Greetings. I'm Josh Tyson, co-author of Age of Invisible Machines, the first best-selling book about conversational AI. Age of Invisible Machines explores the learnings of industry veteran and OneReach.ai CEO Rob Wilson. Each week, Rob and I bring in a guest to continue the conversations we started in the pages of our book. Last week on the Invisible Machines podcast, we talked a lot about AI agents, and this week we're going to show you AI agents in action. You'll see multiple agents swarm around a complex scheduling automation.
and then we'll give you a look at how these agents work together behind the scenes. Our guest is Annie Harshberger, lead experience architect at OneReach.ai. We've talked about Annie on previous episodes because her journey is unique and informative. She came to OneReach.ai with a biology degree and limited technical expertise. She was, however, an expert problem solver, and using the know and low-code tools on OneReach.ai's Generative Studio X platform, Annie has quickly become one of the most prolific automation architects in the world. On this episode, she's going to share a live demo of AI agents, so I recommend watching on the Invisible Machines YouTube channel.
The idea for this demo came from our conversation with Paul English, which kicked off season three of the podcast. Here's Paul describing the use case in that episode. I am very interested in having each one of us have an agent that can negotiate on our behalf. Probably the simplest example that's often cited, but I've yet to see a solution to this. I'm thinking about building it myself, is if the three of us are trying to meet, the way it works today, even with Calendly and Google Calendar, that you can say, propose a time.
we still have to look at some options and choose them. And then we have to accept it into our calendar. If we all had agents like AI assistants, we just have the three of them talk to each other. They'll pick the time and put it in our calendar. And we've never even looked at the options. And I haven't seen anyone do that yet. I'm thinking about building it myself. All right. Now that we've heard the setup from Paul, let's jump right into this eye-opening demonstration of AI agents with real agency.
Invisible Machines is produced in partnership with OneReach.ai. Their Generative Studio X platform is the only orchestration platform that's been named a leader by every major analyst group. Enterprises and other organizations are already using GSX to create technology ecosystems where they can grow their own digital teammates called Intelligent Digital Workers. These IDWs can be set to work with hundreds of customizable skills utilizing GSX's no-code building tools. Head to OneReach.ai to test drive an IDW and experience the next phase of Generative AI. Annie, thank you so much for joining us on the podcast. We're excited to finally have you here with us.
Yeah, I'm excited to be here. Thanks for having me. Yeah, and Rob, a pleasure as always seeing you here. Familiar faces today. Yeah, you know, and familiar discussion too, I guess, because we've been talking a lot about AI agents. We really wanted to demonstrate kind of what an agent with actual agency looks like. And as I mentioned, we had this kind of cool use case provided by a conversation with Paul English. So we brought Annie here today to kind of show us what that solution might look like. It's a funny thing. I've had several conversations, like many conversations in the last couple of months, but definitely in the last few weeks.
People asking about agents, super excited about agents, but it becomes apparent very quickly in the conversation that they don't really know what they are. There's some puzzle in my brain. What are they excited about exactly? I know what I'm excited about. I don't know what they're excited about. It doesn't seem like it's a bit of an ambiguous topic. I understand why it is complicated and it's a whole new way of thinking about machines and software. But I think it does make sense that maybe we dig into this in detail so people can understand what they're excited about.
Yeah, so today we'll be going through an experience conversation with this agent set up for a meeting schedule, and that'll include coordinating for different calendars for the participants, sending out notifications, collecting all of the required information for this meeting, and actually creating that calendar invite. Before we jump into the demo, could we talk a little bit about your role with OneReach.ai and some of the solutions that you build? Yeah, definitely. So I've been here for about four years now. It's been a great time. And I lead a team of builders who create and build demos in our platform for existing customers and prospective customers, just to show all of our capabilities and make sure that we're getting people what they need.
What was your thoughts when you first started seeing how agents work in terms of how that frames up old software? It is different, right? Yeah, definitely different. It's really just interesting to see how all of those different agents are able to coordinate together and how they can all be available through the one entry point. Exactly, yeah. Yeah, I guess it's one thing to think of. And I think a lot of folks are just there at the single agent concept, like one agent that can complete a task, like we set a password.
Once you realize these agents can communicate with each other and then they can work together to get work done, it sort of just gets super complicated, but also super exciting because you're like, whoa, this is kind of how we work, right? We have all these patterns and skills we know how to do that we've either trained on or we learn in school or whatever. And then we sequence them to get our jobs done. And so now we see this. almost exact same pattern with machines where you start adding these skills and it's able to sequence them to get work done. And then you re-sequence them to get different tasks done because you're starting to give objectives as we've talked about in the last episode versus like instructions. So cool. Let's get into like, I guess, giving people a sense. Is this some ball type bot or a single agent approach?
This is a multi. So we'll, I think, see at least three or four different agents throughout that scheduling process that, like we said, all coordinate together to actually get that meeting scheduled, send the notifications and meeting reminders and collect voting results. How did you like think about breaking up? Like you broke it into three agents. Why three? Like what, what made you like, how did you think about compartmentalizing these, you know, when should I create a new agent? When, when should I just add more options and tasks within a single agent? Like why break it up? And, and how did you think about like how to compartmentalize this different task?
Yeah, so I think the way that we're thinking about it, as well as the broader team who worked on this example and frameworks like this, is looking at those agents in terms of reusability and how certain pieces of that experience will be used, not just when you're scheduling a meeting, but in other skills or I guess completing tasks as well. For example, collecting a poll and voting results on which time works best for everybody and sending out a list of possible participants on Slack. There are a lot of those things that can be used in a lot of other places besides just scheduling a meeting. So I think that's the big thing that we look at when we break them down like that. So it's like a looking forward in a way, like what are some of the other tasks and how do we break these down in it? Not too granular and not too atomic in nature. So interesting. All right, let's jump in and that's jibber jabber for me. Okay. So here is the beginning of our chat for this experience. In this case, we're showing this in rich web chat and this will allow us to access any of the agents that are available. So through this one entry point starting here. So at the beginning, we get a little bit of an intro. And sign in is key, this authentication piece, because that's like structuring who can do what.
identifies the user. So the agent isn't just executing the task. It's, it's executing in context of knowing who it's executing it for, which is a really big piece of, you know, of the puzzle, not just do you have the right to execute this task, but more like the agents actually know who they're doing it for and can have context about what that person's done in the past. And, you know, that's kind of key, isn't it? Yeah, yeah, absolutely. One of the other agents we also have as a part of this experience is looking to see previous meetings that have been scheduled and whether they occurred, we can edit all of those meetings. Yeah, like you're saying, all of that previous conversation, history, context, meeting information is all specific to the user that's logged in. And I think it's important to point out that the authentication process itself is not...
But that's just good old Google authentication, symbolic code. And actually, it doesn't make sense to make that an agent. You're just adding extra costs and LLM executions. So this part's not actually going through the LLM. The authentication is just a separate piece of the solution that exists outside of any LLM interaction. Right. Yep, exactly.
Perfect. Okay. So at this point, we've logged in. We know who I am and we can use that information later on in scheduling meetings and understanding what information or agents skills I have access to. So here, how can I assist you today? I'm going to say, I need to schedule a meeting. Oh, here we can see at the bottom, the agents that are used throughout this entire conversation. Gen Y or Gen X emojis? I can't know. I don't know. I had a past episode. I don't know. We, we discussed emojis and different generations use different emojis for the same thing.
And you can actually, yeah, just through emojis, you can get a good sense of what generation they are, but which one they use for things like thank you. Yeah, we learned that thumbs up emoji can be kind of divisive. Yes, yes. Oh, for sure. Yeah, that's a problematic one. Yeah, thumbs up is sarcastic in a younger generation. Yeah, I agree. Yeah, anytime I thumbs up you guys, you know there's something going on. Yeah, we're like, oh great, she's... Noted. She's making fun of us. Exactly, exactly.
Okay, so what we've seen here, we saw a few of those loaders at the bottom. So like I said, that is the agent that is working in the background or that we're conversing with at the moment. So based on this phrase here, I need to schedule a meeting. We understand that we're going to the collect meeting details agent, which will ask us a series of questions to get different details about the meeting. So here we're starting with subject, agenda, expected outcomes, sort of standard things that you would see for a meeting. So now I can reply, let's say, sure. Subject is... And then Rob or Josh, you want to give me a subject to include here for an example? Oh, um...
We need to do a prep call to discuss an upcoming guest. Okay, prep call podcast. Yes. Okay, perfect. So we've given the subject. As we can see, it's asking me for more information, though, that I haven't provided. So we'll be able to converse, go back and forth as many times as we need until it's collected all of the required information from me. So, okay, we got that subject. What's the agenda and the expected outcomes? The outcome is to, ironically, I think, to have an agenda and a scheduled time for the actual podcast.
To align on an agenda and agree on a date and time to record. Unless you can think of anything else, Josh. Those are the things we usually need to do. Perfect. All right. So we have that outcome. Then the agenda for our prep call. Review the agenda for the podcast and identify any...
hurdles in terms of clearances. A lot of times they have to get clearances from organizations or topics that we have to avoid, things like that. We'll share notes too. And then create any action items is the objective. Great. Okay. So once again, as we can see at the bottom, it's telling us which agent is working in the background. Oh, that's super handy. Yeah, absolutely. Because there will be a few that we'll see throughout this experience. So for now, our meeting details collector has gotten all of that information from us. So we have the subject, agenda, and expected outcome. And then its next objective, like you were mentioning, Rob, is getting the meeting participants. So who do we want on here? We have a whole list of everyone in our Slack. So we can add a lot of people. One thing I guess that's important to point out are two things. The first thing is that those agents...
could be a swarm. There could be, even though it says it looks like one agent is exiting it, it could be like 10 or a hundred threads, the same agent. Like if there's multiple tasks to do, like if it was send out meeting invites, that agent could be running, you know, in parallel, like five times, sending an invite to different people and handling it. So it's, I guess there's two parts to the swarm. There's different agents working together to get a task done, but there's also the same agent.
threading out and multiplying itself to also get work done. And then the second thing I wanted to point out is this dropdown here and the participants were like back to an integration and symbolic, like the LLM is not providing this list of folks. This is something that's getting extracted from more of a traditional data storage system. Yeah. It feels kind of like familiar, like a chatbot experience people have had. It's so many more layers happening. It's just significantly different. Yeah. That this is being conjured on the fly and in context versus like a rigid if then statement all the way through, but it's still not going through an LLM to gather the list of possible participants. Right. Exactly. Perfect. All right. So let's add a few of our guests here. We'll add both of you.
You're a mandatory. So we'll select those. And we can also select for optional meeting participants. We can type here to skip that. Let's let's add somebody else here. We can bring Jonathan Anderson and Elias Parker. Perfect. Yeah. Jonathan. Beautiful. He won't be there. Of course he would. Yeah.
Okay, perfect. So we provided those meeting participants. That was one of our objectives. So we need to know who we're sending this meeting invite to, who we're asking for availability. And then the next thing that'll happen here is it'll give us a summary of everything that we've provided so far. So we have an ID already created for this meeting. We have a summary of our subject agenda outcomes that we provided. And then this is the status of this meeting at this point in time. So next, it's going to ask us for time slots. And it'll also give us our participants that we've added so far. Now we're moving on to another agent. And this now gets stored in context so that for future meetings, again, like you said, we're able to pull this in as a reference point.
Yep. Exactly. Exactly. Yes. Great call out. Okay. So what we saw at the bottom there, we saw time sinker. So this is another agent from our meeting details collector. And then there are a few things, a few objectives that this one has, the duration, the number of time slots we want to offer, and then it'll collect some suggestions from us. So duration of the meeting. Anyone have thoughts on how long this will take? 30, 45 minutes? About that. Okay. Perfect. 30 minutes. That'd be enough. Perfect.
So now it's going to time thinker, which was a different agent there. I could see that. Yep, exactly. And then next thing here. So we have 30 minutes. Now, how many time slots do we want to choose from? So this is saying here we have three options maybe for which meeting slot we want to choose. How many do we want? CQ. Okay. So now that we've chosen how many time slots we want to offer, at this point, it's going to give us a choice. So a date and time picker for which slots we want to be available. So we can select those times here. That'll pull up our date and time chooser. Any preferences from you guys on when we want this meeting to be scheduled?
Fridays are usually good for us. Fridays. Okay. Perfect. So we'll look at Friday. Let's try Friday morning. I think it's also important. Like I didn't have to open up my calendar. Exactly. Just in natural language. It makes sense that it gives me a calendar. This graphical micro UI that's embedded. It is a better experience than just natural language. Yeah. It's an important example of like extra linguistic components. Right. Talk about those a lot, but it's cool to actually just see looking at one right here and see why it matters so much that it can create these components on the fly instead of it just being all, you know, words and letters.
Yeah. And it's cool being able to really see the agents swarming on this activity, essentially, right? Like we want to plan a meeting. There's a few different agents that can pick up different aspects of what we need done. Yep. Exactly. Yeah. And like you're saying, so we have another agent here now, this notification center. So we have collected everything we need for our time sinker. We have the time slots and how long this meeting should be. Now we are at our notification center. So this is what is a message that we are sending to the participants to get their vote on which of these times works better. So it's going to give us a proposed message. We can change that if we want to, but we can also say, yep, this is good. That works.
It looks good to me. Right? Yeah, we could have, I included, asked it to include like, hey, make sure they bring, you know. Pineapple? You know, their lawyer with them or whatever. Or legal representation. Okay, perfect. So last thing here is how much time are we giving the participants to vote for their preferred time slots? So within 10, 15, two hours, that's when the voting will end and then their vote won't count towards the results anymore. Let's make it super short just for the demo. Okay, perfect. Let's do.
three minutes all right so we'll submit that here and then the next thing basically means like you know people submit get back and then anybody doesn't respond if they're not mandatory that the slot will be chosen and um and they'll either you know come or not come yep exactly exactly okay so last thing here should i proceed go for it then we'll see that notification center working to send notifications as it is so all of those participants Yep. Exactly. So then what's happening in the background? Like you were mentioning. So I think we've seen like what, four agents so far? Yes. Yeah. So we've notified all the participants and then I'll pull up my Slack in just a minute here. It's sent me a notification. Rob, Josh, you guys might've seen yours as well. So it's your mandatory participants. So once we notify everyone. Yep. Yeah. Let me pull up my Slack here. Oh, it's in Slack. Okay. I went to email. Oh yeah. Meet Sync Pro. Just reached out to me. Got it. Yep. Perfect. Okay. All right. Are you guys seeing my Slack? Okay.
Yes. Okay. Yep. Perfect. Okay. So in this notification, it'll give us all the information that we just provided as part of that series of agents that are helping us schedule this meeting. So we have the subject, agenda outcomes, the mandatory and optional participants. I'm the meeting organizer, since I'm the one that was logged in and we've suggested two of our time slots. So here we can choose in a dropdown, one or multiple that we're going to say both and confirm. Then it'll record my votes and we'll be able to see how many people are available for which time slots. So we know which one to pick. And once we do that, we'll say, all right, thanks.
wishing you a lovely day. And then once our voting time has ended, we'll get a notification when the poll closes, essentially. It's like, you know, you think about two or three people and it's nice to have, but you think about like 30 and you're like, this is a game changer. Yeah, absolutely. Okay. So since we've notified all of these participants now, it'll also allow us to set up reminders at the start or before this meeting. And this one specifically is, let me know when everybody has finished voting. So even if that's before the five minutes are up, say, let's do it. Yep. Imagine you're like a teacher and you have to like set up parent teacher conferences.
with all your students, you know? What a mess today. And then in this case, you just set out an agent and it's going to schedule them all for you, get it all lined up. And you just watch your calendar just pop, pop, pop, pop, pop, pop, pop, pop, just filling in. That's, yeah, mind boggling. It's cool. Definitely becomes really powerful.
Yeah, seeing that micro UI too, like embedded in the chat is cool because I'm thinking about, you said scheduling parent-teacher conferences. Like you usually get some sort of strange third-party scheduling thing that you have to sign in through Google and then hunt around to find the time slot you want. And never quite sure if you reserved it properly. So this kind of eliminates a lot of the gray area. I think where it also gets even cooler is when you start looking at agents for like filling slots. So somebody cancels, you know, especially if it's like a doctor's office, you think of how many canceled appointments they have. And, you know, if you could just collect up front.
anybody who would be open to an earlier appointment. Or you can see their second or third slot. And if somebody cancels now, you can go, an agent can go back to somebody who that was their second choice or even their first choice. And then offer say, hey, you know, a spot opened up and that can like fill up empty spots. And you don't have a human, you know, having to manage like cancellations and then calling, you know, random people to figure out if they want to take the spot. I mean, usually those spots just go unfilled. You think of the cost globally of every appointment that gets unfilled because there's no time to find last minute.
folks that were willing to reschedule and that agents can now solve that. You just look at that one little microeconomic problem and what that can do to an entire economy. And you just go, that's just one small little use case agents can be deployed against that could move the needle entirely at a large scale. I think sometimes we think really big, like coming up with the next. you know, innovation to solve global warming. And although I hope that does happen very soon, some of these are just pretty small, narrow use cases that if applied to enough, at scale, at large quantities, it could move the needle in a significant way. Yeah. I think even for like small businesses, you know, I always like to bring it back to hair salons, but like my wife has a wait list of clients. If someone canceled an appointment while she was with a client, you know, she's not going to have time.
to reach out to all the clients that are on that wait list. So if the minute an appointment gets canceled, yeah, I can just go through her list. And by the time she realizes the appointment is canceled, it's already been filled. Yeah, it's like something that's sitting there playing a Tetris game. It's just keeping the slats filled in front of you, all the blocks, so that you just have a solid day. So yeah, I guess there's a downside to that where...
Maybe some of those canceled appointments are a respite for a break. So the AI's going to have to be smart enough to keep a couple of empty slots in there for her to catch her breath. But hey, at least it's designed. You know, that's better. Yeah, I can check her and know what to do. Sorry, Annie, keep going. Let's see how these things are going. No, that was great. What's happening in the kitchen? Yeah, so in the last couple minutes here, the voting has closed and it sent me a notification in Slack that the voting closed. So let me pull that up again here.
Go. All right. So the new message that we're seeing here heads up voting time for this meeting has closed. You can check out the results of the vote. And then with this link, then it'll let us see how many people have voted on which times that were provided on it. So it can now just pick a time. It doesn't, or, or even in the loop, or, or it could just provide the vote results and let somebody override. In this case, is this a human in the loop? Someone gets it inside or is it just going to decide on its own the best time?
I think we have both options, actually. Yeah, let's check. So how can we help? There's like a timeout. Like, you can override it, but if you don't, I'm just going to go with the best. Great.
Okay, so here I can say, show me the voting results for my last meeting. Once again, we're starting with the schedule coordinator and then it'll be able to pull all of the meetings I've scheduled recently to show me how many have which status and which have been scheduled. And this one is a separate agent, participants response checker, which we haven't seen yet. And then this will allow us to see who all responded to the options for that meeting. So I could create an agent from here. Once I have enough data of scheduled meetings and votes, I can start to see patterns and predict.
best times for certain participants based on feeding that historical data in there. So now we can sort of predict what time is going to work best for everybody based on past voting behavior, which now we get into the prediction side, which is something Paul English in that episode talked about. I wanted to propose a time that's best for everybody before they vote. And so now we know we're beginning that process of having the information and the data to start moving from asking to predicting, which is, we know. The ultimate user experience is not when we have to ask it, but when it can predict for us what we're going to want. So that's cool. That would just be another agent. Like you said, that just gets into more processing old voting data. It's a pretty straightforward algorithm once you have the data. Okay. Yeah. Well, like starting internally on something like that too, I would think you're afforded more flexibility to make those predictions because you have access to kind of standardized scheduling software.
Yep, exactly. Exactly. So now that we can see here the voting result, so we know what the last meeting is based on who I am and what I've just done in my last conversation, then it's going to generate this grass for us. We've got one for 10 a.m., three for 12 p.m., so that's the one that works best. We can see all of that described here. Then we can see that's also been sent to my email. So let me see if I can pull that up here. And you mentioned too, like, how one of the virtues of kind of separating out different agents for some of these tasks is that reusability component. And you can see it with something like...
this agent that collects voting results, right? Like there's all sorts of things beyond scheduling that you could, you could be voting on within an organization or with customers and that could be a useful agent. Yeah. Yeah, absolutely. And it's recursive in that if it didn't come to a conclusion, it can go back and, you know, re-ask for more optional dates and times like, nope, this isn't good for anybody. Let's try some other dates. And the agent is able to assess that on its own. You're not building in that logic of, you know, it's just, it's just understanding like, oh, I need to go back to this agent and, and run this again. So that.
That reasoning concept now we see in action where the programmer isn't explicitly programming out all of the exceptions. You have just a reasoning engine that's just going to predict or assume the next function or the next agent to call and then run through until it gets its outcome. So it's pretty doggish. It's sort of like playing AlphaGo. It's going to keep rolling until it gets to this outcome that you said.
So yeah, let me go ahead and share how this is all built in the background. So let me pull up my brain. So, all right, what we're looking at here is the IDW with all of the agents that have been used throughout that scheduling experience. So you can see each of these are marked here as a sub-agent. These are all the things that was used as part of that process. We have five agents that have been involved. Yeah, so we can click into this one here. So our schedule coordinator, this is that main agent that's taking in any of those incoming chats.
The first thing we're doing here, this is our greeting message. Let me switch this into edit mode so we can see it a bit better. This is that first message that we saw. So we show this is not going through an LLM. This is, like, no need to take the cost. No need to, I mean, it's just an intro message. It's going to stay the same. There's no reason to run it through an LLM. Exactly. Yeah. And like you were saying, same thing with our login component here. So this is just allowing us to use Google to log in that way and then collect information, understand who this person is and what they have access to and what meetings they scheduled and all those good things. Right.
We're saying, all right, how can I assist you today? So this login name here, this is whoever is logged in. So we'll have information about this person to use throughout the rest of that conversation. And then here is where we're using the caretaker execute step with a list of objectives of what this specific agent does and how it does that. So there's our agent step right there. There's the agent. Exactly. And there's a notebook, instructions of what the agent...
instructed to do and its options, which then are symbolic on the legs. And I guess it's important for folks to know that this can also have dynamic element within it. In this case, it looks static, but there could be dynamic components and pieces of data that could instruct the agent to do things in a dynamic fashion. Yep, exactly. And I think we'll see an example of that in the next agent we go through for collecting the meeting details or getting the time slots that are available and those types of things. Got it. Okay. Yep.
Yep. So here we're listing out all of the objectives in natural language, just telling it how to behave, what it does, important notes that these are maybe things that we shouldn't do or that we absolutely should do. And then finally, we're giving it an example scenario of how to create a new leading and what that process looks like. Got it. So it's objective. We're giving it an objective. We're constraining it to doing, it can't just...
go off the rails and do anything. We're constraining the number of options of ways or tasks or subtasks it can use to get things done. And then we're also giving it some example, let's call it compared to coding, loose instruction on how we wanted to accomplish this so that it does it our way, not its own way. And so we have that combination like you would a person, you're giving them objective, you're giving them constraints, and you're giving them thumb instruction. And we see that in comparison to coding, it would just be all instruction.
very strict rules on what to do. Ignoring objective almost entirely and constraints are just inherent in that. So this is like a whole new way of coding to say, here's your objective. Now, here's the ways you can solve it. Here's the small tasks you can use to get it done. Now, it's like Lego blocks. Here's the Lego blocks to build it. And now try to build an appointment with these blocks.
Yep, exactly. So this is where we're identifying another agent that can complete a task. So for example, when I said I need help scheduling a meeting, we understand that we want to start with that meeting details collector agent. So again, a separate agent that can help me with that. So this option here is just searching for available agent that can help with that task. And then once we find one, that's where we're collaborating. So we'll see this trigger step here. And that's based on a description that each agent has of what they're capable of doing. So it's basically like...
an LLM with a vector looking for agents that fit the description of the task and then calling that agent and passing the task off to it. Exactly. If it's helpful, we can see an example of how those descriptions are set up for each agent. Yeah, that'd be great. Okay, perfect. Yep, let me share that here.
All right. So this is our collection in Lookup that we're using to identify these agents. So we have a document here with a list of all of the agents and what they do. So here we'll see all of the ones that were used throughout that process of creating a meeting. The first one we saw is this meeting details collector, and it'll give you a description here of what this agent does, what it's expecting to collect throughout that conversation and anything else we need. So the more generic this description is, the more likely this agent would be used in other workflows. And then the more specific this description is, the less likely.
other workflows would use it. So getting this right is probably a bit of a tuning thing, huh? Yep, exactly. Yeah, like Reminder is probably a fairly popular agent. Yeah. Yeah. TimeSyncer is a little more niche. And if I understand how you implemented this, even if it passes it to an agent incorrectly, all that essentially happens is that agent just then...
says, oh, I can't. It doesn't necessarily reject the task. It still takes the task, but then immediately goes to collaborate and find another agent. So it's sort of a passing it forward. It doesn't necessarily go back. So it's a bit of a combination of centralized and decentralized in that it passes that task on to another agent. If it turns out that that agent can't execute that task, it'll then go ahead and look for another agent. And then, you know, of course, that could go on a while, depending on how many agents you have. Yep, exactly.
Yeah. Yeah. So like you're saying, these search agent and collaborate options we are using in the other agents as well. So yes, if there's a point in time where either we got there by mistake or we're trying to do something else throughout our meeting scheduling process, that is an option. And that's also how we're able to, like you're saying, sort of pass that off to the next agent that's needed as part of this process. So our time slots, our reminders, and anything else that's included as part of that meeting scheduler. So that's nice, like in the hallucination concept that if it gets the wrong agent, it's not the end of the world. It doesn't just break. It's not brittle. It'll sort of continue to look for the right agent. Yep.
Exactly. Yeah. And part of this collaboration will also include, I think you mentioned this earlier, Rob, passing off helpful or relevant information to that next agent that we've already collected as part of this conversation. So for example, like I said, I need to schedule a meeting for the podcast prep session where our outcomes are this, this, and this. It's going to be able to extract that information and pass it off to our meeting details collector. So like you were saying, it's not going to ask me again, okay, what's the title of your meeting? Since we've already provided that information. Right, right. Yeah, that makes sense.
Yeah, it's fascinating to see this level of nuance because I think a lot of how AI agents are being represented out in the world seems more like, you know, just generative AI being kind of hurled at a problem or at a use case. And this really shows the nuance involved. Like generative tools, LLMs, they're a part of this, but they're used rather sparingly. Yeah. Well, anyway, thank you so much for coming on and sharing this demonstration with us. We've really been wanting to show AI agents with real agency. So this has been an absolute delight. Yeah, practical use case. I think it's... Yeah. Yeah, absolutely. Thanks, guys. Thanks.
All right. Well, thanks again for joining this ongoing conversation about conversational AI. Be sure to subscribe to Invisible Machines wherever you get your podcasts. And if you feel so inclined, please leave us a review on the Apple Podcast app. You can also watch these conversations on the Invisible Machines YouTube channel. This podcast is produced for UX Magazine in partnership with OneReach.ai. Over the past five years, our team of nearly 200 engineers, scientists, experienced designers, anthropologists, and linguists have been developing Generative Studio X, an award-winning platform that has the lone distinction of being named a leader by every major analyst group.
GSX is a complete environment for hosting, creating, analyzing, and scaling your own digital teammates called Intelligent Digital Workers. For an interactive demo of IDWs in action and to learn more about the GSX platform, head to OneReach.ai. This podcast would not be possible without the hard work and dedication of executive producer Elias Parker and producer Kate Timchenko. Our video and audio editor is Michael Lifinov, and we rely on support from the marketing team at OneReach.ai, including Allison Harshberger, Anastasia Nechitalio, and Vera Prokodko. Thanks again, and we look forward to connecting with you next week right here on Invisible Machines.