Boomi CTO Matt McLarty: Expert Guidance on AI Agents

hi I'm James McGuire and on today's webcast we're talking about a tech term that gets used a lot these days but probably without much real understanding the term is this AI agent what exactly is an AI agent does it do real work is it like a human what is it to answer that I'm joined by Matt Mardy the chief technology officer Romi Matt absolutely thrilled to have you with us today yeah great to be here so I think there's there certainly uh you know confusion around the term AI agent I think it's used a few different ways there's you know the robotic process automation world where there's AI agents or AI Bots that that are trained there's a lot of AI agents and call centers that help out the staffers there but how do you see the world of AI agents specifically what distinguishes an agent from a traditional AI solution yeah I think um first off the reason it's so challenging to Define things I think is that this is true for AI General but for AI agents especially there's a giant gap between the promise and the vision of what AI agents can be and what AI can be overall right and then where we are today right so I think you brought up some good examples there everyone I think is familiar with the idea of chat Bots and maybe we don't have the best experience with with chat Bots about you but you know they've never fooled us into thinking they're real I never think oh gosh this is just like a human like no that has never happened to me no no they don't pass the touring test most of the time and I think um and then you know for those familiar with RPA and robotic process automation the idea of bots coming mimicking human behaviors and other type of thing I think the wave of AI agents that we're seeing now why there's so much excitement and people get really far down the road on envisioning what could be is because with llm technology with Gen technology we've already seen that that touring test that idea that machines can behave or think like humans seems to be more possible than it was before I think there's still plenty of time for debate on that topic but I think for this kind of the from a practical standpoint um I think agents generally are about their name it's about agency about having M you know machines or technology components that can can act somewhat autonomously can act um you know through reasoning so that's that's the idea of what agents could do now today they're they're not so capable like I think I'm a big believer in G's law not sure you heard of G's law no with G's law John G this is the idea that every complex system that works has to evolve from a simple system that works so we're the early stage I think the the Grand Vision of AI agents is going to be agent swarms these multiple agents working together figuring out complex problem s each with their individual task abilities or reasoning abilities planning abilities so we can Envision these systems but right now we're really at the early stages of let's just kind of get the get the basic agents right to do simple things but do them in a more reasoned way and I think we have the the technology in place to do that through llms and especially through uh through the tools that they might use to do that so so what's new about agents today and why they're such a Hot Topic is that we've all seen the excitement around what llms can do so what you know geni can do but that's all really in a natural language context so you can go to chat GPT and say I have a cat here's the cat's name write me a Shakespearean Sonet about the cat right does great job with that but that's just language for practical problem solving we want them to get more into that automation space of doing either doing useful tasks on a personal level or in a business context and the way that agents are being built and there there is sort of a general architecture that's being rolled out there around agents they have the llms that they use but in addition to the llms they have data that grounds the llm so you can give more context to what the agent is capable of they've got the ability to define a real clear scope for what an agent is supposed to do I mean an LM can do anything but if you make an agent that's totally focused on let's say onboarding new employees in your organization you know that's that's that's a useful domain and functionality and then the tools that it needs to do the job and I think today agents are the way that you can get the llms to actually do things that add value I think that's from a practical standpoint that's the the reason why they're so exciting so you can sort of think about agents as The Meta mind of an llm and I guess to follow up on that um I'm think there's a startup whose Name Escapes me at the moment and they're trying to build this agent that will do a number of tasks for you all strung together like you want to buy a new refrigerator this thing will shop for the refrigerator it'll get the best deal it'll actually oversee delivery to your house and a few other things maybe it'll cook an egg I don't know um so the thing about this particular agent is that it's got to make some value decisions like well wait I mean delivery to the house should I pick a Tuesday or a Thursday do I know my owner well enough so it's it seems like a real agent has some true reasoning capabilities built into yet yeah I mean and and again I would say that's probably a vision statement right because I think that there's we're really at the art of the possible stage with geni you know overall I think um I think we still there's going to be a reckoning coming with Chi when people or instead of saying hey you could do this with Gen you could do that with geni I think we'll see like in that example the number of times you might require more information from your your the agent's client or user is there's probably quite a few human in the the loop steps there that would be required at this point right because they're only as good as the data that they have and and and only as good as the Precision and contextualization of the data they have right um there's also the cost implications of like it scale if you're a company and you run a bazillion llm calls that's not cheap and then the things that go along that the time response time environmental impact even like there's all that stuff so there would be a right I think that's what you just described that company is striving to do certainly is on the on the horizon that we should be aiming for but I think that even today we're still just getting the basics of like I think I think a really practical step we could take is today a lot of the automation that we do in the technology world is very strict it's very static we statically Define we might have conditional statements or rules rules-based logic that gets into okay if it it's either a b or c m what the agent Frameworks allow you to do is now just instead of um having more rigid decisions you can case statements like you can get into more sophisticated and dynamic routing logic about given a particular context given the set of things that you're able to do tell me what the right thing is to do next or the right sequence of things to do next which is much more Dynamic it may not be as mindbending and and science fictiony as some of the scenarios that are out there but it's it's pretty practical well it would be it would be labor saving so I guess that that brings us to the the question you've talked about this what are the true strengths of Agents specifically how do agents collaborate to perform complex functions in sort of a first of its kind kind of way yeah I think I think this is where um you know what I just described kind of fits fits this bill and what I'm seeing in practice right now and even this is sort of evident in even the tooling that's out there to build agents as much as we talk about agents working with other agents we're still just trying to get that that basic uh setup right now where an agent has a set of tools but um so so that's kind of the state of the industry right now if you're building agents is it assumes that this agent you just give it these tools it's going to do the job and and that's kind of the Practical imp implementations right but if you think about it from the perspective of of let's say you build a software component and you give it a tool to what was the the example you just gave there about buying the refrigerator right you might give it a tool to go or a set of tools to go and look through different retailers to find what the fridges are and pull down the specs and so on you might give it a tool to actually make the purchase and get your payment information and you might give a tool to track shipping or something you know like those are all different tools that you provide and you could see where you could string together you could build an agent that would you know understand the user preferences come and do that all those things mhm it's not a big leap though from the perspective of the agent that's looking at these tools to say an agent can be a tool right so I think I think this is where I think we'll kind of Step our way towards this future where because you're not really you don't care if it's a tool or if it's a person doing the job for you right you what you care about is I'm delegating a task right you delegate a task to a tool or you could delegate a task to to another agent so I think once we get this like I think the seeds are already planted there for how we get to that multi-agent future but today I think we're still kind of at the crawl stage of how we actually build these practical agents well so you're actually envisioning a world in which uh agents might be managing agents am I am I correct in thinking that well I think I or agents are building agents perhaps agents are building agents agents are calling agents I mean there there are people like there's there's lots of research going on and experimentation around how you know you could give an agent one tool and that tool would to build tools that so there's these meta scenarios that are out there but I think um when we think about how these things actually scale out practical terms um we we sometimes we need to take a little bit of time we need to let things play out and I think it's where I expect that even though we know that we're going to get to a point of total autonomy and meta autonomy around like you said autonomous systems that manage autonomous agents and so on right I think um the uh we're ways from that but that's okay I think you know what we're spending a lot of our focus on is today I think a lot of the agent um a lot of the agent technology that's out there is very self-contained like even some of these systems where you have multi-agents you have to run all the agents in a very proprietary contained way we know that the future having worked in the Enterprise space for a long time you're going to need a lot of interoperability between these agents and that's something that's not really clear out there so I think um we're focusing on how do you get groups of not just groups of Agents but groups of agents that might live you know some of over this hyperscaler cloud and this Cloud over here and you've running these proprietary agents and python Frameworks how do you get those agents working together and I think what I was just describing about about agents looking at other agents as if they're just more tools is a way to way to get there and my my background is heavily in the uh in the API space and you know we had a we had our big boomy world event in May was my that was my big py tagline was to say the AI economy is an API economy I mean we've spent a lot of time in the last more than a decade turning core software capabilities inside organizations into API so that they can be called that's really the essence of how software functionality becomes tools and I think as we API enable these agents that we're building that's as well how how things are going to all sort of be orchestrated and interoperate that is a fascinating angle and I know we could really get down to the Weeds on that one but it's really we are living in an API economy and things are calling other things and apis are going Every Which Way um that's not a question I just think it's as fascinating I I want to make sure we talk about uh real world implementation how do you see customers Implement AI agents what are the challenges in particular the major headaches we've already touched on one of the big I mean one of the big ones is just like this is great we have technology that can act on its own and then immediately this is dangerous we have technology that can act on its own right um so I think that uh there's a there's a lot of you have to be really careful about hallucination because llms are by Design Crea they're creative in a way and they like they don't like to tell you they don't know the answer they're a lot like humans right right right an ego aot is like an egoistic human yeah exactly like I'm not gonna I'm not going to admit I don't know the answer I'm just going to make it up we've seen we've seen that before um even just auditability and explainability like who did what I think that this is one of the again looking at that interoperability uh with that interoperability mindset where we live you know with Boomie we we're connecting to all sorts of different systems and our role in this AI agent ecosystem is to be like a tool provider as well as you know just kind of helping facilitate communication through these agents you get some pretty complex authorization scenarios there like systems need to be authorized to talk to other systems but at the same time a lot of the permissions on four systems are based on user identities and making sure that people have the right privileges for that in an aged world how does that work are you going to issue credentials for the agent to to like to act actually like have a have an active directory entry for an agent or are you going to give the responsibility for all the actions of that agent to whoever the end user is right so there's some pretty complicated authorization considerations and I think those are the types of next order problems that um organizations will will hit as as they get deeper into the complexities well it seems like they the agent would need authorization to go at a certain up to a certain level uh you couldn't rely on the end user you the agent would need to have the ability to have pretty pretty fair amount of privileges wouldn't they yeah and this is this is where again though I think the experience in the API space helps because we've had um we we've run into similar problems before not necessarily with autonomous software components but more when you've got an end user who's using a system that ultimately you're going to end up in some backend system you don't want to just have your propagate your credentials all the way through so we got token mechanisms and to do that and we've got even ways of of authorizing like the application to talk to the application but then also propagating the identity of the end user so I think it'll become a combination of that right now where I would be leaning is you we're not I don't think we're ready yet for I think there was what was the quote from IBM in 1979 a paper saying something like you know you can't hold machines accountable for anything so you can't hold computers accountable for anything so you can't rely on them to make important business decisions they're just not accountable so I think the accountability at this point at least we still need to attribute actions to whoever the end user was that triggered it so in the case of the ordering the refrigerator if there's some screw up along the way there you know you can't you're you can't really blame the agent you're going to have to put it back on whoever issued the order to to get true um you know I want to ask about the the future of AI agents and Enterprise and I you sing you speak it brings up the idea that there's a lot of complexity going on here I mean it has always been really complex when you think about an agent having certain amount of privileges certain credibility know credentials and then that brings in the issue of cyber security for the agents are they secure can they be hacked and what a system is we're building some systems that are like wildly complex sometimes I think the the complexity of systems May collapse unto themselves but that's let's maybe let's not go there um but we are we are kind of there right like I think yeah we are we are anyone working in the Enterprise knows that it's never like a one system failure that causes the biggest problems it's a little it's a little blip somewhere in a highly connected system that ends up with this sort of progressive meltdown S I mean this is just one of those uh things where we software vendors we're we're guilty ourselves of coming in and saying simplify simplify you know I think that there this is something we've done recently like we're just embrace the complexity like there's no way around it I think that complexity gets a bad name so I've been talking about complexity for a long time there's a great paper from mid 80s uh by Brooks on there's no silver bullets right how we're always saying it's complexity is actually a great thing like the world is built on complexity biological systems are complex systems computer systems are complex systems we solve really complex problems like there's the incredible things that we can do with computers is a result of our understanding of complexity but you have to deal with complexity in the right way you almost have to dance with complexity in the right way and I think that agents like many other um many other components or or a agentic architectures like other architectural practice has to be it's it's another modular architecture approach to software systems and just like SOA 20 years ago and microservices launched 10 years ago like we all have to recognize there's ways of dealing with these complex systems so I think if you respect the complexity and understand that these agents are just more components that are going to fit into very complex landscape and you're honest with yourself about it you'll do better than assuming that you can just gloss everything over with one agent that's going to hide you know all the complexity goes away it never goes away but you can provide good levels of abstraction to help you deal with it more effectively yeah it's a good point we'll never get away from the complexity we're gonna have to deal with it and and there it is and as you said you know nature systems itself is highly complex so there you go uh I want to make sure I get your thoughts on the future of AI agents and Enterprise where do you see us going in the next say 1 to three years uh I want to be able to say to you my bot will call your Bot you know that will probably happen what what do you see in the immediate future oh yeah well I think there's another I keep I keep bringing these quotes up there's another quote about what we we always overstate what we think is going to happen in the next couple of years and understate what happens in the next 10 years so it's hard to put a timeline on it I think that but I think the reason for that um you know I just Co co-authored a book that was that was put out recently called un Enterprise we get into this a little bit it's hard to sort of predict the the future because there's so many pieces that have to come together so I think understanding the like the authorization stuff I said the API stuff like there's all these schools of thought that have to converge around AI so we can see what the what the soup is that where we that we bring all these ingredients but I I think that in the Enterprise space there's a lot of experimentation going on I think that we will go through phases of the ual happy path will be useful maybe simple but useful agents that can function together will probably have some more inside the Enterprise multi-agent systems to begin with from a from a ecosystem standpoint you will see these Frameworks of hey you can if you run everything on this platform then you can do these agents working together but I think will it happen in three years I hope so I think that we'll have to break through that and and like you said start to get into more of a of an ecosystem opportunity where if I'm a company that's built some fantastic agent that's really good at a certain job then how do I offer that out so that it can be consumed by these other other companies and Enterprises and as an Enterprise how do I how do I go get best to breed technology that's really the big push that that that we're helping with and I do think I do think that's where ultimately things will go how fast we get there we'll see but we have paved the road with apis already and I think that will be the path to getting this interoperability of Agents no doubt about the apis I the road has been paided by apis um man I think you said it fascinating stuff re really really interesting um thank you for sharing your expertise and please come back and talk with us again sometime I would love to thanks for the opportunity

This transcript was generated automatically from the video's captions and may contain errors.