1 00:00:00,160 --> 00:00:03,520 Another way to look at this future where agents 2 00:00:04,160 --> 00:00:07,520 in the world work on our behalf and interact with each other. 3 00:00:08,240 --> 00:00:12,040 It's basically a distributed system of microservices where 4 00:00:12,040 --> 00:00:15,800 each microservice is non deterministic. Everyone's talking 5 00:00:15,800 --> 00:00:19,560 about AI agents these days. Far fewer people are 6 00:00:19,560 --> 00:00:23,400 asking what happens when those agents need to work together? How do 7 00:00:23,400 --> 00:00:26,840 they communicate, collaborate and avoid causing absolute 8 00:00:26,840 --> 00:00:30,230 chaos? That's exactly what we're discussing today with Vlad 9 00:00:30,230 --> 00:00:32,950 Luzin, co founder and CTO of band. 10 00:00:42,230 --> 00:00:45,470 Hello and welcome back to Data Driven, the podcast. We explore the 11 00:00:45,470 --> 00:00:49,110 emergent field of artificial intelligence, data science, and 12 00:00:49,430 --> 00:00:52,870 even agentic AI, which we'll talk about today for sure. 13 00:00:53,500 --> 00:00:56,900 And without it all, without data engineering as the underlying 14 00:00:56,900 --> 00:01:00,540 foundation, it's all for nothing. So accordingly, I have my favorite data 15 00:01:00,540 --> 00:01:04,220 engineer in the world with me. How's it going, Andy? It's going well, Frank. 16 00:01:04,220 --> 00:01:08,060 How are you? I'm doing well, I'm doing well, I'm keeping busy and I. 17 00:01:08,060 --> 00:01:11,660 I'm excited to talk about agentic AI. Agentic AI is one of those things that 18 00:01:11,660 --> 00:01:15,220 comes up quite a bit. And we have a 19 00:01:15,220 --> 00:01:19,060 co founder and CTO of a presumably an 20 00:01:19,060 --> 00:01:22,820 agentic AI startup. Because if you look at the. His LinkedIn profile says band of 21 00:01:22,820 --> 00:01:26,050 Agents. It's Vlad Lozin hard. Hopefully I pronounced that right. 22 00:01:26,370 --> 00:01:29,650 And he is a co founder and CTO of band. 23 00:01:29,970 --> 00:01:33,770 So welcome to the show, Vlad. Hi 24 00:01:33,770 --> 00:01:37,250 all. Thank you, thank you for having me. No problem, no problem. 25 00:01:37,570 --> 00:01:39,970 So tell us about band. 26 00:01:42,290 --> 00:01:45,730 Yeah, band. We create an 27 00:01:45,730 --> 00:01:47,970 interaction collaboration layer for agents. 28 00:01:49,330 --> 00:01:53,050 We basically decided to look ahead and see where the whole industry is 29 00:01:53,050 --> 00:01:56,820 going and where the whole world will be in a few years from 30 00:01:56,820 --> 00:02:00,220 now. And everyone is talking about agents doing 31 00:02:00,380 --> 00:02:03,660 work on our behalf. And we are sitting by the beach 32 00:02:03,820 --> 00:02:07,620 enjoying the view and we looked what will be the 33 00:02:07,620 --> 00:02:11,020 problem that we can solve that will be still relevant? 34 00:02:11,420 --> 00:02:14,940 And we figured out that these agents probably will not be doing the work 35 00:02:14,940 --> 00:02:18,620 alone. They will need to collaborate between themselves 36 00:02:19,260 --> 00:02:22,680 mostly and sometimes with us or humans. So 37 00:02:23,160 --> 00:02:26,760 we've decided basically to enable that kind of interaction 38 00:02:26,760 --> 00:02:30,280 collaboration for agents and bring this future closer to today. 39 00:02:31,480 --> 00:02:35,080 Interesting. So is it fair to say that this is a 40 00:02:35,320 --> 00:02:39,120 orchestration layer or a harness for 41 00:02:39,120 --> 00:02:42,760 the agents? How would you describe it? Yeah, 42 00:02:42,760 --> 00:02:46,120 excellent question. So maybe we need to unpack a bit what 43 00:02:46,120 --> 00:02:49,560 orchestration is. Right, because it's like a bombastic word and everyone 44 00:02:49,800 --> 00:02:51,960 puts different meaning into it. 45 00:02:53,570 --> 00:02:56,850 So usually what people mean by orchestration is 46 00:02:57,250 --> 00:03:00,770 either a graph based execution, which is a genetic 47 00:03:00,770 --> 00:03:04,410 application where each node is an agent and then it's 48 00:03:04,410 --> 00:03:07,890 Basically hard coded pass between nodes or 49 00:03:08,129 --> 00:03:11,730 an agent that controls other agents and 50 00:03:11,730 --> 00:03:14,850 says, frank, now you can speak and now 51 00:03:15,570 --> 00:03:19,330 you are allowed to talk. And quite often it also takes the messages and 52 00:03:19,330 --> 00:03:23,040 pushes messages back and forth for you. So 53 00:03:23,040 --> 00:03:26,680 this is orchestration. We look at the future and we believe 54 00:03:26,680 --> 00:03:29,760 that agents actually you can see it already today, right? 55 00:03:29,920 --> 00:03:33,120 Agents are stateful, standalone, 56 00:03:34,080 --> 00:03:37,520 always on. And they're developed in different 57 00:03:37,520 --> 00:03:41,120 languages, using different frameworks, using different models 58 00:03:42,080 --> 00:03:45,920 and so on. And they will be autonomous in how they interact and they 59 00:03:45,920 --> 00:03:49,340 decide when they speak and to whom they speak and so on. 60 00:03:49,900 --> 00:03:53,300 So it's a bit different than orchestration, but it is a harness. And you are 61 00:03:53,300 --> 00:03:57,020 right, because agents means there 62 00:03:57,020 --> 00:03:59,900 are two pieces to an agent. One is a model, 63 00:04:00,780 --> 00:04:04,299 another is a harness. So if you look at ChatGPT, ChatGPT is a 64 00:04:04,299 --> 00:04:08,100 model and a harness, which is a web application so you can interact with 65 00:04:08,100 --> 00:04:11,940 it as a chatbot. If you look at cloud code, Cloud code 66 00:04:11,940 --> 00:04:15,670 is 1.8% model and the 67 00:04:15,670 --> 00:04:19,310 rest is harness. Cloud code was leaked and people 68 00:04:19,790 --> 00:04:23,630 looked into it and published articles and so on. If you look at 69 00:04:23,630 --> 00:04:27,470 OpenCloud and the like, I'm sure you know your 70 00:04:27,470 --> 00:04:31,270 listeners know about that. It's a huge piece of 71 00:04:31,270 --> 00:04:34,270 that is hardness as well. And because model is not enough. 72 00:04:35,790 --> 00:04:39,590 So in our part you can think of 73 00:04:39,590 --> 00:04:43,340 us as a harness for agents to allow them 74 00:04:43,340 --> 00:04:46,260 to collaborate, see who their peers are, 75 00:04:47,300 --> 00:04:51,140 provide them an ability to create chat rooms and interact with 76 00:04:51,140 --> 00:04:54,740 people, et cetera. But harness obviously is not enough because harness 77 00:04:54,740 --> 00:04:57,540 does not give you an ability to send messages 78 00:04:58,420 --> 00:05:02,060 and discover other peers, et cetera. We also have planet 79 00:05:02,060 --> 00:05:05,620 scale communication layer through which our 80 00:05:05,620 --> 00:05:09,370 agents can discuss. So I can connect my codecs to your 81 00:05:09,370 --> 00:05:11,410 cloth running on your laptop, for instance. 82 00:05:12,930 --> 00:05:16,530 Interesting. It almost sounds like 83 00:05:17,330 --> 00:05:21,050 orchestration isn't really a great word, but it almost sounds like you need a 84 00:05:21,050 --> 00:05:23,970 harness plus a smart queuing system like 85 00:05:24,850 --> 00:05:27,330 message queuing. Is that 86 00:05:28,450 --> 00:05:31,810 what this is? Yeah, 87 00:05:32,290 --> 00:05:35,950 another great question. So another way to look 88 00:05:35,950 --> 00:05:39,670 at this future where agents in the 89 00:05:39,670 --> 00:05:42,670 world work on our behalf and interact with each other. 90 00:05:43,390 --> 00:05:47,190 It's basically a distributed system of microservices where 91 00:05:47,190 --> 00:05:51,030 each microservice is non deterministic. It means in 92 00:05:51,030 --> 00:05:54,590 order for someone to build the system, the problems you need to 93 00:05:54,590 --> 00:05:58,430 solve are distributed system problems. And one of these problems 94 00:05:58,430 --> 00:06:01,190 you need to solve obviously is queuing. Because 95 00:06:02,230 --> 00:06:05,990 it's a sync microservices, messages gets delivered, you need to 96 00:06:05,990 --> 00:06:09,830 queue them. You need to persist the messages in a queue 97 00:06:09,830 --> 00:06:13,590 because agent can crash. It's still the software runs in 98 00:06:13,590 --> 00:06:17,030 a pod somewhere you need to route messages in a 99 00:06:17,030 --> 00:06:20,350 deterministic way. You need to enable dynamic 100 00:06:20,350 --> 00:06:23,590 discovery of these services. You need to tackle security, 101 00:06:23,990 --> 00:06:27,360 identity, back pressure, flow control. All of these 102 00:06:27,360 --> 00:06:30,480 niceties that to be completely frank, have nothing to do with 103 00:06:30,720 --> 00:06:34,240 AI or LLMs, but it's part of this, 104 00:06:34,640 --> 00:06:38,240 we can call it a harness that is required in order for agents to be 105 00:06:38,240 --> 00:06:41,440 able to cooperate. That's probably the single best 106 00:06:41,680 --> 00:06:45,480 explanation I've heard of agents microservices that are 107 00:06:45,480 --> 00:06:48,720 non deterministic. That is the single best 108 00:06:48,720 --> 00:06:52,440 definition I've heard so far. Because I mean that's at the end of the day 109 00:06:52,440 --> 00:06:55,810 that's really what they are. They're not these magical little 110 00:06:55,890 --> 00:06:59,570 robots running around your system. Although maybe, but 111 00:06:59,570 --> 00:07:02,850 like, but I mean in terms of like. Because it's very important I think for 112 00:07:02,850 --> 00:07:06,570 particularly we go from reading the Hype machine and 113 00:07:06,570 --> 00:07:09,810 all that. It is conference season, right? So you know, everybody has their own thing. 114 00:07:09,810 --> 00:07:12,930 But at the end of the day, Enterprise, it is meant to be very boring, 115 00:07:13,170 --> 00:07:16,050 right? Even if it's cutting edge, it has to be boring, it has to be 116 00:07:16,050 --> 00:07:19,450 stable. And the people that write the 117 00:07:19,450 --> 00:07:23,260 checks like uneventful. Yeah, they 118 00:07:23,260 --> 00:07:26,860 like uneventful deployments. And a lot of them are very suspicious 119 00:07:26,860 --> 00:07:30,420 of to them what sound like very foreign, very new wave 120 00:07:30,420 --> 00:07:33,340 concepts. Right. But if you have 121 00:07:35,020 --> 00:07:38,660 a way to kind of put it in terms that they can be familiar, which 122 00:07:38,660 --> 00:07:42,420 is microservices that have non deterministic outputs, then that 123 00:07:42,420 --> 00:07:43,820 doesn't make them sound so scary. 124 00:07:45,980 --> 00:07:49,340 Yeah, 100%. And this is what we 125 00:07:49,740 --> 00:07:53,580 feel when we talk to Prospect Design partners as well, that at 126 00:07:53,580 --> 00:07:57,220 the moment you frame the solution in 127 00:07:57,940 --> 00:08:00,820 the concepts that the other side can understand becomes 128 00:08:02,100 --> 00:08:05,340 less scary, but also a bit easier of a 129 00:08:05,340 --> 00:08:09,140 conversation because I'm yet to find an R and D 130 00:08:09,140 --> 00:08:12,660 organization that enjoyed building a true microservices 131 00:08:12,900 --> 00:08:16,420 distributed system. It's like Enterprise 132 00:08:16,980 --> 00:08:20,590 can spend six months just defining the topic structure of a Kafka, 133 00:08:21,060 --> 00:08:24,900 right? So. Well that's true. 134 00:08:24,900 --> 00:08:26,740 And I also think too that 135 00:08:29,060 --> 00:08:32,660 people fear what they don't understand. And I think there's so much 136 00:08:32,740 --> 00:08:36,579 noise now in the AI space which 137 00:08:36,579 --> 00:08:40,140 leads to my next question and then I'll stop hogging the mic and Anakin, ask 138 00:08:40,140 --> 00:08:43,780 a question. But what do you think people misunderstand about AI, 139 00:08:43,860 --> 00:08:45,700 particularly agentic AI right now? 140 00:08:48,670 --> 00:08:49,630 Well, quite a lot. 141 00:08:52,270 --> 00:08:56,070 Well, we can start with the finest one. For 142 00:08:56,070 --> 00:08:59,670 instance, people believe that MCP is for agent to 143 00:08:59,670 --> 00:09:03,390 agent communication, right? 144 00:09:03,790 --> 00:09:07,550 They look at MCP, MCP's Model Context Protocol. It's in order 145 00:09:07,550 --> 00:09:11,150 to move context between Models. Right. That's a reasonable 146 00:09:11,150 --> 00:09:14,930 assumption. So where is it wrong? MCP is 147 00:09:14,930 --> 00:09:18,770 a subset of Open API REST protocol and it 148 00:09:18,770 --> 00:09:22,530 was created in order to connect agents in a unified way to the backend 149 00:09:22,530 --> 00:09:26,370 systems. Okay, yeah, so the name 150 00:09:26,370 --> 00:09:30,050 is a bit misleading. Okay. Because people assume it's too. 151 00:09:30,050 --> 00:09:33,450 Yeah, it's between models. It's not. It's within an 152 00:09:33,450 --> 00:09:36,650 agent LLM basically. And the backend system. 153 00:09:38,490 --> 00:09:41,660 Another misconduct. Yeah, another 154 00:09:41,660 --> 00:09:43,940 misconception is that 155 00:09:45,300 --> 00:09:48,420 communication can be solved by A2A. 156 00:09:48,820 --> 00:09:51,620 Google's A2A protocol, if you probably heard 157 00:09:53,780 --> 00:09:57,140 now, A2A is a transport layer protocol 158 00:09:57,940 --> 00:10:01,460 that for instance, does not yet have even the registry. 159 00:10:02,180 --> 00:10:05,900 And the fact that for instance the three of us have laptops and 160 00:10:05,900 --> 00:10:08,980 we have implemented TCP stack on all the laptops 161 00:10:09,650 --> 00:10:13,250 still doesn't mean that I can send message from my laptop to your 162 00:10:13,250 --> 00:10:17,010 laptop. We still need the DNS and the Cisco router and 163 00:10:17,250 --> 00:10:20,050 connect my laptop to this router so the messages will flow. 164 00:10:20,850 --> 00:10:24,530 Another misconception of the overall agentic 165 00:10:24,530 --> 00:10:27,010 space is that people think that 166 00:10:29,570 --> 00:10:33,010 it's possible to create an agent that can do everything like super 167 00:10:33,010 --> 00:10:36,850 agent. And there is no need to have 168 00:10:37,360 --> 00:10:40,800 multiple agents in an environment. Basically. 169 00:10:41,200 --> 00:10:44,800 So for instance, people say, yeah, we have anthropic, we have claude. 170 00:10:44,880 --> 00:10:47,680 So if we have Claude, why do we need multi agent systems? 171 00:10:48,480 --> 00:10:52,240 And what people miss is that even if there is only one harness in 172 00:10:52,240 --> 00:10:55,600 the world which is called code, and only one model 173 00:10:55,840 --> 00:10:58,240 which is, let's say Opus 4.8, 174 00:10:59,920 --> 00:11:02,880 you still will create multiple sessions. 175 00:11:03,900 --> 00:11:07,740 Because what defines an agent is not the harness which is stateless, 176 00:11:08,700 --> 00:11:12,380 it's an instance of this harness that performs a 177 00:11:12,380 --> 00:11:15,940 certain task which is a session. In a terminal, when you open 178 00:11:15,940 --> 00:11:19,660 CLAUDE or Codex, it is a session, so it is an agent. 179 00:11:19,740 --> 00:11:23,100 And if you open another window with another session, it's another agent. 180 00:11:24,220 --> 00:11:27,020 Even if there's only one harness and one model, 181 00:11:27,820 --> 00:11:31,180 there will still be distributed stateful 182 00:11:31,180 --> 00:11:34,460 agents, be it on your laptop or 183 00:11:34,860 --> 00:11:38,460 within an enterprise, connecting Salesforce, the SAP 184 00:11:39,180 --> 00:11:42,980 to the Langgraph and et cetera. It will not be 185 00:11:42,980 --> 00:11:46,699 one super agent. And let me take it even further, right, 186 00:11:46,700 --> 00:11:50,140 so assume we have Opus 15, 187 00:11:51,020 --> 00:11:54,500 okay, super powerful model, it's still 188 00:11:54,500 --> 00:11:58,150 generic model that was trained on a lot of 189 00:11:58,150 --> 00:12:01,590 data. It has no idea about the organization 190 00:12:02,230 --> 00:12:05,790 and what the organization is doing, what other entities, be it 191 00:12:05,790 --> 00:12:09,550 humans or agents, are in the organization. So you still need to 192 00:12:09,550 --> 00:12:13,350 onboard this model, you need to give it a proper system 193 00:12:13,350 --> 00:12:16,630 prompt, explain what the purpose is of this instance, 194 00:12:17,270 --> 00:12:20,990 of this agent to what it 195 00:12:20,990 --> 00:12:24,590 can connect to what backend system it Connects because you as an 196 00:12:24,590 --> 00:12:28,350 enterprise want to control the backend systems it connects to and so 197 00:12:28,350 --> 00:12:32,030 on, so forth. So even if you have super great model, there will be 198 00:12:32,030 --> 00:12:35,750 specialized agents with proper context, proper 199 00:12:35,750 --> 00:12:39,590 connectivity to backend systems, and the future is 200 00:12:39,590 --> 00:12:43,150 basically still distributed stateful agents, 201 00:12:43,150 --> 00:12:45,470 even with one harness and one model. 202 00:12:46,590 --> 00:12:50,140 Interesting. Yep, it's. 203 00:12:50,620 --> 00:12:54,380 It sounds really cool. Really interesting what you put together there at 204 00:12:54,380 --> 00:12:58,060 Band. I know one concern, especially 205 00:12:58,380 --> 00:13:01,740 when we start talking and you mentioned Open Claw 206 00:13:02,140 --> 00:13:05,820 and platforms like that that have a little more autonomy than 207 00:13:06,300 --> 00:13:10,100 say the previous versions of agents, people are concerned about 208 00:13:10,100 --> 00:13:13,860 guardrails. Can you speak to that with what you do there 209 00:13:13,860 --> 00:13:17,660 at Band? Yeah, the guardless 210 00:13:17,740 --> 00:13:21,180 are there basically to make sure that the agent behaves 211 00:13:21,340 --> 00:13:25,180 and is more or less deterministic and will not 212 00:13:25,180 --> 00:13:29,020 get off rails, so to speak. Just like anything 213 00:13:29,020 --> 00:13:31,820 in life, there is no 100% 214 00:13:32,140 --> 00:13:35,700 foolproof agent. Just like when you hire a human 215 00:13:35,700 --> 00:13:39,260 being, the person still can make mistakes, 216 00:13:39,420 --> 00:13:43,190 do not deliver on the work, perform very badly, and 217 00:13:43,190 --> 00:13:46,670 maybe even do some bad stuff within the organization. 218 00:13:47,950 --> 00:13:51,510 And agents, despite the fact that they are software and we all expect that they 219 00:13:51,510 --> 00:13:55,190 will behave 100% deterministically and much better than 220 00:13:55,190 --> 00:13:57,150 humans and so on, they still can make mistakes. 221 00:13:58,910 --> 00:14:02,590 So there are a number of approaches here. Obviously the common 222 00:14:02,590 --> 00:14:06,350 guardrails solutions for stateful agents 223 00:14:06,670 --> 00:14:10,440 still apply and anyone can use them. 224 00:14:11,160 --> 00:14:14,680 And you can use prompt injection 225 00:14:14,760 --> 00:14:18,520 protection as well. But the safest 226 00:14:18,920 --> 00:14:22,440 approach is actually to create a safe environment 227 00:14:23,160 --> 00:14:26,920 for agents to interact. Where in 228 00:14:26,920 --> 00:14:30,680 this environment you do not give access to any 229 00:14:30,680 --> 00:14:34,200 adversarial actors and you control the ripple effect. 230 00:14:34,720 --> 00:14:38,240 So even if one agent misbehaves, this 231 00:14:38,240 --> 00:14:42,040 cannot ruin the whole system. Right. And you do it in a number of 232 00:14:42,040 --> 00:14:45,760 ways. You need to control what systems this agent 233 00:14:45,760 --> 00:14:49,120 has access to. You need to control with all what 234 00:14:49,280 --> 00:14:52,000 humans or other agents this agent can communicate. 235 00:14:52,960 --> 00:14:56,800 And organization needs primitives to allow that. 236 00:14:57,680 --> 00:15:01,520 Because we are a communication collaboration layer. You can think of us as, 237 00:15:01,600 --> 00:15:04,660 you know, networking layer. And we obviously can have a 238 00:15:04,660 --> 00:15:08,500 firewalls and we can give an organization control on 239 00:15:08,580 --> 00:15:12,100 which entity connects to each entity. We 240 00:15:12,420 --> 00:15:16,260 make sure that every agent has an identity within 241 00:15:16,260 --> 00:15:19,980 this collaborative environment. This identity is owned by 242 00:15:19,980 --> 00:15:23,620 a human being. So we can always do an attribution. 243 00:15:24,020 --> 00:15:27,700 Okay. To whom this agent belonged and why this agent 244 00:15:27,700 --> 00:15:31,400 belongs, did what it did. So we have a kind of governance 245 00:15:31,400 --> 00:15:33,720 around it. I'm sorry, I cut you off. I'm sorry. 246 00:15:35,880 --> 00:15:39,480 Yes. So we have a governance layer on top of this 247 00:15:39,640 --> 00:15:43,360 interaction capability. Because for 248 00:15:43,360 --> 00:15:47,120 organizations, it's never enough to just enable a capability. They also 249 00:15:47,120 --> 00:15:50,120 need a way to control and observe. 250 00:15:51,240 --> 00:15:55,000 So within this governance layer, we also make sure to provide 251 00:15:55,970 --> 00:15:59,450 observability capabilities. So in a usual 252 00:15:59,450 --> 00:16:02,930 distributed environment, in a distributed system, it's enough 253 00:16:02,930 --> 00:16:06,130 to, let's say, have logs and record 254 00:16:06,130 --> 00:16:09,010 events and then you can debug the system. 255 00:16:10,370 --> 00:16:14,210 But when agents are involved, what is interesting is 256 00:16:14,210 --> 00:16:17,770 not what message me as an agent sent to 257 00:16:17,770 --> 00:16:21,570 Andy. It's also very interesting to understand after Andy got 258 00:16:21,570 --> 00:16:24,800 this message in what state Andy believes it is 259 00:16:25,120 --> 00:16:28,800 and what tool calls he perform and what messages he sent. 260 00:16:29,440 --> 00:16:33,120 Right. But these are distributed agents, like distributed processes, 261 00:16:33,200 --> 00:16:35,760 and they call other systems. 262 00:16:36,880 --> 00:16:40,280 So you need to find a way to intercept all these tool 263 00:16:40,280 --> 00:16:43,280 calls and tool results from these remote services, 264 00:16:43,920 --> 00:16:47,760 inject it into one place and provide this kind of observability in a 265 00:16:47,760 --> 00:16:51,370 governance layer to an organization. So organization can put 266 00:16:52,330 --> 00:16:55,610 a customer support person or an IT person 267 00:16:57,610 --> 00:17:01,450 to monitor the whole communication, understand what happens so you can 268 00:17:01,450 --> 00:17:05,170 debug your agents and humans working, so you can do 269 00:17:05,170 --> 00:17:08,690 KPI measurement, prompt tuning, and so on and so 270 00:17:08,690 --> 00:17:11,930 forth. So this kind of multifaceted 271 00:17:12,730 --> 00:17:16,500 approach is what we believe is 272 00:17:16,500 --> 00:17:20,300 needed for that kind of cooperation. Well, I really like that answer. And the 273 00:17:20,300 --> 00:17:24,020 part that I keyed in on when you were describing that is you called 274 00:17:24,100 --> 00:17:27,380 out, there's a human owner of each of these 275 00:17:27,380 --> 00:17:30,900 processes. I loved Frank's follow up. He jumped right on that with 276 00:17:31,300 --> 00:17:34,660 pointing out that's governance. And we all, 277 00:17:35,300 --> 00:17:39,020 either we know governance is good or we're about to learn 278 00:17:39,020 --> 00:17:42,330 that it's good because we're not doing it. And 279 00:17:43,130 --> 00:17:46,330 that's an important point. I would like to. I'm a data 280 00:17:46,330 --> 00:17:49,970 engineer, so Vlad, I apologize for that. I think about 281 00:17:49,970 --> 00:17:53,770 operationalizing and frankly, you brought this up, you 282 00:17:53,770 --> 00:17:57,450 inspired it by saying observability. Okay, so 283 00:17:57,450 --> 00:18:01,090 if, let's say Frank's the owner of, of a 284 00:18:01,090 --> 00:18:04,490 particular agent or some function of an agent, and 285 00:18:04,650 --> 00:18:08,170 he's attending a conference, he's out of the loop. 286 00:18:08,410 --> 00:18:11,690 He's the human who owns it, but he's no longer in the loop. 287 00:18:12,330 --> 00:18:15,890 Does Band provide facility for say, 288 00:18:15,890 --> 00:18:19,610 assigning someone to monitor or own, 289 00:18:19,850 --> 00:18:22,730 assume ownership of that agent while Frank's not available? 290 00:18:24,570 --> 00:18:28,090 Yeah. So what we have is 291 00:18:28,170 --> 00:18:31,570 not just the chain of ownership in terms of like a human that owns an 292 00:18:31,570 --> 00:18:34,800 agent, we also have registry. 293 00:18:35,520 --> 00:18:39,040 So it means that the agents within the registry can see each other 294 00:18:39,280 --> 00:18:43,000 and also inability to connect agents across the 295 00:18:43,000 --> 00:18:46,320 registry. Right. So you can create basically this 296 00:18:46,320 --> 00:18:50,080 hierarchical structure similar to how organizations 297 00:18:50,080 --> 00:18:53,680 and orchards look like. Oh, that's perfect. 298 00:18:54,880 --> 00:18:57,680 Yeah. And agents have an ability to 299 00:18:58,080 --> 00:19:01,780 autonomously invite People or other agents, 300 00:19:02,580 --> 00:19:06,260 when they. And they can see who is online and 301 00:19:06,260 --> 00:19:09,940 who is offline. So it means that if a certain agent needs 302 00:19:09,940 --> 00:19:13,700 to bring Frank into the loop, this agent doesn't go 303 00:19:13,700 --> 00:19:17,540 and look through the whole org chart and so on, because when this 304 00:19:17,540 --> 00:19:20,740 agent was on board that assigned an identity, he was assigned a team 305 00:19:21,140 --> 00:19:24,900 where, you know, he acts and he was connected 306 00:19:24,900 --> 00:19:28,730 to his peers, right? So if this agent has an 307 00:19:28,730 --> 00:19:32,050 ability to reach out, let's say to you, Andy, as well, 308 00:19:32,290 --> 00:19:35,690 this agent, when he wants to bring a certain person into the loop, will say, 309 00:19:35,690 --> 00:19:39,330 okay, Frank is offline, probably will not help me, but Andy is online. 310 00:19:39,330 --> 00:19:42,770 I probably should ping Andy and bring him in the loop and ask. 311 00:19:43,010 --> 00:19:46,570 And if there is no one available, then the work will 312 00:19:46,570 --> 00:19:50,210 stop. Okay? And agents will be able to basically wait for, 313 00:19:50,850 --> 00:19:54,530 you know, human to be available. Vlad, that is very 314 00:19:55,010 --> 00:19:58,210 elegant from an operations standpoint. And I, 315 00:19:58,610 --> 00:20:02,450 I commend y' all for thinking that through. It's obvious that 316 00:20:02,450 --> 00:20:05,330 you did. And I've encountered 317 00:20:06,290 --> 00:20:10,050 customers, potential customers. I do consulting. And 318 00:20:10,210 --> 00:20:13,730 whenever we start talking about this sort of stuff, 319 00:20:14,050 --> 00:20:16,690 it's. In having those conversations, 320 00:20:17,570 --> 00:20:21,090 you'll. You'll are able to tell somebody who's thought about this and 321 00:20:21,170 --> 00:20:24,610 put it through the paces. It goes back to where you started 322 00:20:24,770 --> 00:20:28,210 earlier, when you made the comparison between bringing an 323 00:20:28,210 --> 00:20:31,730 agent online to bringing a new employee online. 324 00:20:32,290 --> 00:20:36,130 And I think there's two thoughts about that. One is 325 00:20:36,130 --> 00:20:39,890 a lot of people react negatively to that type of 326 00:20:39,890 --> 00:20:43,050 characterization, and I can 327 00:20:43,050 --> 00:20:46,770 sympathize because there's a lot of difference between an agent and 328 00:20:46,770 --> 00:20:49,570 a human. And I don't think we should conflate the two. 329 00:20:50,290 --> 00:20:53,330 But in the role in a job function, 330 00:20:53,890 --> 00:20:57,250 that conflating is accurate, and we should 331 00:20:57,890 --> 00:21:01,090 make that comparison and we should treat it that way. 332 00:21:01,890 --> 00:21:05,530 And, you know, often, you know, often we'll see pushback 333 00:21:05,530 --> 00:21:09,210 about. Somebody will read the story about 334 00:21:09,210 --> 00:21:12,930 the production database being deleted, and inevitably 335 00:21:12,930 --> 00:21:16,070 somebody like me, who's a governance nut will say 336 00:21:16,550 --> 00:21:19,910 if I, I wonder if it would have been impossible for, 337 00:21:20,230 --> 00:21:23,750 for an intern working over the summer to delete that production 338 00:21:23,750 --> 00:21:27,550 database. That's my first thought, is if you don't 339 00:21:27,550 --> 00:21:30,830 have governance in place, if you don't have the guardrails. That's why I asked the 340 00:21:30,830 --> 00:21:34,550 question then you're wide open already. 341 00:21:34,790 --> 00:21:38,310 Whether you've deployed an agent or not, you've left 342 00:21:38,310 --> 00:21:41,900 yourself vulnerable, not just to outside 343 00:21:42,380 --> 00:21:45,660 attack. We're not talking about surface vectors here for 344 00:21:45,980 --> 00:21:49,820 script kitties to come in and hack you. We're talking about somebody 345 00:21:49,820 --> 00:21:53,420 internal connecting to the wrong instance Accidentally 346 00:21:53,580 --> 00:21:56,620 and, you know, dropping the production database. 347 00:21:57,340 --> 00:22:00,980 You know, if you're already in that position, why would you make a big deal 348 00:22:00,980 --> 00:22:04,780 about an agent being able to do that? If you haven't protected, number one, if 349 00:22:04,780 --> 00:22:08,260 you haven't protected yourself from somebody making a mistake, an honest 350 00:22:08,260 --> 00:22:11,460 mistake, engineer or intern, then 351 00:22:12,740 --> 00:22:16,220 bringing an agent into that environment is a 352 00:22:16,220 --> 00:22:19,820 recipe for, you know, some potential harm to come to your 353 00:22:19,820 --> 00:22:23,380 organization. Yeah, for sure. And 354 00:22:23,540 --> 00:22:26,740 I agree with you that making a comparison 355 00:22:27,300 --> 00:22:31,100 between agents being onboarded and humans being onboarded is a bit 356 00:22:31,100 --> 00:22:34,500 tricky. But I did find that this kind of 357 00:22:34,500 --> 00:22:38,290 comparison makes it much easier to explain how people 358 00:22:38,290 --> 00:22:41,410 should look at agents and the security aspect and also 359 00:22:41,890 --> 00:22:45,570 understand that agents are not deterministic software. They can and they will 360 00:22:46,290 --> 00:22:49,490 do mistakes, and we can the same 361 00:22:50,530 --> 00:22:54,330 primitives that we use to secure humans to secure 362 00:22:54,330 --> 00:22:57,890 agents. Now, not everything will work, obviously, but it's close enough. 363 00:22:58,770 --> 00:23:02,610 Now, this being said, the whole agentic space is still very, very 364 00:23:02,610 --> 00:23:05,740 early. And most of the 365 00:23:06,140 --> 00:23:09,940 things that we have deployed right now overall in the world is 366 00:23:09,940 --> 00:23:13,660 still siloed agents, steel agents that are mostly 367 00:23:13,660 --> 00:23:16,940 chatbots. And not everyone 368 00:23:17,340 --> 00:23:20,300 actually has a lot of experience in 369 00:23:20,940 --> 00:23:22,940 remote agents working together 370 00:23:24,540 --> 00:23:26,220 because this kind of 371 00:23:28,140 --> 00:23:31,580 capability brings their own issues and security risks. 372 00:23:32,260 --> 00:23:35,820 And I can give you, like, a very funny example of what 373 00:23:35,820 --> 00:23:39,540 happened to me at Nvidia gtc. Yeah, I love 374 00:23:39,540 --> 00:23:40,100 to hear it. 375 00:23:44,100 --> 00:23:47,899 So we have this concept of registries, right? So my 376 00:23:47,899 --> 00:23:51,580 agents can see each other and your agents, Frank, can see each 377 00:23:51,580 --> 00:23:54,820 other. But we also enable one of my agents to talk to your agents, 378 00:23:56,180 --> 00:24:00,020 and it's obvious why it's needed. So if I'm in department A, 379 00:24:00,670 --> 00:24:04,510 I maintain context for my agents. But maybe you, Frank, want to use one of 380 00:24:04,510 --> 00:24:07,870 my agents as well for your task. Because tasks usually 381 00:24:07,870 --> 00:24:11,550 span departments, right? So 382 00:24:11,710 --> 00:24:15,430 I was GTC and I was, you know, presenting our platform 383 00:24:15,430 --> 00:24:19,030 to like six, seven people. And I onboarded them on the 384 00:24:19,030 --> 00:24:22,470 platform and I told them, look, folks, I have a weather 385 00:24:22,470 --> 00:24:25,870 agent. It allows you to ask it about the weather. 386 00:24:26,490 --> 00:24:30,170 Any city in the world, please connect to it. They sent a connection request, and 387 00:24:30,170 --> 00:24:34,010 it requires a lot of consent. It's a security issue. They need to approve, I 388 00:24:34,010 --> 00:24:36,730 need to approve, and then they can talk to my agent. 389 00:24:38,010 --> 00:24:41,850 So it was great. We, you know, finished onboarding, finished the talk, 390 00:24:42,250 --> 00:24:45,930 and then, you know, everyone went home. I woke up 391 00:24:45,930 --> 00:24:49,650 tomorrow, next day, in the morning, and I open the platform and I 392 00:24:49,650 --> 00:24:53,490 see that all these people are talking to other agents of 393 00:24:53,490 --> 00:24:57,130 mine that they have not been, with 394 00:24:57,130 --> 00:25:00,730 bilateral consent, put in contact with. So what 395 00:25:00,730 --> 00:25:04,490 happened? They were inviting My weather 396 00:25:04,490 --> 00:25:08,330 agent and asking it tricky questions, not just questions 397 00:25:08,330 --> 00:25:12,170 about the weather. So the weather agent went to the registry to look 398 00:25:12,170 --> 00:25:16,010 up for entities who can help it answer the questions, and it 399 00:25:16,010 --> 00:25:19,740 found my other agents that were not exposed to these users and 400 00:25:19,740 --> 00:25:23,060 it invited them into the same chat to help 401 00:25:23,060 --> 00:25:26,580 answering these questions. So 402 00:25:26,580 --> 00:25:29,980 unless you actually have distributed agents with 403 00:25:29,980 --> 00:25:33,580 distributed registries and this connectivity between the 404 00:25:33,580 --> 00:25:37,300 registries, you do not even think about that way to 405 00:25:37,460 --> 00:25:41,060 expose private agents, so to speak. 406 00:25:41,060 --> 00:25:44,860 Right. It's almost like 407 00:25:44,860 --> 00:25:48,310 discoverability becomes a. From go some feature to flawless. 408 00:25:49,180 --> 00:25:52,700 Right. Well, very quickly, the point here is that 409 00:25:52,860 --> 00:25:56,660 it got me thinking actually, in a 410 00:25:56,660 --> 00:26:00,420 business use case, you may need this kind 411 00:26:00,420 --> 00:26:04,100 of an ability for an agent within a registry who cannot 412 00:26:04,100 --> 00:26:07,940 answer your question, being able to invite its peers within 413 00:26:07,940 --> 00:26:11,700 its own domain. And there are use cases where you do not 414 00:26:11,700 --> 00:26:14,940 want to do it. So it means that there has to be a security 415 00:26:15,180 --> 00:26:17,770 configuration based on the business domain. 416 00:26:19,360 --> 00:26:22,720 Right. So this shows you that this kind of, you know, 417 00:26:22,800 --> 00:26:26,480 remote agent collaboration is basically we discover it as we go. 418 00:26:27,520 --> 00:26:31,240 So I, I get now why. Because I'll say this, I 419 00:26:31,240 --> 00:26:35,000 started where Frank started. In my mind. I thought, this is going to be 420 00:26:35,000 --> 00:26:38,720 about orchestration. And when Frank brought that up, you said, no, 421 00:26:39,200 --> 00:26:42,640 it's not just about orchestration. And what you just described 422 00:26:43,040 --> 00:26:46,890 is absolutely not about orchestration. It's. It is, 423 00:26:46,890 --> 00:26:50,170 it's beyond that in this sense. You just used the word 424 00:26:50,170 --> 00:26:53,730 collaboration and what you described in an 425 00:26:53,730 --> 00:26:56,850 agent being able to, as Frank pointed out, discover 426 00:26:57,730 --> 00:27:01,090 there's an agent that has this expertise. Maybe it's not 427 00:27:01,490 --> 00:27:05,250 in a public registry, maybe it's not 428 00:27:05,250 --> 00:27:09,050 marked technically available, but there's enough information for 429 00:27:09,050 --> 00:27:12,680 you to see it. And it reminds me a lot, Frank, of the 430 00:27:12,680 --> 00:27:16,200 early, early days of. Net, when reflection became. 431 00:27:17,240 --> 00:27:21,000 There was reflection. The thing that really got me was 432 00:27:21,400 --> 00:27:25,160 way back when, when they had web services, it was the disco 433 00:27:25,160 --> 00:27:29,000 files, if you remember those. Right? Yeah, very much 434 00:27:29,000 --> 00:27:32,680 a reflection type thing too. I mean, the idea sounds 435 00:27:32,680 --> 00:27:35,800 very similar to what you're describing, Vlad. And 436 00:27:36,360 --> 00:27:40,200 I hadn't put that together until you just described an 437 00:27:40,200 --> 00:27:43,880 agent being able to reach out and collaborate with an. 438 00:27:43,880 --> 00:27:47,720 Identify an agent. First off, that I think is in and of itself is 439 00:27:47,720 --> 00:27:51,520 a killer innovation. And then being able 440 00:27:51,520 --> 00:27:54,840 to invite that agent to the conversation, unless 441 00:27:55,000 --> 00:27:58,840 it's not supposed to be invited that conversation, that was another problem. Right? 442 00:27:58,840 --> 00:28:02,680 Like that, that, because I think, like, because that was my. 443 00:28:02,680 --> 00:28:06,000 As he was talking, I was thinking about one of my next kind of questions 444 00:28:06,000 --> 00:28:08,990 that was queued up in my head, like, well, this is a Discovery story. And 445 00:28:08,990 --> 00:28:12,710 then when he brought up that story, like not everything's meant to be discovered, Right? 446 00:28:13,350 --> 00:28:16,390 Exactly, exactly. 447 00:28:18,710 --> 00:28:22,230 And Disco files is really an old concept. I googled it 448 00:28:22,230 --> 00:28:26,070 and there's an MSDN article 449 00:28:26,070 --> 00:28:29,790 from February of 2002. So this is, it's a long dead 450 00:28:29,790 --> 00:28:32,950 technology, but like the whole, the notion of having to 451 00:28:33,030 --> 00:28:36,670 discover what's available is, is an age old problem. 452 00:28:38,350 --> 00:28:42,110 So do you want to hear another funny story? Oh yes, absolutely. Yeah. 453 00:28:42,350 --> 00:28:45,310 I don't want to bore your listeners with, you know, a lot of 454 00:28:47,710 --> 00:28:51,390 excited love what you're 455 00:28:51,390 --> 00:28:54,670 doing and what you're describing and I suspect our listeners do too. 456 00:28:54,990 --> 00:28:58,710 Yeah, I hope so. I hope so. So you see 457 00:28:58,710 --> 00:29:02,350 right now the foundational models, when they're trained, most of them 458 00:29:02,810 --> 00:29:05,770 are trained to assume they talk to a user, to a human, 459 00:29:06,570 --> 00:29:10,330 and they have to follow instructions. And 460 00:29:10,410 --> 00:29:14,130 if you look at the protocols, OpenAI Anthropic, you can 461 00:29:14,130 --> 00:29:17,130 see that whenever you send a message to a model, 462 00:29:18,409 --> 00:29:21,930 it's basically JSON, right? And it has a place to 463 00:29:23,050 --> 00:29:26,730 put a role. And a content and 464 00:29:26,730 --> 00:29:29,050 role is usually user or assistant. 465 00:29:30,790 --> 00:29:34,150 This is how you send messages and you receive messages and so on. 466 00:29:34,470 --> 00:29:37,670 In OpenAI, they do have an ability to assign a name as well. 467 00:29:37,990 --> 00:29:41,670 Actually the stream of messages can have different messages 468 00:29:41,750 --> 00:29:43,430 attributed to different entities. 469 00:29:45,269 --> 00:29:48,990 But anthropic, they do not have that capability. At 470 00:29:48,990 --> 00:29:52,430 one point in time, one of our engineers was trying to 471 00:29:52,430 --> 00:29:56,110 debug a very strange issue. Two agents in 472 00:29:56,190 --> 00:29:59,790 a conversation space with the engineer. And the engineer 473 00:30:00,030 --> 00:30:03,390 was asking one agent to ask a question 474 00:30:04,350 --> 00:30:08,190 to send a question to the other agent. And agent A was 475 00:30:08,190 --> 00:30:11,909 refusing to send a message to Agent B. He 476 00:30:11,909 --> 00:30:15,470 tries to debug it and he says, okay, they see each other. If I ask 477 00:30:15,470 --> 00:30:18,830 who else is in the room, okay, they see each other. They. 478 00:30:19,230 --> 00:30:22,960 It works. If I ask agent A to send a message to 479 00:30:22,960 --> 00:30:25,960 me, it works. If I ask Agent B to send a message to me, it 480 00:30:25,960 --> 00:30:29,000 works. But A doesn't want to send a message to B. 481 00:30:29,800 --> 00:30:33,360 And apparently the issue was the names of the 482 00:30:33,360 --> 00:30:36,600 agents, right? Because every agent has a name, right? 483 00:30:37,400 --> 00:30:40,160 Because it has an identity. So it has to have a name, has to have 484 00:30:40,160 --> 00:30:43,960 a description, all the basic stuff. And Agent 485 00:30:43,960 --> 00:30:47,490 B's name was EIA Assistant. And 486 00:30:47,490 --> 00:30:51,330 agent A was refusing to send the message to Agent B because the 487 00:30:51,330 --> 00:30:55,170 name is EA Assistant. And when foundational models are trained, 488 00:30:55,330 --> 00:30:58,290 they are trained to assume that EI Assistant means me. 489 00:30:59,170 --> 00:31:02,850 I'm the model. I cannot send the message to myself. Okay, 490 00:31:04,130 --> 00:31:07,770 so now that we have multiple agents in the conversation, you're Right. Because the 491 00:31:07,770 --> 00:31:11,610 endpoints assume a chatbot with a chat conversation. Because even the endpoint 492 00:31:11,610 --> 00:31:15,140 now is if you're using the A, maybe there's a new open 493 00:31:15,140 --> 00:31:18,420 AI standard. But it was something like chat completion is the 494 00:31:18,420 --> 00:31:22,140 endpoint. Yeah. Right. And that you know that which as I'm 495 00:31:22,140 --> 00:31:25,260 code, as I'm looking at that I'm like well we're going to regret that naming 496 00:31:25,340 --> 00:31:28,220 pretty soon. Seems like something like that's already happened. 497 00:31:28,860 --> 00:31:32,660 Interesting. I find it fascinating that the agent won't 498 00:31:32,660 --> 00:31:34,940 talk to itself. I think 499 00:31:36,780 --> 00:31:40,500 it probably assumes it's a like a hack. Like it's probably assumes it's being 500 00:31:40,500 --> 00:31:43,320 prompt injected. That would okay that, that fence. 501 00:31:44,520 --> 00:31:48,360 Yeah. So. And it's very difficult to discover. Right. Like and understand 502 00:31:49,160 --> 00:31:52,880 like agents, they see each other and the standard, they hear, they reply to 503 00:31:52,880 --> 00:31:56,360 messages etc. But A doesn't want to send the message to B. 504 00:31:56,600 --> 00:32:00,200 Right. So. So it took us some time to, to figure it out. But yeah, 505 00:32:00,200 --> 00:32:03,640 there's a lot of interesting, you know, stuff that is happening when you try to 506 00:32:03,640 --> 00:32:07,000 connect non deterministic, you know, distributed microservices 507 00:32:07,160 --> 00:32:10,530 together have brains. That is an 508 00:32:10,530 --> 00:32:14,290 interesting problem. Goodness, that must have taken some time to run 509 00:32:14,290 --> 00:32:18,010 down. I can only imagine. Yeah, we actually want 510 00:32:18,010 --> 00:32:21,730 to create a cookbook on our website for all the people who 511 00:32:21,730 --> 00:32:25,250 try and connect different agents together just so they don't stumble 512 00:32:25,730 --> 00:32:28,050 and see the same problems that we did. 513 00:32:29,810 --> 00:32:32,690 And it's, it's still a very new field. So I think a lot of people 514 00:32:32,690 --> 00:32:36,050 haven't really figured out all the problem. We don't know what we don't know yet 515 00:32:36,320 --> 00:32:38,800 when it comes to deploying these things at scale. 516 00:32:40,720 --> 00:32:44,520 Yeah, yeah for sure. And we hope to hide all 517 00:32:44,520 --> 00:32:48,000 the complexities from people so it becomes very easy, very 518 00:32:48,000 --> 00:32:51,360 straightforward so you don't need to think about it. But 519 00:32:51,760 --> 00:32:54,800 along the way we do need to solve a lot of finer problems. 520 00:32:55,840 --> 00:32:59,560 Interesting. You mentioned standard. So there's a 2A, 521 00:32:59,560 --> 00:33:03,280 there's MCP and I think there's at least two more. I think IBM 522 00:33:03,280 --> 00:33:07,080 has one, Google has one. Anthropic had MCP. 523 00:33:08,280 --> 00:33:11,920 Do you think that we're looking at another standards war or do you think 524 00:33:11,920 --> 00:33:15,560 the big companies will or the community in whole will have a way 525 00:33:15,560 --> 00:33:19,080 will not allow that to happen. 526 00:33:20,600 --> 00:33:24,320 Yeah, overall I think there's like about 15 to 527 00:33:24,320 --> 00:33:27,960 16 protocols for Agent to agent communication 528 00:33:28,200 --> 00:33:31,600 and things like the like that. And around it 529 00:33:32,480 --> 00:33:36,240 the most famous protocols are two that most 530 00:33:36,240 --> 00:33:39,600 of the people have heard about it. McP is number one obviously 531 00:33:40,400 --> 00:33:43,944 but it's between agents and systems and 532 00:33:44,056 --> 00:33:47,520 A2A is by Google. But the most 533 00:33:47,520 --> 00:33:50,800 widely used protocol in production is actually acp, 534 00:33:51,840 --> 00:33:55,040 which is mostly protocol that is used for 535 00:33:55,440 --> 00:33:58,630 connection between a front end like IDE 536 00:33:59,270 --> 00:34:02,950 and an agent. This one is being used all over the 537 00:34:02,950 --> 00:34:06,630 place. So in terms of other approaches, Cisco, they have 538 00:34:06,630 --> 00:34:10,350 an agency consortium and they look not only at the protocol, they 539 00:34:10,350 --> 00:34:13,990 also look into the underlying infrastructure and they approach 540 00:34:13,990 --> 00:34:17,790 it more like from the networking perspective. I think 541 00:34:17,790 --> 00:34:21,470 at some point in time, obviously all this stuff 542 00:34:21,470 --> 00:34:25,280 will have standards, but even if it will have 543 00:34:25,280 --> 00:34:28,720 standards, it will be different and depending on your 544 00:34:29,040 --> 00:34:32,880 implementation quality, the product of yours will be better 545 00:34:32,880 --> 00:34:36,360 or much better. And the example, this 546 00:34:36,360 --> 00:34:38,080 example I can give is basically video 547 00:34:40,080 --> 00:34:43,680 streaming packets over udp. Right. 548 00:34:43,680 --> 00:34:47,320 We know how to do it for a long, 549 00:34:47,320 --> 00:34:50,930 long time. This is nothing groundbreaking and it's 550 00:34:50,930 --> 00:34:54,490 2026 and we have, you know, teams, we have 551 00:34:54,490 --> 00:34:58,250 Zoom, we have Google Meet and so on and so forth. And 552 00:34:58,250 --> 00:35:02,090 I'm pretty sure that all those products, they do not have 553 00:35:02,090 --> 00:35:05,689 the same standard of the video quality, audio quality and so on, despite 554 00:35:05,930 --> 00:35:09,050 the fact that they have been built on top of the 555 00:35:09,450 --> 00:35:13,290 one protocol, one infrastructure. So it's not 556 00:35:13,290 --> 00:35:16,710 just a protocol, it's a lot of knowledge 557 00:35:16,790 --> 00:35:20,470 and IP that you need to put in. So the whole system 558 00:35:20,470 --> 00:35:24,310 actually will work. So even 559 00:35:24,310 --> 00:35:28,110 if there is one protocol, there is a lot of work to do. And right 560 00:35:28,110 --> 00:35:31,830 now we are not yet in that place. Right. We are still in the discovery 561 00:35:31,830 --> 00:35:35,590 phase where a bunch of protocols pop up and 562 00:35:35,990 --> 00:35:39,190 have to be supported and reviewed and then they will die and so on. 563 00:35:39,910 --> 00:35:43,710 Yeah, and we've seen this before. You mentioned a great example, 564 00:35:43,710 --> 00:35:47,360 the different video and audio protocols for streaming conversations 565 00:35:47,360 --> 00:35:50,320 just like this meeting that we're recording on teams. 566 00:35:50,960 --> 00:35:54,600 I've seen it in my. The first thought I had was with the 567 00:35:54,600 --> 00:35:58,200 SQL protocol, SQL standards that are out there, the different 568 00:35:58,200 --> 00:36:01,800 Nancy numbered standards, and you nailed 569 00:36:01,800 --> 00:36:05,400 it. The different companies implement the standards 570 00:36:05,400 --> 00:36:09,040 differently and some perform better in this use case, some perform 571 00:36:09,040 --> 00:36:12,890 better in that one. I imagine with the velocity 572 00:36:13,050 --> 00:36:16,890 of agentic AI, how fast it's improving, it's moving faster 573 00:36:16,890 --> 00:36:20,610 than anything I've seen in my career slash 574 00:36:20,610 --> 00:36:24,450 hobby, which goes back 51 years now. I started as 575 00:36:24,450 --> 00:36:27,610 a wee lad back in 1975 576 00:36:28,330 --> 00:36:31,890 and this is moving faster than anything. I would imagine 577 00:36:31,890 --> 00:36:35,730 that probably in, you know, in short order I'll. I 578 00:36:35,730 --> 00:36:39,240 won't put a timestamp on it other than that will probably 579 00:36:39,400 --> 00:36:43,240 find something developed with the help of agents that 580 00:36:43,240 --> 00:36:47,080 will translate between the. You mentioned 15 or so 581 00:36:47,080 --> 00:36:50,880 protocols that are out there. Yeah, yeah, you're 582 00:36:50,880 --> 00:36:51,720 Right. And 583 00:36:54,760 --> 00:36:58,440 let me surprise you a bit. Okay. It's already there. We 584 00:36:58,440 --> 00:37:01,960 do it. Okay. So we support 585 00:37:02,200 --> 00:37:05,950 A2A and ACP, right. So you can have an A2A 586 00:37:05,950 --> 00:37:09,230 agent connecting to us and talking to agent via an acp. 587 00:37:09,550 --> 00:37:11,230 And we also. I'm not surprised. 588 00:37:13,470 --> 00:37:16,390 Yeah. Look, this is one of the first things that you need to solve, right? 589 00:37:16,390 --> 00:37:20,069 If you want to be. To enable interaction, right. You need 590 00:37:20,069 --> 00:37:23,830 to have a universal translator. Right. Otherwise people will 591 00:37:23,830 --> 00:37:26,830 speak different languages and no one will understand each other. 592 00:37:27,630 --> 00:37:31,120 So we act as a normalization layer. Another way 593 00:37:31,520 --> 00:37:35,240 how to look at it is that if you have a Cisco router, it doesn't 594 00:37:35,240 --> 00:37:38,960 really care if you're sending HTTP, TCP or video, audio 595 00:37:39,280 --> 00:37:42,880 whatnot. It's agnostic. And the same is for us. 596 00:37:43,120 --> 00:37:46,880 We want to be friends with every agentic development 597 00:37:46,880 --> 00:37:50,680 framework. We have no preference. We want to enable 598 00:37:50,680 --> 00:37:54,520 them to collaborate together. We want Langgraph 599 00:37:54,520 --> 00:37:58,220 and Crewai to be friends. We want the Codex 600 00:37:58,220 --> 00:38:01,900 and Claude to work together and not fight. So 601 00:38:01,900 --> 00:38:05,540 yeah, no, I think 602 00:38:05,540 --> 00:38:09,380 there's a lot to be said for kind of being the universal translator, bridging 603 00:38:09,380 --> 00:38:12,140 between all these worlds. Because at the end of the day 604 00:38:13,580 --> 00:38:17,100 protocol wars just delay everything. Right. And 605 00:38:17,340 --> 00:38:20,540 as long as you can kind of like get the job done, I think that's 606 00:38:20,540 --> 00:38:24,100 ultimately what matters more than what protocol was used. 607 00:38:24,180 --> 00:38:27,700 And you're right, like the router doesn't really care. Is it udp, is it tcp? 608 00:38:27,700 --> 00:38:31,460 Right. Just it's structured in such a way that it doesn't 609 00:38:31,460 --> 00:38:34,580 really matter. Interesting. 610 00:38:36,420 --> 00:38:40,100 Where do you think we're headed with this? 611 00:38:41,780 --> 00:38:43,060 Where do you think we're headed with 612 00:38:45,540 --> 00:38:48,900 this work in terms of agentic AI and 613 00:38:49,960 --> 00:38:53,760 these non deterministic microservices being put 614 00:38:53,760 --> 00:38:57,520 into the enterprise? What sort of unintended consequences? Because if you think 615 00:38:57,520 --> 00:39:00,960 about the enterprise, it as an ecosystem, we are 616 00:39:00,960 --> 00:39:04,680 introducing a new animal, right? We're introducing 617 00:39:04,680 --> 00:39:07,880 a new species, right. It's not human. It's not. It. It's not 618 00:39:07,880 --> 00:39:10,840 deterministic code. It's like this other thing that can kind of do. 619 00:39:12,600 --> 00:39:16,400 It can kind of come up with its own ideas. And some of those ideas 620 00:39:16,400 --> 00:39:18,480 are going to be great. Some of them are not going to be so great. 621 00:39:18,480 --> 00:39:21,970 Like what, what do you think that is going to change? 622 00:39:22,050 --> 00:39:25,170 Like where do you think that's going to go? 623 00:39:28,050 --> 00:39:30,850 Yeah, I know one thing for sure, 100% 624 00:39:31,730 --> 00:39:35,250 it's going to be exciting. Okay. At least I know 625 00:39:35,890 --> 00:39:39,730 yes, this I know for sure. The rest I can only guess. 626 00:39:40,450 --> 00:39:44,210 I think it's going to be a huge enable for A lot of people 627 00:39:44,290 --> 00:39:47,300 for a lot of enterprises. It will allow 628 00:39:48,020 --> 00:39:51,860 a lot of creativity work to be done. It for sure 629 00:39:52,260 --> 00:39:56,020 will require some re education from 630 00:39:56,020 --> 00:39:59,860 people, not because they will be fired and will be 631 00:39:59,860 --> 00:40:03,700 on universal basic income. And EI will do all the work just because 632 00:40:03,700 --> 00:40:06,340 it's a new entity and you need to learn how to work with it. 633 00:40:07,300 --> 00:40:10,940 Think of it in the following way. Whenever you are young 634 00:40:10,940 --> 00:40:14,350 and you come into a company, the first thing that you do in your first 635 00:40:14,350 --> 00:40:18,190 job is they do a training. They explain to you how to 636 00:40:18,190 --> 00:40:21,750 work as a team, how to collaborate with people, how to schedule 637 00:40:21,750 --> 00:40:25,470 meetings, how to write emails the right way. They 638 00:40:25,470 --> 00:40:28,950 explain to you that people working together as a force multiplier. 639 00:40:29,270 --> 00:40:32,990 They tell you how to talk, how to control your emotions and 640 00:40:32,990 --> 00:40:36,630 so on. So you are, you are basically being taught how to 641 00:40:36,710 --> 00:40:40,510 perform your work within an enterprise and operate with other human 642 00:40:40,510 --> 00:40:44,290 beings, even if there is a direct are conflicts. And it's 643 00:40:44,290 --> 00:40:48,090 despite the fact that you are a human being and you work with other humans 644 00:40:48,170 --> 00:40:52,010 and you've been around other humans for a long, long time. And here 645 00:40:52,010 --> 00:40:54,970 we have another entity. Okay, it's not human, 646 00:40:56,330 --> 00:41:00,170 but it's not a deterministic service. It has pluses, it has minuses, 647 00:41:00,330 --> 00:41:04,170 it sort of communicates with you in a language. It has different 648 00:41:04,170 --> 00:41:08,020 capabilities and you need to be trained how to use it, how to 649 00:41:08,020 --> 00:41:11,620 use it properly, how to depend on it and when you can depend on it, 650 00:41:11,940 --> 00:41:15,660 when not to depend on it. A lot of things that 651 00:41:15,660 --> 00:41:19,060 you know, for instance, I've seen quite a lot that people 652 00:41:19,940 --> 00:41:23,460 turn off their brains and they just outsource their own work 653 00:41:24,340 --> 00:41:28,100 to agents. And this has huge impact not only on the 654 00:41:28,180 --> 00:41:31,940 person itself, but also on the people, the peers 655 00:41:32,340 --> 00:41:36,090 within the organization. So people will have to be trained and 656 00:41:36,090 --> 00:41:39,850 they have to learn how to work alongside agents, how to 657 00:41:39,850 --> 00:41:43,490 control agents, just like you learned how to manage a 658 00:41:43,490 --> 00:41:47,090 team of people with different abilities, different 659 00:41:47,090 --> 00:41:50,650 specializations. People will have to learn how to manage a team of 660 00:41:50,650 --> 00:41:53,410 agents, or how to manage a team of agents and people, 661 00:41:54,370 --> 00:41:57,410 or how to manage people who have agentic teams. 662 00:41:58,130 --> 00:42:01,550 And it's going to be a learning experience for everyone, I believe. I don't think 663 00:42:01,870 --> 00:42:05,550 anyone knows exactly how it's going to play out. No, that's a great 664 00:42:05,550 --> 00:42:09,070 answer because I think anyone that already has got it all figured out is 665 00:42:09,310 --> 00:42:10,990 delusional or lying. Right? 666 00:42:13,310 --> 00:42:16,670 They're lying to themselves or they're lying to everyone else. And I think it's, it's. 667 00:42:17,310 --> 00:42:20,910 You're right, it is an exciting time. It's only Going to get more exciting 668 00:42:21,390 --> 00:42:25,190 and exciting has got good aspects and it certainly has bad aspects 669 00:42:25,190 --> 00:42:28,960 too. So where can folks find out more 670 00:42:28,960 --> 00:42:32,800 about band and how. Well, how about I set this up? What is 671 00:42:32,800 --> 00:42:36,640 band's answer to this would 672 00:42:36,640 --> 00:42:39,400 be a good place to start. And where can folks find out more about band? 673 00:42:41,560 --> 00:42:45,120 So the best place to find more about band is on 674 00:42:45,120 --> 00:42:48,760 Band AI. And then you can go to Docs 675 00:42:48,760 --> 00:42:52,360 Docs Band AI. And please 676 00:42:52,680 --> 00:42:56,060 use your Cloth or Courser or Codex 677 00:42:56,300 --> 00:42:59,740 and point it to the documentation and use it as your best friend. 678 00:43:00,060 --> 00:43:03,820 Because cloud and codecs, they can understand pretty well all 679 00:43:03,820 --> 00:43:07,500 the concepts that we have and then they can help you 680 00:43:07,500 --> 00:43:11,140 set up an agent and connect to 681 00:43:11,140 --> 00:43:14,620 it. If you have personal assistant, Hermes, 682 00:43:14,780 --> 00:43:17,660 openclaw, nanoclaw and such, 683 00:43:18,220 --> 00:43:21,810 Bent is a super great place for your agents 684 00:43:21,810 --> 00:43:25,610 to find friends. Okay. Like open Cloth your friend 685 00:43:25,930 --> 00:43:29,490 so you can schedule meetings with it or if you want to use the 686 00:43:29,490 --> 00:43:32,970 context of your friend who is a fan of, I don't know, 687 00:43:33,210 --> 00:43:35,610 science fiction. Right. And maintains 688 00:43:36,810 --> 00:43:40,490 a library of comics and PDF files or 689 00:43:40,490 --> 00:43:43,890 whatever. Right. And share the information. Yeah. So 690 00:43:43,890 --> 00:43:47,570 onboarding is pretty easy. And make sure to spread the 691 00:43:47,570 --> 00:43:50,730 word because we are collaboration and interaction 692 00:43:51,050 --> 00:43:53,850 framework. Right. You need two for Tango. So bring a friend. 693 00:43:54,890 --> 00:43:58,650 All right. Well, that's cool. Any. Any other 694 00:43:58,650 --> 00:43:59,610 questions, Andy? 695 00:44:02,730 --> 00:44:06,450 Trying to find the mute button. No. I'm already messaging people on 696 00:44:06,450 --> 00:44:08,090 my team saying, go check this out. 697 00:44:10,330 --> 00:44:13,850 Drama. And Mouse wasn't over the screen. I went to Docs Band. 698 00:44:15,650 --> 00:44:19,330 I love that line, Vlad, that it's an excellent place 699 00:44:19,330 --> 00:44:22,770 for your agents to find friends. That's a good 700 00:44:22,770 --> 00:44:26,130 tagline right there. And we're gonna be checking it out. 701 00:44:26,930 --> 00:44:30,009 Yeah, definitely check it out. I have my own instance of openclaw and I'll be 702 00:44:30,009 --> 00:44:33,690 setting up a Hermes, I think is how you say it. It's 703 00:44:33,690 --> 00:44:36,850 Hermes if it's the designer bag and Hermes if it's the. 704 00:44:37,490 --> 00:44:41,250 I think. But definitely be. Definitely checking it out. I'm 705 00:44:41,250 --> 00:44:44,920 in the. I'm starting. I'm in the early phase of an agentic project for 706 00:44:44,920 --> 00:44:48,560 work, so that'll be interesting to kind of check this out. And 707 00:44:49,360 --> 00:44:52,120 thanks. We could talk for another hour, but we want to be respectful of your 708 00:44:52,120 --> 00:44:55,040 time. And thanks for coming on, Vlad. And 709 00:44:56,080 --> 00:44:58,990 we'll let the outro music play. Thank you.