Graph Exchange: Nikita Ivanov, From Agentic RPA to Self-Leaning Autonomous Agents
Recording: Graph Exchange: Nikita Ivanov, From Agentic RPA to Self-Leaning Autonomous Agents
Perfect. All right. So, we're all set up. So, it's pretty cool to see how many people on Wednesday night uh in San Francisco. So, a couple of words about myself. Uh my name is Nikita. Alex, thanks a lot for introduction. Um I've been in the space for quite some years
I've been in the NLP space for the last 15 years. Um, we can share the stories. I've been in obviously engineering space for since 23. I'm involved with a couple of startups. Barry are hands-on uh both as a founder and CEO and adviser and uh an investor. So I basically touched this subject from all different perspective. And what I want to talk to you about today is again we only have about 15 20 minutes. So we're not going to go deep anywhere but I want to give you my perspective of what where we are today in terms of our agentic development and we development agents and where I think we're moving again it's a lot of my perspective uh personally what I see from startups from investment standpoint what I see from startups basically trying to make it happen in real life very few succeeding and I think you know I'll point out a couple of ideas why this is happening so where we are we are in what you know colloqually at least I call is an RPA stage an RPA is a robotic process automation it's been terminology it's been with us for about a decade this point um it's predominantly been used up until recently for the automation so if you Google automation you know you inevitably come with RPA robotic process automation and um I think a couple of years ago when agents start coming up as a conversation topic we as a community almost automatically graft the same idea that we've had you know how many of you guys remember the tools like if then then this right you know yeah so quite a few hands but uh it's surprising how few of you guys actually know what it is but it's actually very close to what we are trying to do with agents today so the idea is still there when agents came up as a sort of a you know conversational topic and idea we kind of adopt the same same principles and RPA is basically one of the kind of a major thing about This RPA is the basically hand rolling logic
I'm sure most of you if you touch agents in the last I would say two years 23 24 even today uh you probably seen this gazillion amount amount of visual tools anywhere from N8 you know flow-wise and length flow where you can connect boxes here and there you can basically create this manual logical flow. Uh that's what RPA is and that's what is a dead end today. Um, you know, I I was given the same talk about, you know, a couple of months ago in December, and I basically we came up with very interesting ideas. If you guys do chain of thought prompting manually, you're likely working on a dead-end project at this point because this is essentially what's rapidly going away. And there's a multiple reasons for that. Um I'm sure most of you again if you're touching the agendic development manually hands-on day in day out you probably can relate to me that it's unexpectedly very complex development from engineering standpoint you're juggling too many things you're you're trying to do somebody told you that you know you have to do the multimodality you have to do multi- aent communication somehow you're trying to do that you're trying to test and maintain all this complexity around multiple agents and how to communicate and how to do all of these different things and all this prompting maintenance and all of a sudden you step back and you look at it and this is god damn it it's too complex. That's why by the way if you look at industry majority of agentic project fail basically people try they get excited you know they get the chart working they maybe get some basic rag working of itself is a can of worms but basically very very very very few agents today in production. This is a very telling sign that we have an idea
We have a definitely use case but you know we're just not there yet because the current implementation is fairly complex. Couple of things on the slide on the on the left side of me. Uh too hard to implement uh fundamentally no self-learning. Uh this is one of the biggest issues that we kind of quickly realized that you know we want to actually add new functionality new skills to our agents and if you do the RPA based stuff you have to code it. the minute that you're coding this logic. You're basically back to square one with this complex complexity and maintenability issues. Uh no autonomous behavior again something we kind of learned in the last I would say 12 months plus give it a take that to have something useful. You got to have a bit of autonomy
You cannot be relying completely on request response pattern. It works in the chat. But how many of you guys use chat beyond just a basic chart? Very few, right? Maybe the wipe coding and cursor something like this. But fundamentally the life is a lot more complex than request response. You got to have some autonomy. You got to have agent that have internal clocks that can basically wake up and have an idea. Uh it's not a joke anymore. We kind of have bench a bunch of use cases for it
And it's hard to implement this when you have this RPA based logic where you're literally, you know, rolling out Python code to create your branching classification and whatnot. I this is a bit of a controversial thing. I do believe that multi- agency is a dead end. Um I don't believe that's where we're going. I mean uh I've been part of the two projects that tried to do this prior to MCP with MCP doesn't really matter. Uh it's basically it's fundamentally dead end. I want to remind you some of you old enough to remember that in the five seven years prior we had a whole micros service slumber party right went to very high to the very bottom. Right now nobody talks about it
Just before that, if you're old enough to remember serviceoriented architecture, how many of you guys old enough to remember that? Spent billions of dollars were in nowhere. Just before that, now really, really, really adults here in the room could remember Corba. I think Alexa remembers that. Yeah. Okay, perfect. So, most of you guys young enough, Google it. There was exactly the same attempt to do exactly the same thing and it failed exactly the same way. So we're trying to do right now exactly the same thing with agents trying to basically orchestrate a a you know predominantly remote uh entities like agents and in some shape of work in some form it may or may not work a A2A is a good protocol we can look at how it behaves and how it advances but fundamentally uh I don't think it's a right approach technically and most importantly we don't need it you know most of what multi- agency can do for us we can do with normal MC CP style protocols that basically enable a frontier models to reason with a bunch of tools and that's where I think uh we are moving
So this is not a some kind of you know Ethereum right this is my take what I'm observing as a matter of fact what I'm building what I'm advising people to build and what I'm seeing from investment standpoint we are moving in a very interesting direction so I've seen at least I would say two brand new projects in the last I would say six months I was like literally involved with and there's a strong movement toward a what called a monolith AI monolith where There is a one frontier model predominantly at this point 3.7 clot because it's about the only hybrid model that does chart and reasoning in the same model. So you don't have to switch between between several and all of the essentially chain of thought prompting all of the you know uh plan creation and execution is pushed towards the model. So fundamentally again remember this very simple axiom if you do coot manually you're likely doing something wrong. There should be no CO2 code in your codebase. All that should be up to the model and something like CL37 can do it pretty effectively. I mean obviously O3 can do it very effectively too but then you have to juggle chart and the chart mode. Um it leads to very dramatic improvements in terms of productivity once you have a single model. Once you're not monkeying around with the Coot implementation and relying on the model uh all of a sudden your implementation becomes very simple
So some of the project in the deep research space actually utilize that and that's why you see this you know projects you know um manos you know you look at the codebase it's practically nothing there there and it's a very advanced project by results and that they basically adopted that you know sort of the same same approach of basically have a monolith um what comes with the monolith is idea of a broadbased core skills right you know what happened in one of the biggest events in the last year and a half is functions right? You know our models can now call functions and now we have MCP to kind of simplify this process standardize all that about this. So all of a sudden we can give hundreds upon hundreds of basic tools to the model to operate and that really switches the way how we think about it. You know imagine that we can do a an advanced coot right in the model. We don't have to write ourselves but attach bunch of MCPS. Now we will talk about separately why many MCP don't work yet but fundamentally imagine they do they will you can attach you know a a long list of MCPS and all of a sudden your model has a gazillion amount of tools available to it and it has internal COT built into it. So your life as an agent developer becomes infinitely simpler and more productive. You know a couple of things and then you know I have this human logger there. I mean some of the things is a little bit proprietary to us
For example, in humon we give our you know in the humon we develop AI workers basically next generation of things right and it's based on the plug you know our own agent implementation called blackbox what we do interesting is we we give our agent its own hardware and its own software access which is very important agents have to have access to the hardware so we deploy them in a dedicated isolated docker uh through docker basically dedicated isolated environment and this agent basically has its own hardware like networks and everything else in file system and operating system. It has its own software. It can fundamentally can install software and not yet but through either large language model, puppeteer or something like this can actually access. Puppeteer already works large language models don't yet. But fundamentally think about this with this new idea and it's not only us by the way you can deploy agents in a very isolated environment and give it its own hardware, its own software that works today. This is not a proposed it's it's we have a working code base that does exactly that. Our agent has access to the all operating system can you know features can create files file system can do whatever he wants to do we can figure out you know network connector users if we need to uh so that's actually also kind of a part of a zygeist where I'm seeing at least one of the project is moving rapidly because you know every time you know when you run you want to run a python generated code which is a very big deal right most of the new skills develop and I'll talk about in a second is about the writing python code that's as simple as this uh and when you execute python code written by by LLM. You got to have a normal environment and typically this environment you can use something like E2B but fundamentally the best way is to have a clean image that you're running operating system and let the Python run whatever it needs to run
Pretty cool idea but basically something that is again part of this new Z guys. probably the biggest um the biggest delta if you will is this you know what I call what I call a recursive unsupervised self-learning this is pretty big deal now you probably have heard multiple directions there's a world models uh basically the sort of next iteration of what we have today and those have a pre you know in model learning we can't even working on that because clearly we realize that you know this whole idea of a knowledge cut off is a bit of a idiocy. Why do we have a cut off and then why have to wreck everything after that? So why do models don't learn? So we kind of work in the next iteration but it's whatever the world models are five 6 7 10 years from now maybe today we need to have a solution and you know couple of companies human one of them developed this solution for a unsupervised self learning there's a lot of proprietary stuff here I'm not going to be able to tell you but here's an idea it's actually by the way it's a very cool that it's a very uh kind of inspired by um woming replication if you guys into the sci-fi and astrophysics you guys know what the f for non replication and how to explore a cosmos. But anyways uh idea is pretty simple. Uh remember this term uh we call the positive hallucination. You can ask a model you know give it a task to your model. Let's say you basically have a task and you you ask a model hey go ahead and create me a plan to execute this whatever the request is right. You can also ask it to hallucinate specifically on purpose and basic say hey imagine which functions would you like to have to complete this task and we call it positive hallucination
So model surprisingly is pretty good at it. I mean obviously Mel's pretty good at hallucinating but you can actually uh prompt it properly and ground it properly to positively hallucinating and model will give you hey I'm missing the function B A B and C and then you get okay fine great uh you get this function A B and C then you basically say look do I have those A B C let's say I don't have them right so I'm asking the same model in context give me a specific specification function A you go ahead make another call you write the code for this function through LLM get the code compile test deploy locally and Ask again. This is the recursion. You ask again, hey model, create me uh plan for execution and give me the functions you don't have. And by the way, you already have function A. And by the way, take a look at what else you missing. And so model can basically uh go in a loop like this. And surprisingly, and by the way, this is already live code in the live system
So it's not a you know theory. This way essentially through this local loop, model can quickly tell you what it's missing. You can quickly develop and give it back. In this way model learns a new skill not model but I'm sorry but agent itself learns a new skill. Those skills can be shared and they can do lots of different things in the background to do something like this. And obviously autonomous behavior again small thing but nonetheless uh agents should have its own clocks. You know we see this a lot of you know in MCP world but it's beyond MCP. You know agents have to be developed in a way that they have clock they have their own life cycle and heartbeats and can schedule tasks
Last thing, you know, uh this is actually it's not an extension of things and it's a as I mentioned it's a more of a evolutionary branch of a agents. We're definitely working towards AI workers. Again, this not a replacement. It's completely in parallel development. We'll need both agents and AI workers. Probably the biggest thing AI agents they improve individual productivity and AI workers improve organizational productivity. So think about you know you know AI workers as a hirable AI. Another way to look at it a softwaredefined workforce
Remember this combination of words. We're going to see it a lot more here a lot more in the next two three years is a softwaredefined workforce. Pretty big deal. Imagine companies building their own workers private or public. This is definitely happening. Google thinking this way. Human obviously think this way. That's what we're trying to do
And then I know at least another two or three companies here in Bay Area working in very much the same ideas. So I'll leave this uh slide in uh the last one. Uh we hiring anybody interested in good jobs, interesting things. Uh both of these companies do hire and I'm going around uh for the rest of the evening. So you guys want to chat a little bit more, I'm here. Thanks. [Applause]