change significantly. Like there are a lot of geeks in the room, so I'll say product management has fundamentally changed. We used to use Figma. All of that is gone. And the reason is it used to take that you have to write a template.

00:45:35:17 You have to figure out a schema. You have to figure out the design document and then a PRD and then do. The cost of a mistake that used to be pretty high, because you take 2 or 3 months, but now you could actually write. There is no cost to mistakes, which means you've fundamentally accelerated the entire cycle to hours.

00:45:52:12 So, think about the amount of disruption, the tools that you have used. The last thing I would say — someone asked about enterprise use cases. If you are building an agentic software for enterprise, make sure that you are not focusing on saving money only because right now, everybody is running that. Saving money is table stakes. It should be about making more money and removing risk. If you can do that, it fundamentally changes.

REED ALBERGOTTI:

00:46:14:06 That's great. I think we're basically out of time unless anybody has 30-second comment.

JACOB WILSON:

00:46:20:23 Just on personal use, coming back to your question like, everything I can think of to try and apply it to just like, is exposure therapy too, right. For myself. Like, keep myself grand grounded too. So, I'm not living in all of the hype around all this stuff too. But I mean, like, I'll use it to help my son or my daughter with math homework that I can't remember.

00:46:40:07 Right? Like, taking pictures, like, have it describe the problem or like, the new way they teach kids math is interesting to me. So, like trying to understand that, also to like, just totally random stuff. Right. I think what's cool too is like, more of the agents are coming. If you're like a sports fan, like there are more things coming into mobile apps and things like that for like how you experience games and all that stuff too.

00:47:04:02 So I think it's pretty cool. I try and apply it everywhere I can.

REED ALBERGOTTI:

00:47:06:18 Yeah, well, I think I have a better sense of the agentic world after this conversation. So, thank you all for your excellent input. Thanks for coming.