Why Your AI Agents Can't Talk to Each Other (Yet) — Vlad Luzin, BAND

AI Engineer · 17 min · 196 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

00:01[music]
00:13Okay.
00:14Hi everyone.
00:14Can you hear me?
00:16A lot of people, yeah.
00:17So, please be quiet.
00:20So, my name is Vlad.
00:21I'm co-founder and CTO at Bend.
00:24And today I have four topics I want to discuss with you.
00:27First, thesis.
00:29Thesis of our company, what we believe believe in.
00:32AI evolution from adversarial agents to loop engineering and beyond.
00:37Technical challenges that we experience right now that should be solved right now.
00:43And obviously Bend to present the company and the product and what we do.
00:48Let's start with the thesis.
00:50First of all, we believe that the future belongs to AI to AI communication within a business, between businesses, and between consumers and businesses.
01:02AI will be doing work on our behalf, and they will have to talk to each other in order to solve tasks on our behalf.
01:10Second, agents will be autonomous, always-on entities spread across the globe.
01:18Now, 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.
01:32Demo time.
01:34Agents will communicate in a conversational space.
01:38They will see each other.
01:40They will be distributed.
01:42They will receive tasks from us or from any system.
01:45They 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.
01:57This is how the future of AI to AI communication is going to look like.
02:02So, while we are talking hold this in your head.
02:07Let's talk a bit about adversarial agents.
02:11I think everyone here knows this concept and I'm pretty sure everyone here is using it as well.
02:18Where are we using it?
02:19If you are using Claude or Codex, I'm pretty sure you're running multiple sessions working on the same task.
02:26One is planning, another is reviewing.
02:28One is writing code, another is reviewing code.
02:32So, 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.
02:47Let's talk about loop engineering.
02:50A number of talks on this conference about loop engineering, a number of articles, uh pretty new concept, right?
02:57But I would like to distill it uh to its essence.
03:00So, what is loop engineering?
03:03Basically, it's the same as the previous concept.
03:06It'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,
03:22and um this piece of code basically prompts the stateful agents back and forth, and you trigger it.
03:30Now, why are we using these multiple agents when we do this work?
03:35Because of the underlying architectural limitations of a transformer, specifically a concept is called single agent bottleneck, right?
03:44So, why we are not running one agent in one session that is doing everything?
03:49For instance, confirmation bias, right?
03:52Attention dilution, context fragmentation, recall degradation.
03:56So, 1 million tokens or 2 million tokens context will not help you in terms of the performance.
04:04But, let's talk about multi-agent systems, right?
04:06So, we already understand that us running multiple sessions, these are basically multi-agent systems running.
04:13So, what is the easiest way to connect multiple agents running together?
04:20Messaging platforms, right?
04:21We already have Slack, right?
04:23Anthropic released Slack agent, we have Teams, we have Discord, we have WhatsApp, we have Telegram, and so on, right?
04:30So, this is a solution.
04:32Well, in order for you to connect an agent to a Telegram, five steps.
04:38In order for you to connect to Discord, seven steps.
04:43In order for you to connect your agent to Slack, eight steps.
04:48WhatsApp, 11.
04:51And these are not simple steps.
04:53They are all manual, and you need to read a book in order to be able to set it up.
04:59And even if you do, you get only one thing and one thing only.
05:05What is this thing?
05:07An agent talking to a person, which is usually you.
05:11This agent cannot talk to any other agent, this agent cannot talk to any other person without you jumping through hoops.
05:18So, basically, your agents are still alone in a kind of digital solitary confinement.
05:29But, smart people will say, "I know the solution.
05:33Protocols, right?
05:35MCP, ACP, A2A, and etc. This is the solution, right?
05:44Well, 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.
05:54Multiple servers calling other agents' tools through MCP.
06:00And if you want, maybe this agent that you called as a tool can call another agent as an A2A.
06:06Sounds super simple.
06:07Okay, everyone knows what A2A is.
06:09Google have a very huge marketing budget.
06:13But let's take the take a look at the complexities.
06:16MCP means stateless calls.
06:19It means you cannot go back to the agent that you talked to and ask it again what happened.
06:26A2A one way unless you implement both directions.
06:31So, agent A to sends a task to agent B. It's client-server.
06:35If you want an agent B to send a task to agent A, it's client-server.
06:38So, you need to implement on the same side both a client and a server.
06:43Chaining REST API calls, you will meet timeouts.
06:47And 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.
06:59And if you're doing this, you're basically building planning.
07:02You're not building a multi-agent system.
07:04You're wasting your time on a boring boring stuff.
07:09Now, let's summarize what we know.
07:11We know that multi-agent systems are already here.
07:14Everyone here is using multiple agents.
07:17We know that the messaging platforms are not a solution.
07:21They are built for humans and they actively block you from connecting a bot to a bot.
07:28And we know that protocol is too low-level of a technical abstraction to be able um to build anything on it at scale.
07:39But how hard can it be to build a system that can allow multiple agents to connect together.
07:45It's not that difficult, right?
07:49Well, 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.
08:00And distributed systems are not easy by definition.
08:03Forget agents, okay?
08:05Deterministic software, it's a pain in the butt.
08:08Add to this agents that are microservices and non-deterministic, okay?
08:13And you have a lot of fun.
08:15In order for this um to be solved, what you need to actually to address is a transport layer, right?
08:24The transport has to be real-time.
08:25It has to be ordered because LLMs expect ordered messages.
08:30You need to handle retries and so on.
08:32You need to handle persistency and hydration because your agent is a microservice.
08:37It can fail, the pod can crash, and then all your multi-agent system goes kaput.
08:42So, you need to handle this.
08:44You need to handle something very new, runtime binding.
08:48If 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.
09:00Otherwise, agents will not be able to work as a group towards the same goal and the same task.
09:07And obviously, you cannot work at the level of IPs and ports, or URLs that you have to manage, or even pub/sub topics.
09:16You 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.
09:32And 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.
09:42They need identity, they need observability.
09:44Observability 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,
09:55right?
09:55And it's very difficult to provide this kind of observability across distributed systems.
10:00So, 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.
10:17I would like to introduce Bend.
10:20We actually took all the stuff that we talked about and solved it.
10:24So, you don't have to solve it.
10:26So, you can super easily connect all your agents within an enterprise or your Hermes and Open Cloak.
10:33I would like to unveil a product that we are launching this week.
10:39And this is a global interaction and collaboration layer worldwide for any agents, any platform, any language, any environment.
10:52And it's not peer-to-peer, it's multi-peer.
10:56And we did not forget humans as well.
10:58Okay, if you are human, you can still connect and talk to your agent.
11:01This is completely fine.
11:03And if we open the hood quite a bit, what we see there is that it's not just communication layer.
11:09We have a registry, so your agents can see each other automatically without you jumping through hoops.
11:14Persistency, channels, message filtering.
11:17Instead of doing it on the client side, we do it within the platform.
11:21Security, observability, and so on.
11:24Now, the fun part.
11:26Okay, let me show you a few demos.
11:28Uh the time is short, so demos are recorded, but we have a booth.
11:32Please come and see it live or just go and connect.
11:35Um the demo will be two users in two different browsers connected to the communication platform.
11:41One is me, Vlad.
11:42I have a personal agent Claude running on a Mac.
11:45And another user is Mike, Claude code in a terminal session, okay?
11:50Uh Claude SDK and uh Landlord.
11:54Demo number one.
11:56On the top window, it's me, my account.
12:00And I have a personal assistant.
12:02Below, it's a different user connected to the same interaction layer with no agents at all.
12:09On 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.
12:20And let's see how fast it's going to to be.
12:26So, first of all, we go and we uh spin up an agent.
12:31Seconds later, this agent card already appears with any within an account of Mike.
12:36We spin up a Landlord.
12:38Second later, this agent card appears as well.
12:41From this point on, these agents know of each other.
12:44They understand that they can interact with each other.
12:47But we want to do something cooler.
12:49We want to take a Codex agent and connect this agent to my personal assistant running on my Mac, so they can actually interact together.
13:00We 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.
13:13The request was sent and it was accepted by me.
13:18And right now we have a Codex agent that can invite my personal assistant into a conversational space and interact with it in real time.
13:27This is completely stateful agents can be deployed anywhere in the world and they are.
13:33And what you see here, this is real time.
13:39And 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.
13:48So, 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.
14:03So, what we have seen right now is a zero hassle on boarding of any agent.
14:08Try it at home.
14:09Tell me how it how how how it went.
14:13So, another example.
14:16We have a user Mike.
14:18And this user still has the Codex running connected uh to the platform and now we want to run a Claude code terminal session.
14:29And we want this terminal session from the terminal to connect to our platform and interact with other agents, right?
14:37It 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?
14:49They can talk to your agent, they can talk to your your agent and so on.
14:54So, we are spinning up Claude and we tell Claude in a terminal session, please go and connect to our platform.
15:02Same concept.
15:03This 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.
15:13And from now this agent can actually create a conversational space as an owner, invite other agents in this specific case, Equidex, right?
15:25And it will invite Mike as well.
15:27So, it's uh you know, happy family, and they can start working on some engineering tasks.
15:35Whatever it is.
15:37The other way that people call it, whatever you see here, is loop engineering.
15:44Okay?
15:45But, 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.
15:58These 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.
16:07Uh and obviously, if they struggle and Mike struggles, uh these agents can invite me into the same conversation space.
16:15I can help them, assist them.
16:18Um and it's pretty easy.
16:23That's it.
16:25Thank you very much.
16:25I'm right on time.
16:27Come to our booth, LG17.
16:29Take screenshots, take QR codes.
16:31One 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.
16:41And we have Works that actually came here because they like orchestration, they do not like collaboration.
16:47They have stolen all the swag.
16:49So, if you want a swag, please come to Works, get a sword, and help us fight them.
16:54Thank you.