WEBVTT NOTE Sentence-level transcript of https://www.youtube.com/watch?v=toq-jyGLZDk NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/toq-jyGLZDk.html. A cue ends where the next begins, or 2 s after its last word. s1 00:00:01.309 --> 00:00:03.309 [music] s2 00:00:13.200 --> 00:00:14.200 Okay. s3 00:00:14.200 --> 00:00:14.680 Hi everyone. s4 00:00:14.680 --> 00:00:16.200 Can you hear me? s5 00:00:16.200 --> 00:00:17.160 A lot of people, yeah. s6 00:00:17.160 --> 00:00:19.600 So, please be quiet. s7 00:00:20.240 --> 00:00:21.800 So, my name is Vlad. s8 00:00:21.800 --> 00:00:24.040 I'm co-founder and CTO at Bend. s9 00:00:24.040 --> 00:00:27.320 And today I have four topics I want to discuss with you. s10 00:00:27.320 --> 00:00:29.320 First, thesis. s11 00:00:29.320 --> 00:00:32.960 Thesis of our company, what we believe believe in. s12 00:00:32.960 --> 00:00:37.320 AI evolution from adversarial agents to loop engineering and beyond. s13 00:00:37.320 --> 00:00:43.320 Technical challenges that we experience right now that should be solved right now. s14 00:00:43.320 --> 00:00:48.560 And obviously Bend to present the company and the product and what we do. s15 00:00:48.560 --> 00:00:50.680 Let's start with the thesis. s16 00:00:50.680 --> 00:01:02.120 First of all, we believe that the future belongs to AI to AI communication within a business, between businesses, and between consumers and businesses. s17 00:01:02.120 --> 00:01:10.600 AI will be doing work on our behalf, and they will have to talk to each other in order to solve tasks on our behalf. s18 00:01:10.600 --> 00:01:18.280 Second, agents will be autonomous, always-on entities spread across the globe. s19 00:01:18.280 --> 00:01:30.840 Now, these are very nice words, but I want to spend a minute just to paint a picture in your head how this type of communication between agents actually is going to look like. s20 00:01:32.200 --> 00:01:34.520 Demo time. s21 00:01:34.800 --> 00:01:38.240 Agents will communicate in a conversational space. s22 00:01:38.240 --> 00:01:40.400 They will see each other. s23 00:01:40.400 --> 00:01:42.200 They will be distributed. s24 00:01:42.200 --> 00:01:45.520 They will receive tasks from us or from any system. s25 00:01:45.520 --> 00:01:57.040 They will be able to look up peers in the registries, invite these peers in into conversational space, delegate the tasks, gather enough information, and and back to us. s26 00:01:57.040 --> 00:02:02.600 This is how the future of AI to AI communication is going to look like. s27 00:02:02.600 --> 00:02:07.080 So, while we are talking hold this in your head. s28 00:02:07.479 --> 00:02:11.600 Let's talk a bit about adversarial agents. s29 00:02:11.600 --> 00:02:18.240 I think everyone here knows this concept and I'm pretty sure everyone here is using it as well. s30 00:02:18.240 --> 00:02:19.680 Where are we using it? s31 00:02:19.680 --> 00:02:26.640 If you are using Claude or Codex, I'm pretty sure you're running multiple sessions working on the same task. s32 00:02:26.640 --> 00:02:28.600 One is planning, another is reviewing. s33 00:02:28.600 --> 00:02:32.200 One is writing code, another is reviewing code. s34 00:02:32.200 --> 00:02:46.320 So, basically what is happening, you have two stateful agents running, working on the same task, and you are basically a Cisco router and a switch moving packets between these two stateful agents. s35 00:02:47.480 --> 00:02:50.080 Let's talk about loop engineering. s36 00:02:50.080 --> 00:02:57.080 A number of talks on this conference about loop engineering, a number of articles, uh pretty new concept, right? s37 00:02:57.080 --> 00:03:00.959 But I would like to distill it uh to its essence. s38 00:03:00.959 --> 00:03:03.680 So, what is loop engineering? s39 00:03:03.720 --> 00:03:06.560 Basically, it's the same as the previous concept. s40 00:03:06.560 --> 00:03:22.519 It's you running multiple agents, hopefully stateful agents and not stateless, but instead of you being a router or a switch, you outsource this to some piece of Python or TypeScript that someone wrote as an open-source project, s41 00:03:22.519 --> 00:03:30.600 and um this piece of code basically prompts the stateful agents back and forth, and you trigger it. s42 00:03:30.920 --> 00:03:35.000 Now, why are we using these multiple agents when we do this work? s43 00:03:35.000 --> 00:03:44.720 Because of the underlying architectural limitations of a transformer, specifically a concept is called single agent bottleneck, right? s44 00:03:44.720 --> 00:03:49.400 So, why we are not running one agent in one session that is doing everything? s45 00:03:49.400 --> 00:03:52.160 For instance, confirmation bias, right? s46 00:03:52.160 --> 00:03:56.320 Attention dilution, context fragmentation, recall degradation. s47 00:03:56.320 --> 00:04:03.880 So, 1 million tokens or 2 million tokens context will not help you in terms of the performance. s48 00:04:04.480 --> 00:04:06.560 But, let's talk about multi-agent systems, right? s49 00:04:06.560 --> 00:04:13.760 So, we already understand that us running multiple sessions, these are basically multi-agent systems running. s50 00:04:13.760 --> 00:04:19.799 So, what is the easiest way to connect multiple agents running together? s51 00:04:20.079 --> 00:04:21.320 Messaging platforms, right? s52 00:04:21.320 --> 00:04:23.200 We already have Slack, right? s53 00:04:23.200 --> 00:04:30.120 Anthropic released Slack agent, we have Teams, we have Discord, we have WhatsApp, we have Telegram, and so on, right? s54 00:04:30.120 --> 00:04:32.640 So, this is a solution. s55 00:04:32.640 --> 00:04:38.840 Well, in order for you to connect an agent to a Telegram, five steps. s56 00:04:38.840 --> 00:04:43.160 In order for you to connect to Discord, seven steps. s57 00:04:43.440 --> 00:04:48.600 In order for you to connect your agent to Slack, eight steps. s58 00:04:48.760 --> 00:04:51.280 WhatsApp, 11. s59 00:04:51.280 --> 00:04:53.680 And these are not simple steps. s60 00:04:53.680 --> 00:04:59.720 They are all manual, and you need to read a book in order to be able to set it up. s61 00:04:59.720 --> 00:05:04.960 And even if you do, you get only one thing and one thing only. s62 00:05:05.160 --> 00:05:07.080 What is this thing? s63 00:05:07.080 --> 00:05:11.520 An agent talking to a person, which is usually you. s64 00:05:11.520 --> 00:05:18.440 This agent cannot talk to any other agent, this agent cannot talk to any other person without you jumping through hoops. s65 00:05:18.440 --> 00:05:28.120 So, basically, your agents are still alone in a kind of digital solitary confinement. s66 00:05:29.760 --> 00:05:33.919 But, smart people will say, "I know the solution. s67 00:05:33.919 --> 00:05:35.919 Protocols, right? s68 00:05:35.919 --> 00:05:44.120 MCP, ACP, A2A, and etc. This is the solution, right? s69 00:05:44.120 --> 00:05:54.320 Well, if you are a proponent of that kind of a for connection for connecting your sessions, this is how your multi-agent system is going to look like. s70 00:05:54.320 --> 00:06:00.320 Multiple servers calling other agents' tools through MCP. s71 00:06:00.320 --> 00:06:06.400 And if you want, maybe this agent that you called as a tool can call another agent as an A2A. s72 00:06:06.400 --> 00:06:07.800 Sounds super simple. s73 00:06:07.800 --> 00:06:09.920 Okay, everyone knows what A2A is. s74 00:06:09.920 --> 00:06:13.200 Google have a very huge marketing budget. s75 00:06:13.200 --> 00:06:16.640 But let's take the take a look at the complexities. s76 00:06:16.880 --> 00:06:19.760 MCP means stateless calls. s77 00:06:19.760 --> 00:06:26.280 It means you cannot go back to the agent that you talked to and ask it again what happened. s78 00:06:26.280 --> 00:06:31.280 A2A one way unless you implement both directions. s79 00:06:31.280 --> 00:06:35.520 So, agent A to sends a task to agent B. It's client-server. s80 00:06:35.520 --> 00:06:38.800 If you want an agent B to send a task to agent A, it's client-server. s81 00:06:38.800 --> 00:06:43.240 So, you need to implement on the same side both a client and a server. s82 00:06:43.240 --> 00:06:47.040 Chaining REST API calls, you will meet timeouts. s83 00:06:47.040 --> 00:06:58.840 And obviously thing a lot of things are missing like discovery, draft state of A2A, queues that you still need to implement and put behind all of that. s84 00:06:59.120 --> 00:07:02.440 And if you're doing this, you're basically building planning. s85 00:07:02.440 --> 00:07:04.280 You're not building a multi-agent system. s86 00:07:04.280 --> 00:07:09.080 You're wasting your time on a boring boring stuff. s87 00:07:09.280 --> 00:07:11.360 Now, let's summarize what we know. s88 00:07:11.360 --> 00:07:14.680 We know that multi-agent systems are already here. s89 00:07:14.680 --> 00:07:17.840 Everyone here is using multiple agents. s90 00:07:17.840 --> 00:07:21.520 We know that the messaging platforms are not a solution. s91 00:07:21.520 --> 00:07:28.160 They are built for humans and they actively block you from connecting a bot to a bot. s92 00:07:28.160 --> 00:07:38.520 And we know that protocol is too low-level of a technical abstraction to be able um to build anything on it at scale. s93 00:07:39.120 --> 00:07:45.840 But how hard can it be to build a system that can allow multiple agents to connect together. s94 00:07:45.840 --> 00:07:49.240 It's not that difficult, right? s95 00:07:49.240 --> 00:08:00.080 Well, you see, connecting remote agents being a two sessions as a process is running on your laptop or my cloud to your codex is a distributed systems problem. s96 00:08:00.080 --> 00:08:03.360 And distributed systems are not easy by definition. s97 00:08:03.360 --> 00:08:05.360 Forget agents, okay? s98 00:08:05.360 --> 00:08:08.000 Deterministic software, it's a pain in the butt. s99 00:08:08.000 --> 00:08:13.200 Add to this agents that are microservices and non-deterministic, okay? s100 00:08:13.200 --> 00:08:15.600 And you have a lot of fun. s101 00:08:15.600 --> 00:08:24.000 In order for this um to be solved, what you need to actually to address is a transport layer, right? s102 00:08:24.000 --> 00:08:25.880 The transport has to be real-time. s103 00:08:25.880 --> 00:08:30.040 It has to be ordered because LLMs expect ordered messages. s104 00:08:30.040 --> 00:08:32.479 You need to handle retries and so on. s105 00:08:32.479 --> 00:08:37.200 You need to handle persistency and hydration because your agent is a microservice. s106 00:08:37.200 --> 00:08:42.680 It can fail, the pod can crash, and then all your multi-agent system goes kaput. s107 00:08:42.680 --> 00:08:44.240 So, you need to handle this. s108 00:08:44.240 --> 00:08:48.040 You need to handle something very new, runtime binding. s109 00:08:48.040 --> 00:09:00.680 If you want to connect a LAN graph to a codex or cloud to a crew AI, you need to connect a thread ID to a session ID to execution ID to a run ID of this multiple agents by and map it up to this IDs together. s110 00:09:00.680 --> 00:09:07.240 Otherwise, agents will not be able to work as a group towards the same goal and the same task. s111 00:09:07.240 --> 00:09:16.960 And obviously, you cannot work at the level of IPs and ports, or URLs that you have to manage, or even pub/sub topics. s112 00:09:16.960 --> 00:09:32.200 You need to rise above all of the technical details to a different level of abstraction, conversation, participants, channels, or rooms, message routing that is tailored to agents as the first citizens. s113 00:09:32.200 --> 00:09:42.640 And of course, all of that is still not enough for an to use this kind of product or communication software because every organization wants governance. s114 00:09:42.640 --> 00:09:44.840 They need identity, they need observability. s115 00:09:44.840 --> 00:09:55.080 Observability not in a message that was sent from agent A on server A to agent B on server B. But also what tool calls, okay, happened after a certain agent received this message, s116 00:09:55.080 --> 00:09:55.560 right? s117 00:09:55.560 --> 00:10:00.760 And it's very difficult to provide this kind of observability across distributed systems. s118 00:10:00.760 --> 00:10:15.120 So, without all of that solved, you will not have agents talking to each other and this wonderful future of us, okay, everyone on a universal basic income lying on a beach will not happen. s119 00:10:17.520 --> 00:10:20.120 I would like to introduce Bend. s120 00:10:20.120 --> 00:10:24.160 We actually took all the stuff that we talked about and solved it. s121 00:10:24.160 --> 00:10:26.000 So, you don't have to solve it. s122 00:10:26.000 --> 00:10:32.880 So, you can super easily connect all your agents within an enterprise or your Hermes and Open Cloak. s123 00:10:33.400 --> 00:10:39.600 I would like to unveil a product that we are launching this week. s124 00:10:39.600 --> 00:10:52.600 And this is a global interaction and collaboration layer worldwide for any agents, any platform, any language, any environment. s125 00:10:52.600 --> 00:10:56.040 And it's not peer-to-peer, it's multi-peer. s126 00:10:56.040 --> 00:10:58.560 And we did not forget humans as well. s127 00:10:58.560 --> 00:11:01.200 Okay, if you are human, you can still connect and talk to your agent. s128 00:11:01.200 --> 00:11:03.760 This is completely fine. s129 00:11:03.760 --> 00:11:09.000 And if we open the hood quite a bit, what we see there is that it's not just communication layer. s130 00:11:09.000 --> 00:11:14.120 We have a registry, so your agents can see each other automatically without you jumping through hoops. s131 00:11:14.120 --> 00:11:17.320 Persistency, channels, message filtering. s132 00:11:17.320 --> 00:11:21.000 Instead of doing it on the client side, we do it within the platform. s133 00:11:21.000 --> 00:11:24.400 Security, observability, and so on. s134 00:11:24.400 --> 00:11:26.520 Now, the fun part. s135 00:11:26.520 --> 00:11:28.839 Okay, let me show you a few demos. s136 00:11:28.839 --> 00:11:32.040 Uh the time is short, so demos are recorded, but we have a booth. s137 00:11:32.040 --> 00:11:35.280 Please come and see it live or just go and connect. s138 00:11:35.280 --> 00:11:41.160 Um the demo will be two users in two different browsers connected to the communication platform. s139 00:11:41.160 --> 00:11:42.320 One is me, Vlad. s140 00:11:42.320 --> 00:11:45.520 I have a personal agent Claude running on a Mac. s141 00:11:45.520 --> 00:11:50.320 And another user is Mike, Claude code in a terminal session, okay? s142 00:11:50.320 --> 00:11:53.920 Uh Claude SDK and uh Landlord. s143 00:11:54.560 --> 00:11:56.520 Demo number one. s144 00:11:56.520 --> 00:12:00.160 On the top window, it's me, my account. s145 00:12:00.160 --> 00:12:02.080 And I have a personal assistant. s146 00:12:02.080 --> 00:12:09.240 Below, it's a different user connected to the same interaction layer with no agents at all. s147 00:12:09.240 --> 00:12:20.920 On the right, okay, I'm going to spin up two agents and they will onboard on the platform and connect and become agents that belong to Mike. s148 00:12:20.920 --> 00:12:26.000 And let's see how fast it's going to to be. s149 00:12:26.040 --> 00:12:31.040 So, first of all, we go and we uh spin up an agent. s150 00:12:31.040 --> 00:12:36.320 Seconds later, this agent card already appears with any within an account of Mike. s151 00:12:36.320 --> 00:12:38.560 We spin up a Landlord. s152 00:12:38.560 --> 00:12:41.480 Second later, this agent card appears as well. s153 00:12:41.480 --> 00:12:44.440 From this point on, these agents know of each other. s154 00:12:44.440 --> 00:12:47.400 They understand that they can interact with each other. s155 00:12:47.400 --> 00:12:49.200 But we want to do something cooler. s156 00:12:49.200 --> 00:13:00.720 We want to take a Codex agent and connect this agent to my personal assistant running on my Mac, so they can actually interact together. s157 00:13:00.720 --> 00:13:13.320 We are sending a contact request that requires a bilateral consent from both parties in order for this agent to be introduced inside each uh registry that belongs either to an agent or to a human. s158 00:13:13.320 --> 00:13:18.040 The request was sent and it was accepted by me. s159 00:13:18.040 --> 00:13:27.640 And right now we have a Codex agent that can invite my personal assistant into a conversational space and interact with it in real time. s160 00:13:27.640 --> 00:13:33.920 This is completely stateful agents can be deployed anywhere in the world and they are. s161 00:13:33.920 --> 00:13:38.440 And what you see here, this is real time. s162 00:13:39.480 --> 00:13:48.400 And you can see also that because my agent is talking to another agent of a different user, I have a full visibility into the conversation. s163 00:13:48.400 --> 00:14:01.240 So, no conversation can happen without me seeing it and obviously my agent can invite me into every conversational space and interact with me and as a human in the loop. s164 00:14:03.480 --> 00:14:08.840 So, what we have seen right now is a zero hassle on boarding of any agent. s165 00:14:08.840 --> 00:14:09.800 Try it at home. s166 00:14:09.800 --> 00:14:13.600 Tell me how it how how how it went. s167 00:14:13.640 --> 00:14:16.360 So, another example. s168 00:14:16.360 --> 00:14:18.560 We have a user Mike. s169 00:14:18.560 --> 00:14:29.000 And this user still has the Codex running connected uh to the platform and now we want to run a Claude code terminal session. s170 00:14:29.000 --> 00:14:37.360 And we want this terminal session from the terminal to connect to our platform and interact with other agents, right? s171 00:14:37.360 --> 00:14:49.200 It can be here specifically Claude code in a terminal of Mike talks to a a Codex SDK of Mike, but because it's a global platform, it doesn't really matter, okay? s172 00:14:49.200 --> 00:14:54.120 They can talk to your agent, they can talk to your your agent and so on. s173 00:14:54.120 --> 00:15:02.800 So, we are spinning up Claude and we tell Claude in a terminal session, please go and connect to our platform. s174 00:15:02.800 --> 00:15:03.920 Same concept. s175 00:15:03.920 --> 00:15:13.800 This specific Claude session gets an identity, an agent card and we hide all the complexities in terms of where it is from the network perspective. s176 00:15:13.800 --> 00:15:25.720 And from now this agent can actually create a conversational space as an owner, invite other agents in this specific case, Equidex, right? s177 00:15:25.720 --> 00:15:27.880 And it will invite Mike as well. s178 00:15:27.880 --> 00:15:35.520 So, it's uh you know, happy family, and they can start working on some engineering tasks. s179 00:15:35.520 --> 00:15:37.600 Whatever it is. s180 00:15:37.600 --> 00:15:44.400 The other way that people call it, whatever you see here, is loop engineering. s181 00:15:44.400 --> 00:15:45.400 Okay? s182 00:15:45.400 --> 00:15:58.120 But, it's loop engineering without 500,000 lines of a price book Python or Python script code that you need to bring over so your agents can actually prompt each other for work. s183 00:15:58.120 --> 00:16:07.480 These agents will now right now start to and create a small website, uh and they will interact with each other, and one will review work of the other. s184 00:16:07.480 --> 00:16:15.640 Uh and obviously, if they struggle and Mike struggles, uh these agents can invite me into the same conversation space. s185 00:16:15.640 --> 00:16:18.400 I can help them, assist them. s186 00:16:18.400 --> 00:16:22.680 Um and it's pretty easy. s187 00:16:23.680 --> 00:16:25.000 That's it. s188 00:16:25.000 --> 00:16:25.640 Thank you very much. s189 00:16:25.640 --> 00:16:27.360 I'm right on time. s190 00:16:27.360 --> 00:16:29.920 Come to our booth, LG17. s191 00:16:29.920 --> 00:16:31.920 Take screenshots, take QR codes. s192 00:16:31.920 --> 00:16:41.720 One is for the platform to connect your agents, and one is for the future of the loop engineering that doesn't need not require anything. s193 00:16:41.720 --> 00:16:47.600 And we have Works that actually came here because they like orchestration, they do not like collaboration. s194 00:16:47.600 --> 00:16:49.800 They have stolen all the swag. s195 00:16:49.800 --> 00:16:54.760 So, if you want a swag, please come to Works, get a sword, and help us fight them. s196 00:16:54.760 --> 00:16:56.960 Thank you.