UiPath Forward 2024 | Ed Challis & Taqi Jaffri, UiPath

Clips
News
More from UiPath Forward 2024

Edward Challis

Head of AI Engineering

UiPath

Taqi Jaffri

Sr. Director, Product Management

UiPath

UiPath pioneers agentic automation, empowering autonomous systems to drive innovation and precision

Based on the urge for practitioners to focus on more complex, creative and strategic roles, the demand for agentic automation is growing because it drives efficiency, reduces costs, enhances precision and opens up new opportunities.As a result, UiPath Inc. is in high gear to deliver the agentic automation vision because it triggers precision, consistency, competitiveness, innovation and enhanced decision making, according to Taqi Jaffri (pictured, right), senior director of product management at UiPath.UiPath’s Ed Challis and Taqi Jaffri talk to theCUBE about the revolutionary effect of agentic automation.“Here we are unveiling this vision of agentic automation,” Jaffri stated

play_circle_outline Distinction between bots and agents in terms of defined actions and objectives

play_circle_outline Importance of automation as a prerequisite for successful agentic systems

play_circle_outline Role of specialized agents and orchestration in the new agentic stack

play_circle_outline Trust layer for security, governance, and monitoring in agents and bots

play_circle_outline Navigating Agent Frameworks: Orchestrating Specialists for Cambrian Evolution in Industry

Info
Transcript

Ed Challis & Taqi Jaffri, UiPath

Edward Challis

Head of AI Engineering UiPath

Taqi Jaffri

Sr. Director, Product Management UiPath

The senior director of product management at UiPath, Taqi Jaffri, and the Head of AI Engineering at UiPath, Ed Challis, discuss the vision of agentic automation at the conference. They explain the difference between bots and agents, emphasizing that agents have a goal and decide how to achieve it, while bots follow predefined procedures. The importance of automation as a prerequisite for successful agentic systems is highlighted. They discuss the key components of the agentic stack, including generative AI models, trust layers, and data harmonization. Th... Read more

explore Keep Exploring

What are the differences between a bot and an agent in terms of their actions and decision-making processes? add

What is the importance of RPA in successful agentic systems? add

What does the new agentic stack look like? add

What is the relationship between trust, agents, robots, and the trust layer in the context of secure activities? add

What are some potential challenges and opportunities in the evolution of the industry? add

bolt Powered by CUBE AI

Ed Challis & Taqi Jaffri, UiPath

search

>> Good morning everyone. Welcome back to theCUBE's live coverage of UiPath FORWARD 2024. We are here in day two. I'm your host, Rebecca Knight, alongside my cohost and analyst, Dave Vellante. We've got two guests for this section, Taqi Jaffri, he's the senior director of product management at UiPath. Welcome, Taqi.

Taqi Jaffri

>> Thank you.

>> And Ed Challis, Head of AI Engineering at UiPath. Thank you both so much for coming on theCUBE.

Dave Vellante

>> Thanks for having us.

>> So I'm going to start with you, Taqi. You were on the main stage this morning. I think you're doing seven keynote addresses.

Taqi Jaffri

>> Keynote .

>> And you're demoing here, you're demoing there. Tell us a little bit about what you're doing here at the conference.

Taqi Jaffri

>> It's been fantastic. First of all, thank you for having me. Here we are, as you know, unveiling this vision of agentic automation. And I'm just fortunate to represent the work of a fantastic team. I'm just the face of it. And we are unveiling many things. One of the premier thing is the agent builder, so how RPA developers can build their own agents and deploy them. And we like to say that it's not just about deploying agents, it's about deploying agents that work well. So we are also, as part of the agent builder, giving you the ability to evaluate them, test them, and actually monitor them while they're running. So it's a very, let's say, deeply rooted vision that we have. So just excited to be here and talk about that with you today.

Dave Vellante

>> So if I had to summarize the strategy, I'd say, okay, UIPath strategy is to leverage the core capabilities that it's developed over the last 10 years and installed in 10,000 plus customers and then supercharge them with AI. It's like Jensen says, "We're going to supercharge you with GPUs. Well, you're going to supercharge your customers with AI. You're going to build on top of that." So my question is, first of all, is that an accurate simplified version? And is it necessary to have that underlying infrastructure? And if so, why?

Taqi Jaffri

>> Yeah, I think that is a great question. Yeah, I think at core screening level is definitely a technology which can supercharge the entire digital estate of a business. And in some sense, I would say that automation is a necessary prerequisite. So we heard our chief product officer make this analogy that agentic without automation is like a plane without a pilot or a body without a brain. Without taking actions, without causing changes. You are ultimately just a advisor. You can provide useful text, but you're not actually doing anything that moves the needle. So automation closes the gap for AI and is a really critical component.

Dave Vellante

>> Are you using software robots and automation as interchangeable? Is that?

Taqi Jaffri

>> Yeah. And I'll just take that. The way we think about it, and we actually set this on stage, is agents are nothing without the tools. Because to me, an agent needs to have agency, which means it needs to act. Like Ed, said more than just advise you with some text. It actually needs to take an action. And in order to take action, we believe the best way to do that is with robots. And we've spent, if you think about it, decade or so, preparing for this moment almost by shipping all these robots, which are well governed, they're trusted, they work well, but they're very deterministic and that's a good thing. And now we have this non-deterministic agent, which is launching them and running them. So that's how I think about it.

Dave Vellante

>> Okay. Let me back up. Help the audience understand. Help me understand. Define the difference between a bot or an automation and an agent.

Taqi Jaffri

>> Want to take that or?

Dave Vellante

>> Okay. I mean I think, so a bot is, the actions it takes is completely defined by the control flow that the developer creates. So there's logic and a predefined set of procedures which define the flow. An agent is defined by an objective, a goal. And then it has tools at its disposal, a memory at its disposal to try and achieve that goal. And it does so in a recursive way. So it will maybe get some information, read that information, decide what next step to take. And iteratively do that until it thinks it's achieved its goal.

Dave Vellante

>> And a bot is better than microservices because it's governed, it's secure, it's got a richer set of capabilities, correct?

Taqi Jaffri

>> Yeah, exactly. So with a bot, you not only tell it what to do, you tell it how to do it, you code the bot.

Dave Vellante

>> Yep.

Taqi Jaffri

>> With an agent, you just tell it what you want done and it decides how to do it. It comes up with the plan dynamically. So that to me is the key difference, which is the control flow, the how it achieves the objective is determined on the fly by the agent. Indeed, you can run it multiple times and it might achieve the same objective multiple different ways because sometimes these things are stochastic, non-deterministic, right? So I think that's the key difference. The microservices analogy is interesting because to me that's about factoring or decomposing agents into smaller pieces. And we actually talked about that, Ed, this concept of having these single-minded agents. Maybe you want to address that question about decomposing things.

Dave Vellante

>> And the reason I use that analogy, because they're hard coded.

Taqi Jaffri

>> Yeah.

Dave Vellante

>> Right?

Taqi Jaffri

>> Yeah.

Dave Vellante

>> But whereas I think agents are not right, as you described.

Taqi Jaffri

>> Exactly. Exactly.

>> Yeah. I think there's two ways. The microservice there, I mean a bot gives you more access to more information go via the UI, process unstructured data. They're just purely programmatic. But yeah, agency, it can solve fundamentally new problems. Ones where the data is too variable for you to hard-code specific instructions that maybe the data changes from run to run or situations where there's just too much complexity to define the path. There's too much nuance and the process is a little bit too messy. And then there's a whole bunch of tasks that take up a lot of people's tasks, but each task individually doesn't take up a lot of time. Doesn't happen that frequently. So you might have a whole bunch of travel-related work, but actually you're not searching for a flight hundreds of times. You're searching for a flight, then you're checking the details, you are maybe changing the seats. There's a whole bunch of different actions there. So agents attack those kind of problems. The hard and maybe not worth it with traditional automation or the variability challenge.

Dave Vellante

>> But you guys are making the case, if I understand it correctly, that RPA, automations, bots, whatever we want to call them, is a fundamental, I'll even say prerequisite to successful agentic systems. Is that a fair statement?

Taqi Jaffri

>> Absolutely, because if you just want your agent to, in Ed's words, just advise you, then perhaps don't need robots. You just need a chatbot, you ask a question and just answers you in text, then you don't necessarily need robots. But if you actually want your agent to take action on your behalf, you need to give it these tools. And again, we believe these robots, these robotic automations that UiPath has spent years building are the best kind of tools. So yes, that's a very accurate characterization.

Dave Vellante

>> I mean, I want to build systems of agency that transform my business and give me massive competitive advantage and cut my costs by 10X, and dramatically accelerate my time to an outcome.

Taqi Jaffri

>> All you have to do is tell it to do that.

Dave Vellante

>> That's what I want. That's the future that I want. And my big takeaway of this conference is you need automations to do that. The companies, your customers that have built those automations are in a much better competitive position. So if I can figure out who those... We look at 10 companies and tech stocks, we say, "Oh wow, NVIDIA. Well, great. If we could have called that as many folks did." But if I could understand the customers that are deploying these systems, whether it's GE, GE's actually back on fire, which is kind interesting, I wonder how much your automations have helped. But what I'm saying is if I could bet on those companies that are applying those automations, they're going to dominate their industries. You're going to feed them, they're going do well as a result.

Taqi Jaffri

>> Absolutely.

Dave Vellante

>> But that's a really interesting investment angle that I hadn't really grok'd until I came here.

Taqi Jaffri

>> Yeah. But I'd actually like to double click on something you said where you just tell it what to do and it does, there's actually a really interesting tweet that went really viral by a famous science scientist, Andrej Karpathy. Some folks might've heard of him. It was from about two years ago. And just a one-liner, which is, "English is the hottest programming language."

Dave Vellante

>> Yes.

Taqi Jaffri

>> So it's a really amazing way of saying it, what you just said. But I think there's a corollary to that statement, which is, well, English is also vague sometimes and ambiguous.

Dave Vellante

>> And nuanced. Yeah.

Taqi Jaffri

>> And nuanced, right?

Dave Vellante

>> Yeah. You're right.

Taqi Jaffri

>> Code where you're very specifically telling it. Another really, it's a nuanced point, but I'd like to still try to make it, which is yes, you just tell it what to do, but you, A, have to be very, very specific. And the way we think the best way to do this is to have a bunch of specialized agents, which are single-minded, and they do one thing and then you connect them in an orchestration, which is exactly what UiPath loves to do. We actually spent a bunch of time talking about controlled agency, which is this entire concept. And the previous keynote that we just came from.

Taqi Jaffri

>> Yeah, no one-to-one, you can't have an agent that is an equivalent of a person. An agent will do a much narrower set of tasks and really focus on that. And then the benefit of that is it makes it much easier to test, much easier to validate, easier to improve. So that's kind of the pattern that we're going to see emerge really specialized agents and orchestration to bring it together.

Dave Vellante

>> So what is your vision for how humans and agents and robots will work together? I mean obviously we know English majors are going to have a job security for life now in this new economy, but how do you see them collaborating and how is this going to change workforce dynamics?

>> I think that's a really interesting question. I think time will tell, right? And what we as a society have to do is we have the tools now, we need to just iterate and see what works. Because sometimes the models underlying this is generative AI and sometimes it astounds you with what it can do. And sometimes you're really frustrated-

>> .

Dave Vellante

>> Because it can't do it, deflated. You're either astounded or deflated.

Taqi Jaffri

>> That's true.

Taqi Jaffri

>> And actually even people that work in the field can't confidently predict, I really thought I'd be able to do that, or wow. So I think in every line of business, there's a process of trial and error that actually has to happen. And that's going to take time to propagate through our .

Dave Vellante

>> Can we paint a picture? I love working with you because you go right to the humanization aspect. I want to dig into the stack. So what is the new agentic stack look like? I got infrastructure as a service cloud on prem. That's cool. I have data, whether it's a data lake or a Snowflake. I got a bunch of data. I feel like there's got to be a way to harmonize that data. You can help me understand how you do that. And then you've got this agent control framework. I think you guys are calling it orchestration. It can interpret top-down goals and then it can execute bottom-up outcomes. It's got connectors to all these which you have through APIs or other integrations to backend systems, whether they're analytic systems or CRMs or ERPs. And there's a UI emerging, which is according to Daniel going to coexist with existing CLIs and other interfaces. But there's an NLP interface as well emerging. And that's this rough picture of the application stack that's emerging. How do you guys see it and where do you fit in that?

Taqi Jaffri

>> Oh, there's so much to say. So maybe I'll start bottom-up maybe or top-down. But let's just go bottom-up.

Dave Vellante

>> Sure.

Taqi Jaffri

>> So well first is the actual models, right? Like Ed said, this is powered by generative AI. So we have a partnership with Anthropic that we announced. They have fantastic models. OpenAI of course is doing great work. There's a fantastic work going on in the open source world with Meta's Llama family. So I think Daniel put this perfectly. He said, "We are going to be the Switzerland," I think he said. So we'll work with everybody, although we'll have partnerships with some. So that's the first thing. That's the model. And the model is the underpinnings of all of it. Then above that, there's typically a trust layer, because you want to make sure the model is complying with your policies. If there's personally-identifiable information, PII, that you want to strip out. So we have our LLM trust layer. Others do as well. Ours is pretty good. And then on top of that, you have essentially what you said, these ways of pulling in data. So you just don't query these models directly. You want to pull in some data and then you query them. So there's all these connectors, tools we call them. And these tools go both ways. You pull data in to give to the LLM, but sometimes the LLM responds back and says, "Hey, to fully service your request, I need you to go query a system or something." So these tools are both ways. In the art, this is called function calling sometimes, right?

Dave Vellante

>> Yeah. Sure.

Taqi Jaffri

>> Some of these models natively support function calling. And then I think it's very important to have this conversation that you briefly said, which is what are the best user on top of these agents? So there's of course the conversational experiences that everybody's familiar with, ChatGPT, which is just amazing Claude, all these great experiences. But not everything can be achieved with a conversational experience. And indeed, you're already seeing these agents evolve. Sometimes you ask them a question and they respond with a visual. They respond with a chart now. And then everybody's heard of these things where you can actually send in a voice stream or send a picture to the agent. So these things are becoming multimodal. But I think those are still roughly conversational experiences. I think you have an attachment to a message, but I think it will really evolve. And I like to think about it as of course, we support these conversational experiences. We have autopilot for everyone. You can also build a conversational experience in our agent builder. So this is table stakes, of course we do this. But what I'm really excited about personally is what I call, it's not an official word, but like an ambient agent. Something that does not really have a user experience. It's almost like an activity in a workflow. It's taking information from somewhere and putting it somewhere else with zero user experience. So I think that to me is the stack. I went really fast. Maybe you want to .

>> That was a very comprehensive.

Taqi Jaffri

>> Oh, there's you go.

Dave Vellante

>> Can I ask you a question about that? So the trust layer that includes security governance, is in that layer?

Taqi Jaffri

>> Monitoring.

Taqi Jaffri

>> Yes.

Dave Vellante

>> What's the relationship between the bot? Because we heard yesterday that you trust bots, you're not going to trust the LLM. Makes sense. How does the bot take advantage of that trust layer? Is the bot encompassed in the trust layer? Is it injected into the bot?

Taqi Jaffri

>> Maybe I can talk about that, it's a product question. So I would slightly rephrased that. I think what we said is you can't trust agents, but you can trust robots, right?

Dave Vellante

>> Yeah.

Taqi Jaffri

>> And I think you actually can trust agents if you only let them have agency through robots. I think Daniel said something on stage, don't give an agent your password or something. I think that was just a fantastic way of saying it. But I think of course you can have an agent do secure things through a robot so the robot can have your credentials and all these things. It can be governed. So the way I think about the trust layer is it's the sandbox to both of them. And today before agents, robots can actually use generative AI, can use LLMs. We have these gen AI activities and things like that that all go through the trust layer. So you can put rules saying, here's my business policies, don't violate them, etc. And agents will just go through the same trust layer. So it's the sandbox, it's the controls around the sandbox for both of them.

Dave Vellante

>> So essentially it can't do anything without that trust layer?

Taqi Jaffri

>> It has to go through the trust layer.

Dave Vellante

>> And what about, I'm stuck on this data harmonization. It just seems to me that there has to be a way, I got all these different data sources. How do I know that revenue is revenue and it's not bookings and it's not NRA or quarterly revenue or calendar year revenue? How do I know that? Who figures that out? Is that the agent? Is that the bot? Is there another layer that emerges from a third party? Is it a knowledge graph? Does it matter?

>> I feel that's a problem that needs to be resolved independently of this. Agents and humans, you can view an agent as an approximate human, and humans make those mistakes.

Taqi Jaffri

>> They have the same problem.

Dave Vellante

>> Okay. But that's not your swim lane, what I'm hearing.

>> Yeah.

Dave Vellante

>> But what I call the agent control framework absolutely is. And that's a really high value piece of real estate. And inherent it in there is the ability to access, I think you said this before, access tools.

Taqi Jaffri

>> Yes.

Dave Vellante

>> Whatever those tools may be.

Taqi Jaffri

>> Yes.

>> Exactly. And if you take that revenue example, you might use a robot to get the right revenue number to feed to an agent, and then developer has that design authority. So you're not letting the agent pick the column header that it thinks is most likely to be the revenue number.

Dave Vellante

>> So Rebecca and I were talking in our open this morning about our Nervana is a company like UiPath creates the abstraction across all these stovepipe applications, and we utilize your framework for our agentic system. But the reality of our industry is that Oracle's going to have its agent framework and Salesforce and ServiceNow and Palantir and all these apps are going to be in their own stovepipes. Your strategy, I think is to be Switzerland across those.

Taqi Jaffri

>> Yeah.

Dave Vellante

>> And that's going to be the interesting, I don't even want to call it a battle, it's just the evolution of our industry. I hope that you can succeed wildly.

>> It'll make all of our lives easier.

Dave Vellante

>> as well.

Dave Vellante

>> But these other apps aren't going to make it. They're not going to necessarily try to stop you from doing that, but they're going to want to control their own little domains and they're going to have little hooks and features that attract people.

Dave Vellante

>> But I think we're going to see this Cambrian explosion of agents, and it will make sense for big important applications to have their own agents that really deeply understand their world. But just as you do with people, we have specialists, then you'll need coordinators, managers, middlemen, an orchestration on top of that.

Dave Vellante

>> Yeah, makes sense. Absolutely.

>> Yeah.

Taqi Jaffri

>> It's actually perfect.

Dave Vellante

>> Ed and Taqi, a great conversation.

Taqi Jaffri

>> Thank you very much.

>> Thank you.

Taqi Jaffri

>> Thank you.

Taqi Jaffri

>> I'm Rebecca Knight for Dave Vellante. Stay tuned for more of theCUBE's live coverage of UiPath FORWARD. You're watching theCUBE, the leader in enterprise tech news and analysis.