Why AI Agents Can't Be Trusted Yet (And How to Fix It) | Moe Katib (One)

24:35Anne Dwane:And to be able to accomplish the goal. Correct, yeah. Because we are, like in San Francisco, people are willing to take risk. You know, like, and, you know, we are, you know, we adopt technology faster. And we, like, tend to be on the cutting edge of technology. And, like, we try things. It might not work. The tolerance of failure is pretty high. Like, there's no issues there. But most people are not like this. They are concerned, you know, like, what would happen if it leaked information? What would happen if it sent the wrong email? I don't believe that we have yet the mechanism to, like, we don't have yet that sort of system that creates the trust for people to allow higher adoptions of agentic.

25:20Because I truly believe working agents is going to free a lot of our time. It's going to get us into the sort of future that I really wish that is going to be true, which is the future where we as human, we go back to what we like to do. We go back to talking with other people, going to the communities, getting away from the computer, and trusting that the systems are actually working. And I like to use an example. If I go today into a store and I buy a drink, I use my credit card and I tap on it and I walk away. Nobody is concerned. It just happens, right? Like it's, my credit card could be Canadian and I could be in San Francisco or I could be in Spain and I would do the same tap and I would walk away and nobody will trace me, nobody will say any single word.

26:07Why? Because everybody trusts that the money is going to go to the right place. We need to do the same thing for agents. We need to make it where when I tell my agent, do something, it's gonna do it. And if it doesn't know how to do it, or if it's unsure, I know that it's gonna tell me. Right. And that's really the balance. It's like you don't need the agent to notify on everything because then you get overwhelmed. Now you're creating more work for me. You know, like if the agent is going to confirm everything with me, it's more work for me. There needs to be a balance. And what I found works really well, and this is something I'm building for myself, is you start with a lot of notification and then the agents start understanding my preference.

26:47And I start building up the trust with the agent that is like now, okay, it's replying for my emails. And at the beginning, I'm like, under no circumstances, you reply. You show me everything. And then I start like saying, yeah, this is good. This is good. This is good. This is good. And now, you know, there are certain aspects of things like completely autonomous, like a refund request, for example. I don't even need to know about it. It just happens.

27:08Anne Dwane:So you have trained it bespoke for you. Correct. And so you have onboarded and trained your agent. Okay. Well, let's talk a little bit about your company, One. It's withone.ai. Yeah, so one came as a result of all the knowledge and all the things we've done in the past. So as I mentioned, we had all this knowledge that we accumulated about these integrations. And it's not a secret that integration alone, it's not going to be a moat anymore. It's not a moat. You know, like you cannot be, you can't have a company that is just doing integration and call it a moat or even a SaaS. You know, I believe software is going to become commoditized as a whole.

27:50So I started thinking, where's the value? What am I doing? Where is the value that if someone is using me, what's the value? And I start reasoning over and over that the value is in the capabilities of this. I'm selling the capabilities. And then if I think even deeper, what does that mean? Well, it comes down to effectiveness of these agents. So first of all, I needed to make sure that neither me nor my team are living in a bubble. So I took what we thought is so valuable, which is our knowledge. And I know I am very biased, but our knowledge is really clean. And if you use it with any agent tech, it works.

28:34That's because we vet the knowledge and we have a system to scrape the documentation. Then we put it through a... So we scrape... Let's say, for example, we want to work with Gmail. We go and we scrape all the docs, and then we take whatever we scrape, we put it into other system subsystem to make sure that the knowledge we scraped can actually be executed.

28:54Anne Dwane:Okay, got it. This seems like you're – ideally, integrations would be an easy thing for agents to do because there is good documentation in theory. But in practice, it isn't always the case. So you have made sure that it's agent ready. Yeah. And then also agent always ready, meaning you're maintaining it over time. That's right. And I would argue that, yes, the agent can go on the docs and do it, but it's the Wild West because the agent can read the right doc, but it could also read the wrong doc. It could read an old doc. So you're at the mercy of the index. So what we're doing is we're like, no, we actually are vetting this knowledge to make sure that it is actually accurate.

29:37And then we run different use cases. So we have a system called the Scenario Generator, and it generates different scenarios for every action that we use. So, for example, I like to use the send email example. If we start from an agent that is like non-trained and you tell it like, the user could say something like, write me an email for Emily. That's what the user would say. And the agent would interpret this and it will say, it might write you an email because it might think it's a sample. And then it'll write the email. The agents now are a little more sophisticated. They'll ask more questions, you know, like, but you could have an agent that is not as sophisticated and it could literally write you an email and the email address will be emily at sample.com.

30:20But if that agent has the capability to also send, well, it's sent that email. So you have a tiny little issue there. So what we do is we generate these different scenarios. And it goes from someone who is not, they don't know how systems work, all the way to someone who is very capable and they understand exactly how they're using these tools. And we generate everything in between. And then we generate the success route and the failure route. And then from there, we start seeing the behavior of the agent. Like, for example, send an email to Emily. Well, which Emily? Like, how do I know what Emily you're talking about?

30:56Is this an Emily that we've been working on and I have it in memory? Or is this something I don't have and I don't know and I need to look up in the CRM? If I looked up in the CRM and there's like several Emilys, maybe it could figure out from the context, but should it make a guess? And that's the thing, you know, like, is that there is a balance. So for us, what we're doing right now is we're solving the whole schema problem and like, okay, the obvious stuff. But the scenario generator, what it's doing, it's like solving a little bit of a bigger problem. Because as you could imagine, this is going to generate a lot of knowledge, a lot of data.

31:27But there is almost always one path that is correct that the user wants. And figuring out that path is not an easy thing. But if you don't figure that, you will not have high adoption. We know how to do this for ourselves. But when you need to deploy this to the masses, it needs to just work. They don't need to think about how this works. So we are building the systems that kind of make it work. Got it.

31:50Anne Dwane:And who is the ideal user for one? There are three ideal users for one that feed into each other. We have big companies. This is like companies who have already massive agent tech and massive, in a sense, agent tech solutions such as Vipe coding platforms, platforms that are deploying agents because they want these agents to be connected to integrations and apps and all this. So we help them. It just works because we solve authentication, we solve the knowledge. And now, as I mentioned, we're getting into the skills. This is like where the value really lies. And these are vetted skills. And then we have the second one, which is San Francisco startup building agents to solve very particular niche.

32:38So you're like sales agents agents, marketing agents, you know, I would love to work with these types of startups and allow them to kind of figure out, you know, like, how do you solve this bigger picture? You know, like, this is like, okay, one agent that's doing sales and like focusing on sales. They, of course, you know, like, they're a great ICP for us because their user needs connection and we help them get their, like, we help, we make it really easy for their user to connect to these third party integrations. And then finally, prosumers. They use our CLI, and with a single CLI, you can connect to any tool that you have and build your own agent.

33:16This is the hobbyists, the developers, people who want to tinker and play with the agents and try it for themselves without necessarily a lot of opinions.

33:25Anne Dwane:And how many integrations do you offer today? I believe today I checked, I think we were at 330. Okay. So amazing progress. And you had shared previously that the rate of adding new integrations was a certain pace. And then over the past couple months, that has really accelerated. Can you talk a little bit more about that? Definitely. The important thing for us is more than the quantity, the quality. Because I could easily, like right now, for example, go and add a thousand integrations, but the quality would not be as good. Sure. We have figured out the processes that are necessary in order for us to actually say that this integration is good.

34:07So we start by scraping, but it's not a normal scraping where, like, you would, like, for example, like, when you do a generic scraping, you will miss so much data. For us, we analyze the type of technology that the docs are living in. And then based on these different technologies, we write different scripts that essentially, like, would open different tabs, have different pop-ups. We solve problems such as if you have, sometime you have a schema and you have a nested schema, but sometime you have a schema of a schema, meaning you have an invoice that has a bill inside of it, but the bill has an invoice.

34:39And if you keep opening, it keeps opening. So like if you tell an agent to do that, it'll basically go in an infinite loop. So we have encountered all these different problems and we've solved all these problems to figure out how to get to the actual value of the knowledge that we have there and get everything. But that's not enough because even if you get everything, there would be so much noise. So then we clean it up. We then have different AI system that actually cleans up the docs so that it's repeatable because we figured out a pattern that if you give to the agent, it works every time.

35:08Anne Dwane:Right. Okay. Got it. And then so you're able to abstract away all the complexity, all the maintenance over time for the agent and ultimately the end user. That's right. Yeah. And then, but what we figured out is the balance of the human in the loop. And that's really key, in my opinion. And we see the same pattern happen. At the beginning, literally everything was vetted. And I literally mean it. The first hundred integrations we've added, it was vetted by hand, every one of them. The system were not as good. The models were not as good. And I'm talking this is like in 24. But then we started noticing there's a lot of improvement in the models.

35:49And now we're getting into a point where we're getting roughly a 98 % success at first run. That doesn't mean, of course, that we handle every edge case. But because it's a network and because the knowledge is open source, you start having contribution. So when, for example, someone could be in Spain and they're using our tech and they encounter an edge case that we've never seen before, we've never even thought about. If they want, they can share that problem with us. because most likely their agent would have figured it out. And because that section of the industry, the ICP that is using the CLI, they tend to be technical.

36:29So they tend to correct the agent, which is going to help the other buckets because the other buckets, there would be no that human in the loop. So you want to get the human in the loop, which is what we're doing. And ideally, we're getting all the people who use it also become the human in the loop, which feeds the knowledge, which improves the AI. And then we want the AI labs to train on this data. Like, I want them to train on the data because I truly believe that the better, the cleaner the data, the better off all of us. And that's the thing, you know, like, it's a scary thing to open source the knowledge because, and my whole team, I had to speak with my team for hours, hours, you know, like, until, because, you know, like, at our company, you know, like, there's, I don't, it's not, there's no dictator, you know, like, oh, I say this, you know, I have to convince my team because they built it, right?

37:14I mean, my team built this.

37:16Anne Dwane:Yeah, so walk us through that decision to open source. So I started thinking about the value, and I wanted to make sure that I don't fool myself. And I found that it's really easy for me to fool myself into thinking that I have something when I don't have something. And so this becomes very important for me as a founder. We as founder, we live in a parallel universe. And we have to. We have to because we create this sort of vision and we need to believe it. And if we don't believe that vision, nobody will follow us. You know, like why do people like to work for founders, right? Because you have this big vision and you want to make it happen and like you stand for something.

37:57But if you're talking with normal people who are not aware, they will think you are like, you know, high or…

38:03Anne Dwane:Deluded, yes. Deluded. You're like, what are you talking about? You know, like we work on the computer. You know, like what do you mean? You know, like you want to invent a new thing now? You know, like what do you – it sounds crazy. So for me, there's like a balance for this. You need to like be close to the reality, yet you still need to create your vision of the future. And I find this to be if you don't know what actually is true, your vision will be completely disconnected from reality. And if you're disconnected by, say, some some fraction, that's OK. But if you are completely disconnected, then you're hallucinating.

38:37You know, like I do not want to be hallucinating.

38:39Anne Dwane:What was the hard truth that you needed to make sure you weren't fooling yourself? The hard truth for me was that knowledge is an inevitability, that eventually the models are going to get to the best knowledge. Eventually, people will clean their docks. Eventually, we will have a better mechanism for docks. So the knowledge itself was not a moat by itself because, yes, it felt like it's a moat. It felt that it made our product better at the time, but eventually it's going to run out. So this is a really bold move because you had a business operating on proprietary data to make integrations, and you decided to open source it all.

39:16Anne Dwane:So let's talk about now what's your pricing and what's the value proposition. So our pricing is really simple. My goal is to make integration free for everybody. And I truly mean it. And right now we're trying our best to make that happen. And right now what we're doing is if you're a normal user, you're building your own thing, you're a developer building your startup or any of this, you could add as many integrations as you want for free. There's no connections limit. You could literally add as many integrations as you want. We, of course, have to do a little bit of a rate limit because some people really have used this.

39:53So we have, of course, a rate limit. And then the second tier is 29, I believe, which gives you a little more rate limits. You still get the million request for free. And then above that, we charge on like small amount of money on like extra usage above the million.

40:09Anne Dwane:So for consumption-based? For consumption-based, yeah. But my goal is to like make that completely free because I want to make one like the DNS. You know, like you don't expect when you go to the Internet and you type like www.with1.ai, You don't expect that the DNS company is going to charge you for the lookup of the DNS, the location of the IP. And the same way, integration should be free. You shouldn't pay for the integration. Nobody should pay for the integration. And I want to make that happen. So where we make the money is more so on the more enterprise, the startups, when we're working with them at the scale.

40:50So, okay, I want to deploy the agents for so many customers. I want to do a lot of auth. I want to like, and so we have a product similar to Plaid, which is authentication, but it comes with a lot of restriction. Because the other thing that I want to do and we're starting to do is the user must own their connection in this economy because I want to get trust. And to me, this is really important. And I tell this to my team all the time. In today's world, you go into a place and they ask you to connect to your integrations and you connect your integrations. And now what? If you want to disconnect, how would you disconnect?

41:22You are at the mercy of this company. So we're trying to make it where you own the authentication. So like your email will be the like truth. So if someone is requesting something from you and you want to give them access, at any given point, you could revoke the access. And that is good for the startup. It's good for the companies because it builds the trust. So and again, you know, you have to trust us, you know, like and we are also being very open on how we're doing this. because I want the user who owns the connection to own the ability for them to revoke it, even from us. So one of the things we do, for example, is secret keys are one-time only.

42:00We don't even store them internally. There's a lot of mechanism that we have, and we are also continuously going to be working on to make sure that the user owns the connection. This is going to be very important because we need to trust these companies who are building these agents. And the trust is a circle. It's not one entity that you need to trust. There's going to be many entities that we need to trust. And in order for people to trust us, we need to be open. And we need to tell them what we're doing. And there needs to be no hidden agendas or anything. And that's exactly what we're trying to accomplish at One.

42:32Anne Dwane:So you've got great momentum helping agents do integrations across many platforms. How or why will One be important in the future? For me, it comes down to two things. Quality and trust. Quality in the sense that then they feed into each other. People won't trust a product that don't work. And we will not have a good product if we don't have good quality. So to me, quality matters quite a lot. The quality of the knowledge, the quality of the data. And the quality, of course, it's multifaceted. So one of the things that we're trying to do right now is work with partners. We're not claiming that we know everything.

43:12We're not claiming that we're the best at everything. We're not. But we are good at something. And this something is the infrastructure and creating the systems that actually scales. We're good at this because we've been doing this for quite a long time. But for example, what do I understand about CRMs? Not a lot, really. I can build you a CRM, but I don't understand the actual use cases. But the providers of the CRMs, they understand how this works. They understand the use cases. They understand their users. So we want to work with these partners so that we can get them to do this validated knowledge.

43:44so we can do the same thing that we did for our knowledge to them so that they can give us their validated skills so that we can now have a mechanism where... Because right now, it's the Wild West. You are hoping that things are going to work. And in many cases, it does work, but you don't know what happened in between. You're like, yeah.

44:06Anne Dwane:Or it worked yesterday. It doesn't work today. Exactly. So to scale these agentic systems, The trust needs to exist at a, the most non-technical person needs to trust the system. A baby needs to be able to use the agents and trust that it's going to work. And a completely non-technical farmer working in his farm needs to also trust the system. And to get there, it needs to be a work of many people. And that's why I truly believe, in my partnership program, I want to work with these companies so that we can understand the problem at a bigger scale and then figure out how can we create this open, it will be open data source that's evolving.

44:47And it's open. It's literally open source. Because if we have this, and again, I hope other people do the same thing too. I think it'll be good if other people do that as well. But by having the structured knowledge that is opinionated by the people who understand the problem, we're going to get the trust one step higher. And then every step we do it, it's going to get higher and higher and higher. And then eventually you're going to get into a point where it will no longer be a friction for anyone to have a sales agent, for example. And that gets us to the place where I want to be at, not 20xing, 30xing, 100xing, but getting the job done.

45:27Because that's what matters. And that's what will free people. And that's when we will achieve the ability for the new way of work. And I don't know exactly how that's going to look like. I'm sure we're going to have other problems.

45:40Anne Dwane:Well, what is your opinion on the future of work? You're going to work less, and you're going to have a lot of abundance. So in a perfect world, you will go back to your family. You'll go back to your community. You will have more events, more in-person events. Maybe you'll have a community farm where you go on your farm, and it's all for the community. And the yield of that farm is for the community. And the work that you have to do might be notifications, literally notifications. You're looking at your watch, and you're getting a nudge. I call them nudge, not even notification, because I don't like the word notification.

46:16It's a nudge. It's like your agent is nudging you very gently, and maybe at a certain time only. But while you're getting the nudge, maybe you are at the community farm. Maybe you are at a community hub. You're doing intellectually interesting things. You're painting. You're going to maybe communities that are less lucky and helping them. You're going to a poor community and trying to get them to – you're teaching them. You're taking what we have because we have access to something right now that most people don't have access to. And we are basically like the 0.01 of adopters. And the majority of the world are living in a completely different reality.

47:00And even though we feel that this reality is so true, and it is true, I truly believe that AI is revolutionary and it's going to change the lives of the world. But to me, that's the future I aim for. That's the future I want to make.

47:15Anne Dwane:Okay. So we've got to talk about your personal productivity because I find you very responsive. But you told me you have not logged into your email for a long time. That's right. So tell me about what agents are doing for you today. So I'm trying to do that system for me. So I always tell my team that I am the power user of one. So what I'm doing is I'm trying to understand, because you hear a lot of people saying, oh, my open cloud does everything. A lot of people in San Francisco, yes. I truly believe it's a non-true statement. And no disrespect, I know some people have figured it out, but I think there is a catch that not a lot of people are talking about, is that what is the cost of a mistake?

48:03And to me, for example, if the agent is running everything for me, the cost of mistake for me is very high. It's very, very high. I'm talking with partners, and if the agent mistaken one partner for another partner, leaked some information, it would be not good at all for me. It looks really bad. And I actually have a story that happened to me recently. I have an agent that I built for myself that goes over my email, and I haven't been into my email. And what it does, and by the way, I'm going to open source this system. It's actually very simple, and I like simple systems. It will go and it will grab all your emails for the last six months, and then it will read each email via a smaller model.

48:42And this model will label these emails and figure out how you write and how you reply, and then the different aspects of email you receive. So in my case, for example, I have, it'll create a rubric. So in my case, I have support. We get like people asking questions, you know, how do I do this? How do I do that? I have emails that we send to our customers, like this is the drip. And some people reply to me. And then I have investors. I have partners. And then I have notifications from like GitHub. Nudges, yeah, exactly. So my AI will label each one of them. And when I first started, the AI would just tell me, like, you have like six emails, two needs your attention.

49:27And that's it. It just stops there. So this was great. And then you're like, okay, now what? So, for example, a lot of the support stuff, the agent can reply. Like, we've had so many repetitive questions that we now have semi-perfect Q &A that the agent can automatically reply. And the agent does. I don't even know about them at all. Like I don't even know what's happening.

49:51Anne Dwane:And how do you trust that the agent worked? So the agent is instructed that should it not be 100 % certain about something to nudge me. As I mentioned at the beginning, it was more so like tell me everything. Then, okay, propose something. So like propose this. So I would see what it's proposing. And then I would look at it when it's doing it. So when I needed to do the first refund, it was nerve-wracking because I'm giving it access to Stripe. And I'm like, okay, you're going to make a refund. And I simplify things. So, for example, our refund policy is so simple. If someone asks for a refund, refund them.

50:29Sorry, it's that simple. No question asked. If someone is unhappy or for whatever reason they want a refund, we just refund them, no question asked. So that's easy for the agent. Classify as refund, make the refund. It's that simple. And, of course, Stripe takes care of the security. So there's not a lot of risk in that specific mechanism. And then there are the other areas, like investors, for example, investors relationship. I have a list of VIPs that under no circumstances, the AI replies to them. So this VIP list keeps growing. So like if I'm working closely with a client, for example, I add them to the VIP.

51:06I literally go say, this is a VIP. And then it gets added to the VIP. And then the AI will just tell me what's happening. I also built a memory system that is a graph system where every person that I interact with has its own graph, knowledge graph. So the agent, when it receives an email, it pulls on that graph. It reads everything it needs to read. And the graph has also things to forget. So I tell the agent, for example, forget about that thing. And then it completes it, forget about it. The more I use something, the more it's surfaced into the memory and it ranked based on that. So the agent have access to all this.

51:44so the agent makes these sort of decisions and then of course i give it my preferences but then something happened recently right now i told my agent that like new investors you know like we're not fundraising at this moment so we're like heads down we're like we're heads down working right now we want to get there's something we need to do we need to do it right now but i get a message from an investor and typically my agent just ignores it like it just ignores it but this investor was a good investor. And my agent literally nudged me, saying, I know you said we shouldn't set up a meeting with investors right now, but I really think you should meet with this particular investor.

52:25Wow. Yeah. It was a big VC. So I was like, okay, set it up. And sure enough, set up the meeting in person. And then we get a, in the same day of the meeting, we get a message to reschedule for a different time. Same day, but different time. So my AI nudged me. It's like, oh, this happened. What do you want to do? And I'm like, yeah, make it happen. There's no problem. I checked my schedule. I'm like, yeah, no problem. This works for me. It's no big deal. The AI sends a really rude message. uh so i you know like i i went to the email and i'm reading the email i'm like oh my i felt really bad because this is not something i would send at first i'm like oh my god what am i like it just doesn't feel like me right like it's like so i go back to the ai i'm like man you know like what like what what did you do you know like why did you do it this way and the ai literally just said He's like, well, you said, you know, like you're a CEO, you want to sound confident.

53:27He asked for a reschedule. Your time matters. You know, like, so I wanted to make sure that it looks like you're not happy about it, but you are accommodating. And so I started thinking, I'm like, in a way, he's not wrong. It's like, but it's not how I would do it.

53:44Anne Dwane:So is the right answer, the agent should have its own identity? Because it's sending as you. It's sending as me. Yeah, that's the problem, which I think in the future, I don't know if I, who's responsible? Well, I don't know, but we could have a whole new categorization of gatekeeper. Yeah, yeah. Because at the end of the day, like, I was responsible for that email. And I felt bad, right? Like, I felt bad. Yeah, yeah. It's so interesting. I mean, these are the things we're going to have to deal with in the future. Yeah, yeah, 100%. Tell us about what is One, which is withone.ai. So one is an infrastructure for, it's an agentic infrastructure.

54:24And we're trying to give our users access to everything they need from an agentic standpoint. Right now, we have worked on integration, but we're now getting into the next cycle, which is, so authentication with integration is solved. Knowledge is solved. The next thing is skills. This is like combining multiple different knowledge pieces into a skill that does something that's repeatable. The next one would be a combination of skills that makes a role. But then you have also memory. Memory, I don't think memory has been solved also. So this is something we're trying to also figure out how can we solve memory.

55:05The other thing that we want to get into is how do you deploy these agents? How do you trust that they are actually working? So at the bottom of it, it's infrastructure for agents, agentic infrastructure. But we see it as multiple faceted, multi-bucket faceted integration, which everything you need from integration, authentication to the knowledge, to the access to the communication with this knowledge, and also access control. Right. It's very, very important. Not a lot of people are giving a lot of attention to this. We have refined access control that you can give to your agent. You can say, like, I want this to be read only, write only, Gmail only, send draft only.

55:41Like, you go at the level of one action. And then you can do it across multiple different platforms, which is, again, going to be very important for trust. Because we sell trust. We're increasing the level of trust for people so that they can adopt agentic.

55:54Anne Dwane:Yeah. It's kind of like the elevators of the AI age, right? Like it just has to work. It just has to work. And so your customers are large enterprises, they're startups, and they're individual developers sometimes. Can you provide a use case that's a good sample? 100%, yeah. So there is a really good use case for finance, for example. So one of our early users, and he's a big fan, and also he works at an enterprise company that we're working with. But he has his own use case where he has his own side businesses, and his wife also has the same thing. And he was trying to get his QuickBook information agentically, meaning he would be able to handle all his finances completely agentically.

56:41So he tried many different solutions. I'm not going to name, like, yeah, because I have full respect for all my competitors. But he used a lot of our competitors, and he literally went one by one and tried them, and it just wouldn't work. Something would be either, like, missing some stuff. Others would just not work. And, like, the trust was not there. There are certain tools he tried didn't even work at all. This is his words. he tried at the time Pika and it just worked. And he was like, my agent started pulling the data and he started finding issues and mistakes and stuff. And he built an entire system completely agentically to manage all his business finances.

57:26And I truly mean it. It's an agent that actually takes care of everything. So you could put a receipt, it takes on that receipt, it could like upload it to your quick book and it does the crunching everything that is needed for you to run your business from a finance point of view he built it agentically using one and and to me like when i hear these stories and and there are so many other stories similar to this you know like i have another one of like our early investor also like early adopter you know like same thing you know like he's built so many uh of the the thing that he does on a day-to-day on one and when i hear these stories, it just makes me really happy because it's providing value.

58:07And like when people tell you like, you know what, your tool solved a big problem that I have. It just makes you like, it just makes all the hardship and all like that. It just goes away. It literally goes away because this is amazing. These are some of the ideal use case. And of course, you know, like, as we start working with more partners, we will be able to isolate and more craft these solutions in a way that would be opinionated, but it'll be opinionated, in my opinion, in an effective way.

58:38Anne Dwane:There's a lot of buzz about MCP, and you're doing integration. What's different? So MCP is a phenomenal tool. It was obviously designed by Anthropic Model Context Protocol, and the intentional design for it was me on my computer, I want to connect to something, and I install the MCP, and it connects to something that exists already on my computer. And the intention was for developers. I have a very controversial opinion on it, but it's grounded from engineering. What has happened is that there were massive adoption of MCP, which is great, amazing. But what happened is that a lot of companies got forced into adopting to MCP in a wrong way, in my opinion.

59:26What ends up happening is we have an API, and then most companies started doing an MCP as another layer that mimics the APIs. So now, okay, I already have the APIs, which we perfected. If you're Notion, if you're these companies, you perfected your APIs. It's your offering. But now you have this other layer, which we don't know how it's going to scale. We don't know how it's going to be used. And all it does, it connects to your APIs. But it has its own authentication. Well, guess what? API also have authentication. So what are you doing? Are you passing the same authentication? There's so many problems in the structure because you created this new layer that's completely unneeded.

1:00:04So at scale, you're going to start feeling the pain. So if you are to manage these MCPs, they were never meant to be done this way. It was meant to be that I have my MCP. I put it on my own server. And that's exactly what we do right now. We are stepping away from remote MCP, but we offer our own MCP. Meaning if you want to host the MCP yourself, there will be nothing wrong about that because that's your usage. I would still tell you not to do that. But if this is what you want, it's no problem. You can have the MCP. And the difference between integration and the MCP is MCP in most cases is limited by the number of tools you have.

1:00:42And there are ways now that are making it better where you could search, you could index certain things. But no matter what, you're going to still be capped. Because let's say you want to connect to six platforms. Each MCP is going to eat something. It's going to use some tokens, maybe 1 ,000, 2 ,000, sometimes 20 ,000. We've done a calculation where if you're using four, in the worst case, you're using 80 ,000 tokens. And you haven't done anything yet. Wow. Yeah. And so imagine this. Every single time you're doing something, you're telling your agent hello. Your agent has already used 80 ,000 tokens.

1:01:23Anne Dwane:Interesting. It's not efficient. Okay. And you are someone who is building and seeing agents being deployed. What's your outlook on the landscape of agents? What are the most promising agents, the most promising frameworks, any insider view for us? There are so many innovation happening in this space. And I have been really, like I've been using Anthropic quite heavily. So I'm talking here specifically from my own personal experience. Anthropic seems to like nail it on some sort of a balance I feel on the agent side like from a model perspective where I don't know if the word entity is the right word that I want to use but it kind of feels like an entity yeah it feels like an entity it feels like it has a personality that's like and I've heard a lot of people saying the same thing.

1:02:15ChildGBT will get the job done and And again, I don't want to be harsh by any chance. I mean, they are phenomenal companies, all of them. But from my own personal experience, I noticed that the personality of Claude Asian, Claude Code specifically, it's evolving with me. And I noticed, for example, like pushback, a lot of pushback on things that actually matter. And I'm starting to see this happening. and I think that's intentional. There's some sort of a balance that's being added and I am starting to feel it. And now if I jump ship, I come back because I'm like, it feels like you left your friend.

1:02:59There's this sort of, I don't know about that feeling, but I've heard a lot of other people saying the same thing. I have a lot of friends who tell me the exact same thing. Both can get the job done. But I think, I don't know if they're talking about it, but there's some sort of a personality to it. It feels human.

1:03:18Anne Dwane:Anything you'd like folks to walk away with? Yes. My message to the people who are hearing a lot of noise and a lot of like, you know, you're going to lose your job and like all that rhetoric, no, you're not going to lose your job because if it wasn't for you, we wouldn't be here. We wouldn't be like what we're doing is for the human and the human is the value and the people is the value and the communities is the value and us is the value. We need to make the country better. We need to make us better. And we need to, the human is the value and you are human. And therefore, everything that we're doing from an agentic, from an AI is to serve human.

1:03:56So as far as you're serving other human, you're not going away. And that's literally my message, you know, like remind yourself that you're serving something that is bigger than you. And if that's the case, you're not going to lose your job.

1:04:08Anne Dwane:Okay. Well, we will send people to find you at withone.ai. Withone.ai. Thank you so much, Mark. Thank you so much for having me.

1:04:19Hey, this is Ben Kassnoka, co-founder of Village Global. Thanks so much for tuning in to the Village Global podcast, where we go deep on all of the biggest topics in tech. If you enjoyed this conversation, please subscribe to our YouTube channel. You can check us out on Spotify, Apple, wherever you get your podcasts. We'd love to see you for the next one.