Slack for Agents — Building the Internet of Agents

Hey everyone, welcome back here to another Techstrong TV interview. I want to introduce you to Arik Gumenovsky. Arik is the CEO, I believe, of a company called Band, and they're doing some interesting things around agentics I want to talk to him about.

But first, let's welcome Arik. Arik, welcome to Techstrong TV. It's great to have you here.

Well, thank you so much, Alan. It's great to be here. Thank you for having me.

I know we've been trying to get this done for a while, between I was traveling and it's conference season, but I'm so glad I finally got you on. Arik, before we jump into Band and talk about agents, let's spend a minute or two talking about you. Give our audience a sense of how you got here today.

Yeah, well, pretty straightforward. I'm a startup guy. I've been doing startup for the last 11 years.

Before that, I spent 17 years working for the government. I come from a cybersecurity background, so my two previous adventures were in the cybersecurity space. I co-founded a cybersecurity services company doing a lot of incident response work in North America called Signia.

And then I went on to co-found an identity security company called Aromatic that we sold to Tenable in 2023. I spent two years at Tenable- Sure ... leading product innovation and doing mostly M&A, looking at other interesting startups that we could potentially acquire.

But my dream since I think 2018 was that one day agents will take over the world, and we have to help them to do that successfully. And I've dreamt about building something that we're working on today, but obviously 2018 was way too early. But with all the LLM revolution, I thought that I cannot continue sitting on the sidelines, that it's the time to go and build Band.

So here we are. I love it. When did you leave Tenable?

Roughly a year ago, just when we started this company. Okay, so you were there unfortunately when, Yeah, when Amit passed away. Amit passed away.

Yeah. He was a friend. He was an amazing person.

He was. Yeah, he was an amazing person, great friend, and a great mentor. Yeah, it was heartbreaking.

And I knew Amit a long time. I know Dov very well. Anyway, onto better things.

So you've been dreaming about Band since 2018. What exactly is your dream? What was the dream?

So, back in 2018, I looked at the market and I said, okay, agents are, or maybe we called them machines back then, the word agent wasn't as popular, or algorithms, they're really becoming economic players. Right? I look at the financial industry and the algo trading market, and I saw algorithms trading in very high volumes.

Hundreds of millions of dollars automatically without any human intervention. And I said, okay, why is it limited only to algo trading? Obviously in algo trading there is a clear benefit because of the high volume and fast transaction that everywhere, one day agents will be able to start making economic decisions.

They will sell, buy, negotiate, decide on the price, et cetera. And within the business, make other business decisions, not only about pricing, et cetera. And, I thought, okay, what has to happen to enable this amazing future of what today we call the agentic economy?

And the answer is pretty straightforward. Obviously, the agents themselves and their reasoning has to become more mature, and we have that with LLMs now. But second, agents need to be able to collaborate, just like humans do, among themselves, and also with humans.

And we identified that as a missing piece, the missing piece in the agentic stack, if you like the tap talk. And that's what we are out to solve. So basically, what we are building is an agent interaction layer that can connect any agent out there to any other agent out there.

They can discover each other, and they can collaborate, just like humans do. They can talk. It's a full duplex, bidirectional communication like between you and I, and they can also do that in groups, in multi-player conversations.

Like someone could now join our conversation. Same could happen with agents as well, and we support that. So, I first became kind of aware of this agent-to-agent communication, of course, probably like most people, when Claude first came out.

OpenClaude, whatever it was called in the beginning. And someone did sort of a social network and it went viral. " And one dude was, "Ah, it wasn't really agents, it was people spoofing as agents, but it was really the people," but what have you.

Clearly, today's agents are capable of communicating with each other, which is a good thing. Where that was more of a social network, though, it sounds like Band is more about business. Right?

And how do you view the early OpenClaude social network versus what you're doing with Band? Yeah, sure. First of all, what happened with OpenClaude is obviously amazing.

This is a great ecology. I think overall, OpenClaude's probably one of the three main inflection points that we've seen in the last few years in the evolution of AI. It all started obviously with LLMs.

The ChatGPT moment everyone talks about. And when we started talking to people a year ago, two years ago, about agent-to-agent communication, obviously people thought that this is a far future. But OpenClaude and Mollbook, they opened the eyes for many people out there, and, Alan, you talked about yourself, that agents can in fact collaborate and create value through that agentic collaboration.

And Mollbook was one of the best market educations we could have asked for. But Mollbook is more of a forum or a social network, if you like, for agents. But in the business, but in also in real life, what you need is not a social network.

If Mollbook is the Facebook for agents, we are the messenger or the WhatsApp or the Slack, right? Organizations need collaboration infrastructures where agents can come together and solve business issues in real time, where they can have peer-to-peer and multi-peer conversations. And this is how we view what Band is building, right?

So we are the, if you like, the Slack for agents or the Teams for agents for the enterprise. So Arik, this is all great, right? But with all of the different agents out there, right?

OpenAI has one, Claude has one, NVIDIA has one, OpenClaude has one. There's NanoClaude, there's Perplexity, Microsoft. Everybody's going to have their own agent, Salesforce, ServiceNow.

How do we get all of these agents to talk a common language? Because that's really the key to the communication, right? Yeah.

So first of all, I think you're absolutely right. I think the language issue is pretty straightforward, right? Just like humans, we talk English, right?

And the famous joke about UK and US, right? Two great countries separated by common language, right? I'm from Pennsylvania.

Yeah. Exactly. So I think same goes for agents.

Although now we're discussing with potential customers in Korea and apparently agents in Korea speak Koreans. So it's a little bit- And I'm sure the Chinese will as well, right? Because my lesson- Of course ...

traveling to China a few times is not everyone there speaks English, that's for sure. Of course. But I think the point is that agents must speak in natural language, and this is one of the core elements of our thesis, and there are a couple of reasons for that.

One, because you don't want agents to be segregated in their conversations from humans. You want a joint economy. You don't want a segregated human economy and separate agents economy.

You want that to blend, right? So you want agents and people within the enterprise working together. So humans have to be able to participate in agentic conversations.

That's one. And secondly, you also want to be able to supervise what agents are doing, right? And humans need to be able to look at those agentic conversations and make sure that, first of all, they are correct, that the agents are doing what they were designed to do.

Because ultimately, agents are just like software. There could be bugs, there could be issues, right? Someone has to troubleshoot them.

So in order to be able to do that, you need to look at that in normal language, right? So these are the two main reasons that we would argue that agents would speak a common language. And also that's where the power of all LLM is harnessed in the best way possible.

You don't need now to maintain different schemas and translate between schemas. LLMs can effectively translate everything to English, right? So I think that's the power.

But obviously you need to give all these agents, and your point about the diversity of agentic platforms is absolutely true, and that's one of the other core reasons why you need Band, because every ecosystem would want you to come to their environment. I don't think that Anthropic would like you to use Claude Code together with products from OpenAI because they want to keep the tokens for themselves. And same goes for the hyperscalers and the cloud infrastructure and for the SaaS vendors.

But ultimately, an enterprise would have all of that. Enterprises today already have a lot of agents on the edge, whether these, the coding agents, whether these the OpenClaudes of the world. They have agents in the cloud that are building different frameworks like LangGraph, CrewAI running in AWS, ADK in Google, et cetera.

And then they have, to your point, agents running in SaaS, Salesforce Agent, ServiceNow Agents, SAP Agents, all of these build on these SaaS platforms. And now to bring all these agents together, there are a lot of technological challenges that bubble up when you try to do that, because from an engineering perspective, what you are dealing here with is non-deterministic distributed microservices. And each and every one of these three domains is very hard.

The non-deterministic applications are hard, microservices are hard, and distributed systems is very hard, right? So you need a real solution that can actually allow that to happen across an enterprise. I get it.

So one of the things I've heard counter to that, though, is that human-to-human communication is pretty slow, and human-to-machine communication is slower, but machine-to-machine communication can be very fast, and especially if you strip out what they need to put in there for us humans to understand, right? How does Band deal with that? Yeah.

So first of all, that's a great observation point. I think if you look at multi-agent systems, they're not IO bound. They are compute bound because ultimately every message gets processed by an agent, by an LLM, by a reasoning loop that calls that LLM multiple times, right?

So there are a lot of LLM work that goes before or between agent messages, right? And that's where most of the time is being spent, not on processing the actual messages, right? But to your point, yes, agentic communication will be more speedy and much more high volume than human communication.

That's why you need to build the infrastructure right from the get-go. We believe that the day is not far when on our platform, we will have more traffic than WhatsApp and Discord and Telegram and Signal combined because there will be more agents than humans, and the agents are way more chatty than humans are. They spill out text in masses, right?

Like when you're texting someone, you count every word because you don't like typing. Mm-hmm. They can spill out a full document very, very easily.

So, we have to support all that. By the way, that's why we built our platform from day one on the same tech stack that WhatsApp and Discord are running on, right, which is built for large-scale messaging, global messaging platforms, to support that future scale, right? Because we understand where the industry is going.

Let me ask a little bit about the business model, Arik. How do you charge, based upon how many agents you have, how many humans you have, a combination of it, flat fee? So first of all, it's an amazing question, and obviously we kind of think about that all the time.

And I think whatever we have now probably will not be the last word of that, right? Because that's an evolving market. First of all, we have a freemium platform.

You and I and everyone can go online, log into our platform, and start connecting their agents. And actually, your agent can talk to my agent, over our global communication platform, which is free, just like the messaging platforms are free of charge for personal usage. And then we have an enterprise license.

The enterprise license obviously offers the enterprise not only the communication, the agentic mesh that allows the agents to collaborate, but also the control plane for agent interaction, right? Where the AI ops can see everything that's happening and can enforce policies and guidelines and roles and security capabilities and everything that an enterprise needs to scale multi-agent systems safely across the enterprise. Now, that is obviously licensed, and that is licensed by the number of agents you have.

It's not like a seat model. It's like the band where you fall in because we don't want to discourage people from adding more and more agents to the platform. But usually, an enterprise can forecast how many agents it's going to have in the next year, whether it's zero to 100, 100 to 1,000, 1,000 to 10,000, whatever.

And then that's what they're licensed for, and they pay for that specific tier. There is a subtle nuance here because there are some use cases where organizations have very few agents, think about customer support, but they have millions of customers. Think about a B2C business, right?

Then they have maybe five agents only, but these agents interact with millions of humans and have millions of interactions daily, which trickle also agent-to-agent interaction in millions, right? So we also have the notion of a session where basically agents come together to solve a business task. And ideally, by the way, we would charge everyone by the number of sessions they have because there is a clear alignment between a multi-agent session and a business use case that they solve together and the value we provide.

But most organizations just would have a hard time forecasting that. That's why we go for the number of agents. But in those specific B2C scenarios where they go into millions or maybe hundreds of millions of sessions, there is also an additional fee for each session above a certain number.

But basically today- So it's session based ... we charge by the-- Yeah. But basically today the core B2B scenario is just number of agents and that's it.

It's going to be interesting because look, I've seen, and you've seen these too, I'm sure, that each would basically have 100 to one, right? 100 agents for every person. You can get crazy.

And if it's 100 to one is average, you'll have some people who have 1,000. But nevertheless, I do think this is an important part of the infrastructure of the way these things are going to work. Because agents shouldn't be in a vacuum.

Exactly. Exactly. Look, can you imagine an organization where people are not collaborating today, right?

Like even in the factory where people... What was the first industrial revolution? What was it all about?

It was about putting people into an assembly line, right? Mm-hmm. Creating a workflow.

That was the first industrial revolution, right? And that's great, but we've evolved beyond that, right? Now people in most organizations are not working assembly lines.

We hopefully replace them by not very intelligent machines that can do that, and people are doing creative work and collaborating, et cetera. Now, what are we doing today? We're bringing agents, but the existing multi-agent workflows, or what people are calling orchestration, is basically putting these agents into assembly lines.

But that doesn't allow us to gain the full power of their intelligence. The full power of their intelligence will be when they can actually collaborate in the organization, just like humans do. And at some point we wanted to go with the slogan, "Intelligence freed," right?

So we're kind of freeing the intelligence in the organization. But that's a little bit more future-ey. I like the Band thing.

It's all good. Anyway, Arik, we're about out of time. I don't know if we mentioned the website, though.

It's very straightforward. A-I. I love it.

Please pay us a visit. You can log into the platform. It's free, and you can connect your agents, and your agents can talk to your friend's agents, and you guys can schedule some time out and enjoy and have fun.

I love it. Arik, good luck. Keep us posted.

Come back on and let us know- Thank you so much, John ... what's going on. All right.

Thank you. ai. Check it out.

We're going to take a break. You're watching Techstrong TV.