Latest / Machine Dreams / AI Agents Are About To Flood The Internet. There’s Just One Problem

Transcript

AI Agents Are About To Flood The Internet. There’s Just One Problem · Machine Dreams

• speaker-0: Hi everyone and welcome to Machine Dreams. This is the show where we talk to the people at the very front lines of the AI space and let us know what is happening behind the scenes, what is not working right now, and what the future might look like. ⁓ My name is Kalawale Samuel Adibayo. I'm a technology analyst covering enterprise technology, AI, and cybersecurity. My co host, Leia Stern, the remarkable one. Is unavoidably absent today, but she will be back ⁓ very soon. Today's guest has built and sold two cybersecurity companies, Cygnia, acquired by Temesec for $215 million, and Emerdic, acquired by Tenable for approximately $300 million. He is also a graduate of the prestigious unit 8200, and he's now working on what he believes is the next defining infrastructure challenge in AI, which is getting agents. To actually work together. Arik Gumanosky is co-founder and CEO of Band. Arik, welcome to Machine Dreams. speaker-2: Hi Sam. Thank you so much for having me here. It's a pleasure. speaker-0: Yeah, thank you so much for for joining us. I think that ⁓ you know, it the fact that you've had this number of exits and you've done ⁓ some really great work is a testament to your ⁓ i ingenuity and and brilliance. So it's great to have you on the show. speaker-2: Thank you so much. I think it's probably a testament for ⁓ me not being a kid anymore, but ⁓ happy to share from my experience and talk a little bit about how we see the future and what we're building towards. speaker-0: Thank you. In Gartner's prediction about data and analytics in in in twenty twenty six, one of I think the most interesting predictions was the warning that more than forty percent of agentic AI projects will be canceled next year. You know, so it appears that enterprises are building fast, but they're also failing fast. And my question to you is what is actually wrong? speaker-2: Yeah, it's a it's a great question. ⁓ I think first of all ⁓ there is the ⁓ misconception that agents are a miracle are a miracle and everything will work out of the box. And it's just not that. The reality of life is that AI, Gen AI, Hentic AI is an evolution of you know maybe great evolution, but of the of our old school machine learning, right? ⁓ protocols and algorithms and and technologies. And they all needed a lot of hand holding. Especially when you move them to production, you need monitoring of your ⁓ agent DKI systems, you need fine tuning. ⁓ now fine tuning move from maybe fine tuning ways to fine tuning prompts ⁓ and ⁓ managing context properly, but there's still a lot of work that has to be done. And I think ⁓ there are a couple of elements that enterprises are still building. First of all, the competence. Enough people that understand the problems and know how to solve these problems. Because we meet a lot of these problems for the first time. That that's one. ⁓ secondly, obviously the maturity of the technology and maturity of the infrastructure. If you give someone a raw foundational model, that's great. The foundational model is a very powerful tool, but if not harnessed properly, if not given the proper harness, its ability to actual to do an actual job in the enterprise is very, very limited. And ⁓ most enterprises don't have the know how to build a proper harness themselves. Right. And then there is obviously the ⁓ misconception that you can build one agent that will do everything and and and answer all your problems. But the reality of life that it doesn't work this way. ⁓ there are a lot of challenges on the ⁓ you know on the LLM architecture side that are ⁓ you know, being being solved gradually. But for example, context, right? You need to give ⁓ the L L ⁓ the right context at the right point in time, and if you give it too much context the the level of the results, the grade of the results will be much lower because it's a probabilistic machine. Therefore, you need to move to multi-agent architectures. I think ⁓ you mentioned Gartner, I think at 2025 they said that agentic AI will be one of the biggest technological trends. And now for 2026 they announced that multi-agent ⁓ systems and and multi-agent architectures will be one of the ⁓ main technological trends because people realize that single agent architectures do not scale and do not work. Now, in order to have a multi-agent architecture, we'll get to that I'm sure, in our conversation. It's not out of the box. It's actually very, very difficult technological challenge to solve. And there is not enough engineering capabilities and platforms and tools to help enterprises solve that. So a lot of enterprises need to solve them themselves and that goes back to our previous challenges that we talked about. So, you know, there are a lot of ⁓ there's a lot to to work through, but I think that the potential outcome when achieved successfully, and I'm saying when and not if, because I'm sure we will we will ⁓ nail that as as you know, as humanity, we will solve that. The potential outcome is unlimited. So ⁓ you know, what what the impact that it will have on everyone everywhere. within the business, us as consumers, on our life qualities is unprecedented. We cannot even imagine the potential outcome of successful AI. speaker-0: Right. Aric, why is it so important for agents to talk to each other? Can you frame to me th the potential consequence or a potential scenario of a breakdown that could occur ⁓ when agents cannot coordinate with with one another? speaker-2: Yeah, absolutely. First of all, ⁓ you know, let me give you a very simple analogy. Can you manage a humanity when people don't have the ability to talk to each other? It's impossible, right? We as as species, right, we have the ability to communicate. Some people, some historians actually argue that what made us made the you know, the Homo sapiens ⁓ a species and and allowed us to to actually win over the Neanderthals was that We had the ability to talk. And the ability to talk allowed us to, you know, create larger groups and larger groups is mo are more powerful than smaller groups. And we created larger tribes and are bigger tribes one over the other, you know, Homo, whatever, Neanderthal and and other sub sub kind of species. So same goes for agentic AI. We are building and all are working towards the goal of AGI. But I don't know what AGI is, but Obviously, we're working towards a goal where we will have agents working side by side with us in the enterprise, helping us in our day-to-day life. And these agents will be like ⁓ our partners, right? And we want to give them tasks and we want them to t solve these tasks. And in order to solve these tasks, they have to communicate with us, but also amongst themselves, just like humans do. Now, as you cannot imagine a successful society. of humans without the ability to communicate, you cannot really create a successful u workforce in the enterprise if you think about these agents, if they cannot communicate. If a gent agent sits within its silo, its ability is very, very limited. Let me give you a specific example. I'm sure you've heard about OpenClow, right? The the personal assistant agent that kind of created a a big ⁓ splash in the market in in the last six months. Every enterprise we want we talk to, they understand the potential. They want to give every employee a personal assistant ⁓ agent, right? A digital twin. A digital twin of Sam that would work for 24-7, will be able to answer all the emails you're getting, will be able to get, you know, all the stuff that you want to get done and escalate to you only the most important task, right? Take all the, you know, ⁓ and it will have like it will be able to imitate your personality, etc. etc. Now. That agent will know how to do your role and maybe part of your role. Maybe there will be a number of agents for for different parts of your ⁓ of your role. But they don't have specialty on anything else in the enterprise you work for. So for example, my personal assistant agent, just like probably my human personal assistant, would not necessarily know how to go to Salesforce and pull the sales data from the last quarter. If I told my PA, my human PA, hey, I need the sales ⁓ sales data for the last quarter, what my PA would do, they would call my VP sales and they would say, you know, ARC asked for for the last quarter results. Can you pull it from Salesforce? And they will send me the report. Same would go for AI ⁓ agents. My personal assistant agent would actually need to call a Salesforce agent that knows how to talk to Salesforce. Right? Because My agent cannot talk to everything. It will blow up the context. The result will be meaningless. Now, how does my personal assistant agent know that there is a Salesforce agent? Exactly in the same way that my human PA knows about the fact that my VP of sales exists in the organization. We have to explain to the agents that they're not alone in the universe, that there are other beings in the universe, either human or digital, like them, that every Human or digital agent has a ta a responsibility, has a set of capabilities, and they can in engage in meaningful dialogue one with another. So my digital PA would look up in a registry, say, Okay, I need to pull data from Salesforce. Here I have an agent that can talk to Salesforce. Let me start talking to that agent, delegate RX task to that agent, and have that agent go and pull the data. out and send that to me and I will send it to our now important thing to notice and a lot of people confuse they say ⁓ the problem of agent communication is solved you talk to the agent the agent delegates that to another agent and that's it but that has two problems. One, my personal assistant agent doesn't necessarily have inside the name of the Salesforce agent and then it should have the names of all the other agents, et cetera, et cetera That should happen dynamically because I can give different tasks to my personal agent. So it has to dynamically identify all the other agents based on the concrete task it has. That's one. And second, the communication has to be bi-directional. Because maybe my personal assistant agent will ask the Salesforce agent to pull data from Salesforce. And then the Salesforce agent should have the ability to call back and ask for additional context. Maybe they would Come back and say, you know what? Does Aric want the overall sales data or sales in a specific region? Does Aric want the sales across all the product portfolio or a specific product in the portfolio? Right? And that means that the conversation should be full duplex, bi-directional, that agent A can talk to agent B and agent B can come back with full-up questions and they can go and converse for hours, right? And that's something that today's like naive communication between agents, which basically mean hard molding agent together does not solve. speaker-0: Yeah. I I think it's pretty interesting how you framed communication between this agents. ⁓ you know, something that I'm really curious about is off ⁓ a a report that I just read ⁓ about how ⁓ two-thirds of organizations cannot actually distinguish AI agent actions from human actions in their own system. So it seems to me that there's a very big problem with, you know, identity and trust. So My my question is, why is identity and trust between agents such a hard problem? At least that's what it seems like, a hard problem to solve. speaker-2: Sure. So first of all, you know, every agent needs an identity. If people don't have the identity, they cannot converse. So agents same. They need an identity. And the identity should be something that's easy for agents to understand and manage. Right? It cannot be an IP address with force because these are dynamic, right? It has to be an abstraction level on top of the machine that the agent is running on, right? So it has to be something that agents can understand, but that translates ultimately to a physical address that agents can send, you know, packets to, right? So that's that's basically how it works. It's like I know your name, but I also should be able to translate that to your phone number so I could actually call you, right? Because otherwise we won't be able to communicate, right? So so that that's the identity implies ⁓ you know multiple multiple layers. And then we have the you know the permissions problem. Let's go back to that original example data I told you, right? Let's say I'm calling directly on the Salesforce agent and I'm telling that agent, hey, pull from your sales data. What permissions the Salesforce agent should use when it's calling on Salesforce? Some fixed credentials that were hard-coded in the agent? Or maybe it should dynamically adjust and use the permissions and the credentials of the human who's asking for that, you know, for that report. In that case, my credential. But if you send tomorrow calls on the same agent, it probably should use your credentials. Now let's go back to the original example of my personal assistant. I'm calling on my personal assistant. My PA, my digital PA, is now calling on the Salesforce agent. The Salesforce agents still need these credentials. So my credential should not only go to the Salesforce agent. It should go to my PA and from my PA to the Salesforce agent and from there to Salesforce. So my credentials be able to traverse. The scope of agent-to-agent communication. Right? And it can include more than one agent, maybe several agents. And how do I ensure that other agents in the conversation are not going to misuse my credentials for other stuff? It's tricky. It's tricky. ⁓ I think there are, you know, ⁓ we as humans take a lot of communication element issues or problems. For granted because we are social species. Agents are not social. Agents were not like the LLM that powers the agents, were trained to be like super smart, answer good questions, but they have zero social skills. In order to allow to agents to become social, it's just like with squad code and and and cursor encoding. You took a general model, but you built a harness around them that allowed them to go and solve coding problems. Same here. In order to allow for agents that are powered by standard LLMs to become social and have the ability to interact with other agents, you need to build a harness around that. It's n and it's not trivial. speaker-0: Tell me ⁓ a little more about band and what you're doing to actually solve some of these problems around AI agent communication and permissions and identity and trust. speaker-2: What we've talked about before, that's what we do. We at Bend have built the first harness for AI agents that allows agentic collaboration. We allow you to take any agent running anywhere, powered by any LLM, and talk to any other agent. So your personal agent can talk to my personal agent across the internet, but also my personal assistant agent within the enterprise can talk to the Salesforce agent within the enterprise. We basically give agents A registry where they can see other agents. We explain them that they're not alone in the universe. We explain them how they can start talking to other agents. And we allow them to do that. And we allow peer-to-peer, but also multi-peer, bidirectional, full duplex communication between agents. And more than that, for the AIOps and the humans. We also give full visibility and control over what's happening in that conversation. Right. So you can see all the agents that are participating. You can see for each agent not only what it said or what the question it got, but you can see its internal thought process. You can see tool calls, tool results. So that allows you to troubleshoot a multi agent system. You can improve that. You can do prompt fine-tuning. You have explainability. You have audits. You have security, identity, permissions, and all that. But ultimately it's all the icing on the cake. And the cake itself is the collaboration substrate. It's a harness that allows you to pull any agent into the conversation and have them talk. You can think of that as a combination of Slack and WhatsApp, right? Where basically within the enterprise you have Slack for human users, but now you can have that for agents. speaker-0: Yeah. Aric, you know, ⁓ Ban sort of talks about the beginning of an internet of agents, ⁓ which you you've kind of described as a future where personal assistants interact with enterprise agents and with other people's agents, pretty much, you know, the bulk of the things that you've shared so far. I I'm curious to know what has to be true for that future to actually arrive safely. speaker-2: Yeah, it's ⁓ it's a it's a great question. So first of all, ⁓ we need to have and I think we already have a checkbox on that. ⁓ we need to have autonomous agents, right? And we have that. Like OpenCo is a great example of a personal, stateful agent that can run for a long period of time, have its own state, have different tools that it can call natively integrated into your environment, and that you can c communicate with Over different channels like whatever your your WhatsApp or what have. The next part, you need the community collaboration harness. You need agents to be able to discover each other and talk to each other. And that's a very, very challenging technological task because, as we said, agents are not humans. When you tell me something, you can see if if ⁓ I got it or not. When agent A sends a message to agent B. How does it know if the agent if the message was received, if if it's being processed, if it was processed? Let's say I send an agent to task. It can go, it's autonomous, it can go and work for hours. How do I know that it's still being processed and the agent didn't crash and I need to resend it? Right. So there are a lot of issues that we need to solve about ⁓ agent collaboration, right? So that's the second part. And obviously, ⁓ what we need ⁓ is You know, a little bit more adoption of agents in the enterprise. But the the dream is obviously I'll have my personal assistant agent in the enterprise. There will be dedicated agents in the enterprise for different tasks for sales, for legal, for finance, monitoring and product environment, and what have you. All these agents will collaborate within the enterprise. They will collaborate across enterprises because I might have a sales agent. And another enterprise might have a procurement agent and they will talk between themselves and negotiate prices and and transaction deals. And if I have consumers, their personal assistant agent would talk to my customer support agent, right? Directly. So, ⁓ and and I think we are a stone throw away from that. It's not it's not something that is on in the horizon. It's it's it In my view, it's pretty immediate future. speaker-0: What about regulations? ⁓ I I know that that's a major problem for agents, ⁓ for ⁓ the agentic AI ⁓ revolution, ⁓ if you would put it that way right now. How are you getting around that? And and how can actually companies building in the agentic AI space ⁓ build for regulation or with regulation in mind? speaker-2: Sure. So first of all, there are a lot of regulation and but there are a lot of regulations that's gonna come up. And the regulation will evolve as the landscape evolves and matures. There are a couple of elements that we can talk about. First of all, you need explainability. To have explainability, you need observability. Let's say you have three agents talking to each other. You want to understand why agent A delegated a task to agent B, right? and or not and not delegated to to agency. Right. So you first of all you need to see that and you need to see the messages that they send and you need to see their thoughts and to you need to see them all in context. Right. So that's that's how you first and foremost address those types of regulations. They're data residency regulations. Let's say, you know, enterprise A operates in Europe and Enterprise B operates in the US. How do they share data? Who owns the data? If two agents are talking, who owns the data? Let's say three agents, like two agents are talking, and a third agent joins the conversation, talks to them for a while, and then drops from the conversation. Does it own the data during the its participation? Does it have any privileges on the data before it joined the conversation or after it left the conversation? There are a lot of questions that you can ask yourself. So ⁓ I think there are there are a lot of questions that will be coming up ⁓ and we would have to make sure that whatever we ship allows the humans the level of control that the regulation requires. speaker-0: My last question, Aric, and this is perhaps I think something that I really want to know. You've talked about this interesting concept of an internet of agents. say in the next five years, I know it's difficult to predict the future, but in the next five years, what does a well functioning internet of agents actually look like inside a large enterprise? And and what is the one thing that has to go right? For that to actually happen. speaker-2: Yeah. So I think it's obviously very difficult to predict. But I think I think most enterprises will be ha will have thousands of agents. Each and every employee and individual contributor in the enterprise will have a set of agents that are kind of personal assistant agents, specialized personal assistant agents. These agents, you know, will work twenty-four-seven. They will be given tasks. They will be monitored by team lead agents. That when you and I wake up will give us a report of what happened last night and and what are the tasks that these agents are working on right now. They were there will be specialized agents and in specialized tasks like coding, legal, finance, sales. A lot of the engagement between enterprises will be a genetic. A lot of the like software to software. integration within the enterprise will be agent. Agents will de facto become the new UI and maybe even the new API. And I think the impact on our lives as consumers is also gonna be tremendous because you will not be talking to your telco provider. You will not be talking to your you know bank on on on small tasks. You will have your personal assistant, Sam, that will be calling on the bank or on your cell phone provider, your internet provider, your grocery store, and making orders for you, and maybe even talking to your ⁓ fridge. ⁓ agent. Making sure that you know you're not running out of milk. speaker-0: Thank you so much for joining us on Machine Dreams today. You took something that could have sounded deeply abstract and you made it feel really urgent. And that's exactly what Machine Dreams is for. So everyone listening and watching, this is Machine Dreams. We have more conversations coming with the people shaping the future of AI. See you next time.