Your Agents Are in Solitary Confinement: Why MCP & A2A Aren't Enough — Vlad Luzin, Band

AI Engineer · 17 min · 208 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]
00:13Hi all.
00:13My name is Vlad.
00:14I'm co-founder and the CTO at bank.
00:17Now, before we start let's do a bit of trivia.
00:22Orc C okay.
00:25If C, raise your hand.
00:28K?
00:30Who believes both are correct?
00:33Okay, just a couple.
00:34Both are correct.
00:35One is from the Lord of the Rings, another is from Warhammer game Game Workshop games.
00:44I have in my agenda four topics to cover.
00:47Topic number one, I would like to tell you about our thesis as a company, what we believe in.
00:52Then we'll talk about the AI evolution from adversarial agents to loop engineering and beyond.
00:58Then we'll touch base on the technical challenges of tomorrow that we have to solve today.
01:04And then I would like to present the company, what we do, and what problems we solve.
01:10So, our thesis is future belongs to AI to AI communication within a business, between businesses, and between consumers and businesses.
01:21Agents will be everywhere.
01:22They will do work on our behalf, and they will have to communicate between each other.
01:26They will be written in different frameworks, different languages, and deployed in different environments.
01:31And these agents will be fully autonomous.
01:34And they will communicate with each other without human intervention.
01:39Now, just to paint a picture before I start talking, uh how this type of communication between the agents is going to look like.
01:48So, what you see here is a conversational space that agents can create themselves, so they can be added.
01:53They will receive a task either from a human or from another system.
01:57And they will be able to discover other agents add them as a participant in this conversation space, have back and forth communication, solve the task, and report back to humans.
02:09This is how the future that you all read in the newspapers will look like.
02:15And when people hear this, they think either that this is too far in the future or that I'm crazy.
02:23Okay, that's it.
02:25And here is an example of a conversation I had with a CTO of $10 billion company who said that before thinking about technical solutions to hypothetical problems like multi-agent coordination,
02:37I try to keep things simple and avoid the problem.
02:42Now, let's unpack.
02:43Is multi-agent coordination a hypothetical problem?
02:47Is it possible to keep things simple?
02:51And is it possible to avoid the problem?
02:56Now, let's talk a bit about the adversarial agents as a as a concept.
03:01Um we all use this.
03:04I'm pretty sure everyone here is a developer.
03:06And as a developer, you have two cloud or coding sessions.
03:09One is doing the work, another is reviewing the work.
03:12And you probably have not only two sessions, you probably have multiple tabs and multiple sessions working working at the same time on multiple problems.
03:23You are basically acting as a router, a Cisco router or a switch between two stateful agents that do work on your behalf, but they have no ability to communicate, hence you need to prompt them.
03:37Let's talk about loop engineering.
03:40You've heard the the guy before talked about Peter and Boris, right?
03:44So, what they say to you?
03:45They basically say to you, "Stop being the router between your agents and let your agents prompt each other.
03:54Now, what they're basically saying to you is instead of being the router and copy-pasting stuff, start fighting with Python and TypeScript libraries and different abstraction layers that will invent
04:06how agents should prompt each other.
04:11We also have protocols.
04:13A2A, MCP, ACP, and then other protocols crypto-related.
04:19Simple.
04:20Solves all the problems, right?
04:22So, this is how your multi-agent system looks like probably in production.
04:27MCP, maybe if you're super advanced, it's A2A.
04:31And sounds simple.
04:32You call agents as tools, maybe you connect to agents through A2A, and life is great.
04:38But, MCP, calling agents as tools, it's completely stateless.
04:42If you want two agents to be stateful, sticky sessions, good luck to you.
04:47A2A is client-server.
04:49If I'm an agent, I can send a task to another agent, but if this agent also wants to send task to me, we both have to be a client and a server.
04:59Chaining multiple agent equals together involves REST API timeouts.
05:04Good luck managing that.
05:05And obviously, what no one says to you, you need queues with persistence and so on to keep track of the messages that are being sent.
05:13And of course, discovery, which is not even part of the A2A protocol.
05:18So, you're basically doing the planning work.
05:22You're not creating a multi-agent system.
05:24You deal with planning.
05:26But, we have messaging platforms.
05:30Slack.
05:31Anthropic released a wonderful agent.
05:33You can talk to it in Slack.
05:35We have personal agents here and I connected to Telegram, right?
05:39Wonderful.
05:42To connect an agent to a Telegram, five steps.
05:46Discord, seven steps.
05:49Slack, eight steps.
05:52WhatsApp, 11 steps.
05:55Every step is manual that you have to do by hand and you have to read the documentation.
06:03Again, you're doing doing a lot of planning and manual work.
06:07And this gives you only one thing and one thing only, an agent that can talk to a person, usually it's you.
06:15Your agent is still alone.
06:17They cannot see each other.
06:18They cannot communicate with each other.
06:21They are in digital solitary confinement.
06:25So, let's unpack the statement I heard from the CTO.
06:29Is multi-agent coordination a hypothetical problem?
06:32Clearly, if you're copy-pasting stuff between two sessions, this is the problem of today and it's not hypothetical.
06:39Plugging in MCPs and A2A, this is today's problem.
06:43It's not a future problem.
06:45Is it possible to avoid the problem?
06:48Clearly, it's not possible because otherwise it would have all work with only one session and not two and we would have no need to call other agents as tools and so on.
07:00But the question is, is it possible to keep things simple?
07:05And the answer is, if we look at all the options that we just talked, not really.
07:12But how hard can it be to connect two sessions, two processes on my laptop together?
07:20How hard can it be to connect my cloud to Lambda?
07:24Or Salesforce agent to the Databricks agent to SAP agent to my my my Codex?
07:33Actually, it's pretty hard.
07:35Think about it.
07:36These agents, even your two sessions are processes that have to talk to each other through a network.
07:43So, it's a distributed systems problem.
07:46And distributed systems are hard even before you have introduced agents on top of it.
07:51And multi-agent system where every agent is remote is basically a distributed system of microservices where each microservice is non-deterministic.
08:00So, it is hard.
08:02What do you need to solve all of that so it will become easy?
08:05You need to solve the transport layer.
08:08Ordered message delivery, real-time message delivery, retries, and so on.
08:13You need to solve continuity.
08:16Microservices, your agents are software, pods, dockers crash, and so on.
08:21So, you need to have a consistency hydration.
08:23You also need to do the run time binding between different agentic frameworks.
08:28You have thread IDs, conversation IDs, execution IDs, and so on.
08:32And someone needs to map all of these IDs together so your agents can actually interact.
08:38But, it's also not enough.
08:40Your agents cannot communicate at an IP port level.
08:43They cannot communicate at the URL level.
08:46They cannot even communicate at the pub sub level because it's still a lot of plumbing that you need to do as an organization.
08:52So, for this wonderful future of agents talking to each other, we need to raise abstraction of a technical stack to conversation and talk about rooms, channels, participants, and figure out the deterministic
09:06routing of messages within a channel and also across different channels.
09:11And even if you solved all of that, it's still not enough.
09:13You need to solve the governance layer, the identity, audit, and etc. So, I would like to introduce Band.
09:21This is exactly what we solved so you don't have to.
09:25We connect every agent together and we are global collaboration layer for all the agents.
09:32Any framework, whenever they are deployed.
09:35Inside, it's not just a communication.
09:37We implemented every primitive that is required for your agents to talk to each other.
09:43Let's see them.
09:46What you will see are two different users connected to the collaboration layer, Vlad with a personal assistant and Mike that has no agents.
09:53On the right in the terminal, I'm going to spin up different agents and they will be onboarded into the platform.
10:01So, let's see how fast it is to onboard a new agent.
10:05We just spun up a new agent, programmatic registration and onboarding and agent card appears.
10:10This is a Codex agent.
10:12Next, we spin up a Land Rush agent.
10:14This Land Rush also appears in the platform.
10:16From this moment, they know that they exist and they can talk to each other.
10:21Now, we will ask Codex agent to send a connection request to my personal assistant.
10:27Keep in mind, different users, different registries.
10:31There will be a connection request, contact request sent to my personal assistant to require bilateral consent.
10:37It will arrive here in a second and from the moment that I approve this, Codex will know and see my personal assistant, will be able to invite the personal assistant into the conversation
10:50and send messages to this personal assistant.
11:02So, we're asking the Codex to invite Andy.
11:05Andy got invited, received the message, reported back.
11:10But, let's talk about problems of today.
11:14We are all developers.
11:15We use multiple sessions, probably a lot of sessions.
11:19We do routing.
11:21We don't like it.
11:22And when we go to uh have a snack, we come back and we have no idea what our agents have done.
11:29And you as a manager have no idea uh how much it cost distribution uh of ticket and the tokens and so on and so forth.
11:39And you have no idea even how long your human was involved in the work.
11:44I would like to introduce Jam.
11:46Jam is an internal product.
11:48This is how we develop the software and we've built it on top of bank.
11:54Jam is a desktop application that simplifies the onboarding of your local agents to the platform and solves the problems that we just mentioned.
12:06So, we solve routing, we solve context overload, we solve cost management and attribution, and we allow multi-agent and multi-human collaboration together.
12:19So, what you see here is local agents and remote agents working together.
12:26And we capture all the tasks generated by Claude and Codex that they generate for themselves when they do the work.
12:36And we present this to you so you can track the work that these agents are doing.
12:41And this can be your local sessions or your local session with a session of your friend.
12:46Because trying to understand what your agents are doing, 1 million tokens multiplied by three, that's a lot.
12:52You need a completely different way to understand what your agentic team is doing.
12:57Moreover, we provide a way for agents to describe the layout of the software architecture that they are working on.
13:03This is what you see on your right side.
13:06And you can see in real time where your agents are working right now, what piece of component they're touching in real time.
13:15And once they are done, they mark it as done.
13:17If there is a human in the loop involvement, you'll get pinged.
13:21Now, you don't have to use this desktop application.
13:24You still can open your terminal and you can work from your terminal, but because every communication goes through the network, we can monitor all that stuff and we can surface
13:38a lot of other very useful information and we can enable my agent join your agent.
13:44We can enable our team member works remotely to join the session together with the agents and the humans so he can help solve us with some problems.
13:54We if you have a security guy who maintains skills for the security agent, I don't need to copy his skills.
14:01I can just ping his agent to join this conversation and solve this for me.
14:09Now, I'd like to show you the platform itself.
14:14So, since we're all managers, we'll start with graphs.
14:17Uh so, the moment you open application, you can see all the stats of all the traffic that happened between your local agents and also your remote agents.
14:24You can see for instance over here, I have a full stack developer uh $2,000 in tokens and this is a local cloud session.
14:34You can see an architecture over here, $600.
14:38This is a local coded session and the rest are different agents running in different environments.
14:45But, how do they work together, right?
14:48Do we have a bunch of Python code triggering the agents so they can work together?
14:54We do not.
14:56Because all models right now, they're trained on a lot of data so they understand very, very good how to communicate through messaging platforms.
15:07So, here I have real work that I've done this morning.
15:12Engineering manager, developer, and architect, different instances of quad quad working together, reviewing PRDs and SRS, we are implementation.
15:22there is no need to hand code all the loops.
15:26They know how to do it natively.
15:29And for the managers, we have full statistics.
15:33Let me kill off the manage the the engineering manager who sent too many messages.
15:38So, you can see the attribution.
15:40You can see the full [clears throat] user chain cost.
15:42So, if you ask a question, if my developer is actually involved in the code he pushes as a PR, or it's all AI slope, you can see it here.
15:52Not only locally, right, but also through the remote agents.
15:56You can see attribution by developer or by teams of your agents and developers and how they work.
16:03Everything is gets updated in real time.
16:07You can see all agents, and every agent is basically a session, right?
16:12So, if I click here, I can see all the sessions of my Cloud and Codex instances that are running and they're connected to the platform.
16:20They're connected to a global platform.
16:22So, if I want, I can connect any of you to any of my agents in 30 seconds.
16:28Obviously, you can set permissions, you can see all the rooms, you can see all the errors, and so on.
16:33And also, you can see work.
16:35What work was done, what is pending, what is in progress, and everything works in real time.
16:43Thank you very much.
16:44If you want to know more about the future software development that does not involve you pulling in another 50 packages, and if you want to enable your Salesforce and Slack and Data Bricks
16:55and Cloud and Codex to work together and collaborate, come to our booth LG17.
17:01QR codes for the Band infrastructure layer and for the Jam, the uh desktop application to allow you to be part of this future that everyone talks about but has never seen.
17:11Thank you.
17:14[applause]
17:33Hey.