navarre.aiClarisTalk AI · the FileMaker AI podcast
Agents vs. Workflows – with Ronnie Rios
Agentic. It's the new thing. Cris and Matt invite Ronnie Rios to talk about upcoming conferences, get deep into a use case that led directly to the cutting edge. It might seem crazy for a mature platform like FileMaker, but this is the BEST time to do what we do. Why? Because of the skills set of solving totally different kinds of problems in different industries with what looks like magic. Here i
Apple PodcastsSpotifyYouTubeAll episodes
Transcript
Transcribed by machine (whisper, large model), 12,335 words. Names and product terms are sometimes misheard; the audio is the record.
Claris Talk AI. Welcome to Claris Talk AI, the first edition of 2020, what is it, 5, 6? 25. 25. We have special guest Ronnie Rios from Claris. Welcome, sir. Actually, the 2024 guest of the year. We had our little informal awards. This is true. You're our favorite guest of the year, Ronnie. So it made sense for us to have you back as the first guest of 2025. I pay you well. There's some what he does. But you might want to go listen back to that episode, because when I said you were the guest of the year, Matt had a response. So I'm just saying you got a little homework to kind of. I don't remember. I know you don't. Yikes. I'm sure it was humor, though. I'm sure it was, because I'm always going for that.
We had a couple of great episodes with you. We had a lot of fun with his AI conscious one. If people haven't listened to it, it wasn't one of the most watched. But I really think if you guys want to just have some fun with the AI concept, the three of us talking about that. I was the odd man out, and I still found it to be extremely compelling. I still think about it quite a bit. So check that one out for sure. I've been thinking a lot more about that stuff, too, consciousness and reading other books that are kind of interesting. OK, general agenda. So we've got some plugs and stuff talking about Engage, obviously, talking about a lot of the stuff that's going on with AI.
And I have kind of a big question for you guys. My mission is training, and I'm all about how I can boil down. The complex parts and make them eight lines of code that you can just copy and paste and then implement it. So like my dream would be to allow people who come through my training classes or really anybody who watches videos to implement AI without having to understand all of the stuff to black box a lot of it. You know, I think that's literally the mission that Ronnie is on at Claris is what you just described. So semantic search was successful at that. I mean. It's four lines of code. Four lines of code actually does the work right open. I don't know what they are, but I know that there's four.
Well, they just literally one of them is just turning on the AI account access. So that's not a lot of fun, but that you're burning it on there. But the rest of them, I mean, that huge that performs semantic fine. What's the other one? Kind of, I mean, sort of, sort of, sort of, but yeah, fun set and then the performance semantic fine. Yeah, maybe there's just three. Yeah, yeah. So we're. I'm. So happy to hear that because we work very, very hard, hard, not just engineering team, but with members of the community as well to really narrow it down and make it as easy as possible. Oh, yeah. The fourth the fourth one is actually you have to you have to do an embedding of your query so that you can do a vector search and then you would do the embedding of the found set for the search that you're searching again.
OK, OK, plug, plug, plugs. So conferences coming up in Spain in April, April 25th and six in Valencia. I'm going to be going to a conference and doing my session on AI in Spanish. Oh, no, we're not going to. Yeah, so most likely what it's going to be is a mix of Spanish and Greek because they're both in my head now. No, it'll be an English speak. That's actually speak, ironically. Yeah, because I almost just said a few words in Greek right there. Yeah. OK. Then in May, Alexis Allen is having a conference called Elevate FM in Toronto. It's May 26th and 7th, and I'll be doing a full day training day on AI there. So this way, I think you lit a fire under me, Chris, with, you know, because this is awesome.
You know, it's not FileMaker talk, we're ClarisTalk, man, AI. OK, then the champion class. The main thing that I do is teach this intensive class, 40 class hours over a 20 week period, and dive into all aspects. So, you know, from data modeling all the way through, and actually a lot of focus on AI. The next class starts April, wait, I wrote it down, April 7th. And then the next one after that is September 8th. So yeah, hit me up for that. Let's talk about Engage. Yeah, well, we have the superstar of Engage here with us. So, you know, in the spirit of plugs, Ron, you want to talk a little bit about what you've got planned at Engage? And a little bit of hype, because I think, I don't know, my teases, I think this is gonna be one of the more must see Engages in quite some time.
And let's start off with Ron, definitely top, definitely top 30, nicely done. Yeah, so for those who don't know, the Claris Engage 2025 is happening this March, starting on the 25th, runs through the 27th, Jam Pack, we've got sessions about everything, there's obviously a lot of stuff. We've got a lot of sessions from the community about AI, which I think is very important. I myself have a session on AI on the 25th. It's titled Leveraging AI in Claris FileMaker: Unlocking Data Insights and Capability. And I'm going to be focusing on just practical techniques, showing people how to use AI to unlock the hidden value they have in the data that they have in their current applications, how to add AI driven features into their existing application.
So just be a lot of fun. There'll be a lot of demos, some files shared and stuff like that, so it should be a lot of fun. Cover your ears, Ronny. But for those of you that did attend Engage last year or heard about all the hype, the top thing that you heard about was Ronny's demos that he did last year on practical use cases in FileMaker and AI. His tool set has grown even larger and I'll just tell you, I've seen some of these demos and it's bonkers. And he's got a whole month, the rest of us are freaking out that we only have a month. What Ronny can do in a month, what Ronny can do on a Saturday... It's been blowing my mind for years. And so guys, you got to check that one out.
It will again be, for two years in a row, the most popular session at Engage, no question about it. See, I pay him well. But yeah, I'm really, really excited. This is, again, this is my favorite topic, FileMaker and AI, and the things you can do now, right? And some of the things that we have available. Now we've got semantic search built into the product, which I didn't have last year. So expect some of these techniques to involve that. So I think it'd be a lot of fun. Actually, that's a good way to frame it, Ronny. Last year, you were literally just showing things with insert from URL capabilities, like little API calls, which I think is important because, you know, we talked about it before, but there's nothing to wait for.
Like if you're waiting for something, like what Claris is doing is making it easier. Like what Matt said, like really just lowering the bar, but the tech, the capabilities have been there this whole time. I mean, we've been delivering things to market for four and a half years now, but to see the stuff once. We can lower that bar now, take Ronny stuff to the next level and make it even more approachable. If you, if you guys heard about the session, saw the video or went to it last year, you might've thought, ah, awesome as usual, but maybe not achievable. This one's going to be more awesome and more achievable because of the tool sets that are available inside the product.
So it's gonna be a lot of fun. I'm really, I'm really, really excited. Yep. Learn stuff. Pick the demo files apart. Definitely. Yeah. Maybe this year I'll actually be able to fit in the main room, not the overflow that happened. Cause I'm so packed. So yeah. I'll try to get there. Early. I heard about that. Get there early. Right. In one room. This time they got to pick your theater. Well, the keynote will be the keynote will be over or no, no, no, no. I think the whole reason for overflow is just that the theater was so small and there was just so much demand for it. So they're going back to the, um, like at a hotel ballroom, you know, that we've been used to in the past so that everybody's together.
While that format is back to what we're familiar with. Uh, the rumor is again, Ronnie earmuffs, Matt here. Come here. So the rumor is that engage is the biggest FOMO engage in maybe 30 years. So, um, you got, you want to show up for this one, a lot of interesting things going on in the Claire's community, but don't, don't tell Ronnie anyways, we're back. Um, so the other, the other thing is that, um, at the tail end of it, so usually before a keynote, you've spent a day at training. This year training day is at the end, um, kind of like dessert, you know, like, uh, a nice palate cleanser or an app and upper teeth. Is that the right word? Um, so, uh, oh, oh, I'm thinking after T after T after T yeah.
So, but training day is, uh, afterwards. So if you want to stick around and get some more brisket in your belly for one more night and then stick around for training day, I think spaces are still available. Ronnie is that. Probably. I'm guessing. Yeah. For it. Yeah. For engage. The tickets are, if they're, if they're still available, they're burning, running very, very low. Um, same day for, for training day, there are, there are only a few left if there are any, any left. So if you, if you want to go like literally go right now and get your tickets. Yeah. I mean, I'm, I'm super excited about it in the training day. So amongst various others, you can check out the engaged website for all the different training day details.
But since we talk about Claire's talk AI here in this podcast. Uh, there is an AI training day. It's an entire day in a workshop style for AI. Um, it will cover, uh, some of the basic terminology. So those of you that maybe you hear all these acronyms and you're like, oh, what is reasoning and you know, LLMs and what does all that mean? You'll know all the answers to that right out of the gate. There's some stepping through some basic prompting techniques with data, which is a really important skillset to understand, to be able to do a lot of these in product features. Like the ones that Ronnie talks about, um, and then an entire workshop, like hands-on workshop with guided exercises for semantic search.
So if you've heard all about this and you haven't got a chance to work with it, this is your chance to do it. Um, and then, um, oh wait, I'm sorry. That sounds, dude, this sounds really amazing. You know what? I should probably go to that. Yeah. Wait, you know what? I'm going to go to actually. Um, so yeah, the lead is, yeah, so, uh, it's a mixture. It's a, it's a one, two punch with Ronnie and I, um, I'm starting off the beginning part. Part of it, the couple of things that you guys just heard there, um, but then the big, the, you know, the big bonuses, uh, is Ronnie kicking off the end of it. Uh, so a bunch of stuff, let's just say, we can't even tell you what it is, but it'll become very obvious what Ronnie's going to be doing, uh, for those that show up.
So if you want to, I guess the best way to put it around is like, you're going to be teasing a lot of stuff at a really high level earlier in the, in the conference. And then this is going to be the chance for everybody to really see the under the hood, so to speak, not officially, I don't want to steal that brand, but this is your chance to kind of workshop. Um, peel back the curtain on a couple of things. Is that a fair way to put it? Yeah. I'm really excited about this because, um, one of the things that I, you know, from last year in my session, well, um, I had like four or five different demos and, you know, applications that I was, uh, uh, the AI applications that I showed off.
And while we, I showed how they, how they run very quickly, we couldn't get really get into the weeds of, you know, how the script works and how the calculations were working and how the kind of the, everything really comes together. And so, you know, we did. Best we could with, with the format, we only have, you know, 45, 50 minutes there. Maybe we release the files. People can go home and really take them apart. Training day gives us a chance to like really dive deep into at least one of them, right? The idea is that we'll, we'll give you hands-on we'll go through, um, one of these, uh, use cases, one of these practical examples, and then we'll help you.
Let's build it together, right? Let's sit down, go through this, look at, look at the code and start building these things there. So you have some practical experience building these things I am really excited about being able to dive in. Dive deeper, uh, into at least one of the use cases. This is going to be, uh, one of my favorite co-labs, um, Ronnie and I, uh, collaborate on AI stuff for all the time. We talk about this stuff constantly. Um, he's just been a real fun person to be able to partner in crime and all of this, and to be able to actually have a formal setting for this is something I'm really excited about. So of all the training days I've done, which have been every one of them, uh, this is the one that I'm personally most excited about.
This is the one I would actually attend and we're getting a little pitchy. Dog in this, uh, thing before people think that maybe this is a sponsored podcast, literally, these are genuine enthusiasm. And the reason we're talking about this is just because, well, we have a month and we actually are working on these things like every day to actually get it ready. So I'm excited about that. Very excited. Go ahead. I would say one more thing in addition to the training day, since we had our last podcast, I also have, um, like a semantic search and FileMaker course, both on lynda.com for subscribers. And if you're not a subscriber. And you, and you have access to Claris academy, go there and you can actually get a micro certification for, um, semantic search in a FileMaker and AI semantic search.
So, uh, you probably, if anybody's been on LinkedIn, you see, you know, every day, there's like 10 of those things floating through your timeline. Go ahead and jump on it. It'll only take you a couple hours to consume all the stuff and go get micro cert and annual, and you'll be ready for dev con or I'm sorry, I knew I was going to do that. Um, and, uh, you know, uh, properly set up for some of the, you know, the, you know, new, maybe there's stuff coming that you might want to be ready for. So who knows? I learned a bunch of stuff from going through that. My favorite one, which I didn't wish I was sure was not the case. You can actually combine apparently a text data and image data, and you can, um, you can embed them together and then search on it.
This is true. That's cool. I didn't even, I've been describing the multimodal model, which, you know, has multiple functionalities and yeah, I think that's kind of a hidden gem. Um, is, you know, can you, the capabilities are pretty insane. So, um, yeah. And the metadata. So you're getting both things, you know? Yeah. Anyway. Okay. Okay. Can I hit you guys with a question? Go ahead. Cause this is what I've been working on and I'm, I'm, I'm kind of, uh, the way I describe it as this, I've got asked up going really, really smoothly in this direction and really smoothly in that direction, but I need something that brings it together with semantic search. You can search on tens of thousands of, of things, right?
You can bring in. Gigantic PDF and break it apart by whatever method, um, uh, so that it's data in FileMaker embed it and search on it and get really, really good, relevant results with generative AI. You can do something like send a big Jason blob or some text to the limits of what you can do with the, with an API call, which gets really slow when it gets big. Um, and also the limits are real, right? Uh, and also the token window. And so you can get really good, meaningful stuff from that. So for example, if you've got, um, I don't know, let's use a medical example. Um, you've got a patient record with all of their labs and all over their hospital stays and all the other stuff.
And you want to say like, what, you know, what's interesting about this? What should I be looking out for? Something like that. It's more interesting than a CRM example of how many, what are my top products in this invoice? Um, so those things work beautifully now, but what if you want to do a query like that with generative AI against 10,000 records? Well, first of all, let me frame this up because this is literally almost eerily the exact same conversation we had on a standup with one of our AI customers yesterday. And we had a whole debate about this and the way that we frame this up is you're really talking about two different things there. You're talking about, uh, a rag technique, the, uh, retrieval augmented generation, the retrieval is the key part there versus context stuffing and needs.
It's not an either. I mean, it is kind of an either or, um, but one's not better. It's not better than the other. It's, and you know, so to kind of take what you said and put it into that framework, um, when you say rag, you're talking about, okay, I have a question and my answer is probably over here in a database or in a semantic database or in some documents that I've chunked up, whatever it is. And so you will go right. That's kind of what you're referring to there. And it's more data than you can send in an API call. There you go. You could, you could literally have 10,000 white papers in a date, like chunked up in a semantic database or, and by the way, I, it's not only semantic.
You could just have a database of information. If you go to the chat GPT, for example, page and you make a custom GPT, you can put a huge amount of PDF files in there and then query them. Yeah. So finish up the framing. This'll help when we talk about it. So there's the rag technique, which is go search for data because there's so much data that all you want to get is the stuff that's relevant to the user's question, bring that back and then send that to a language model. That's the inner workings of a chat bot. And that we are all introduced to rag through chat bot because of that. The alternative however, is well take 10,000 papers and stuff them into the context window, meaning like the same call you're making to the, to, to the model via an API, obviously you can't fit 10,000 papers in a context window currently, but you can fit a lot more than you think.
Like the Google Gemini, um, the Google Gemini has 2 million token window. Yeah. Which is, uh, two hour long video, uh, well, sure. Yeah. And that's, that's a kind of a different aspect as well, but, um, but I actually think, I think that might be the solution or one solution to this, but I haven't tried it yet, but anyway, continue. Well, you actually are, I mean, technically you are using memory with your retrieval technique as well too. And, and you're kind of, you're creating memory inside the context thing. So, so there are two different ways to do it. So let's talk about when one might work over the other, the way literally in this standup we were having with the customer yesterday.
The way we boiled it down for the customers, we said, well, look, if you have, and I mean, this isn't like a defacto thing, but this is how we described it to our, our customer. If you have like a deterministic question that you want to ask and you want to have the answer synthesized or whatever it may be, or visualize or whatever it is, I think the deterministic part is, is the important part, because then you can actually do a retrieval or you can do a search and go find the answer. Right? Right. Yeah. Well, yeah. I mean that. Yeah. Right. Yeah. And so let's say though, you're saying like, what is our policy for such and such, right? And you are going to go search the policy document and the answer lives somewhere in there and maybe it lives in multiple different pages.
So your return, depending on the, you know, how wide of a threshold you want to allow comes back with all those. That's just a semantic search though. That's not, that's not what I'm talking about. So what I'm talking about. That's what I'm talking about. Okay. So that's retrieval. Okay. So the other part is if you just said, what is the most common topic in all these meeting transcripts from 2024 that you can't do with retrieval that you would have to, that you could accomplish with the context stuffing. So just to kind of frame up when one thing works over the other, that neither are right or wrong. And sometimes you might use the combination of both. As a matter of fact, we I gotta come up with a better name with it.
But right at the beginning of the last year, me and one of the guys on my team came up with this concept of rag stuffing, super gross, sorry. But what that means is instead of bringing back the chunk, bring back the entire document that has the answer in it and because, and then, then stuff that into the context window. So there's maybe a hybrid approach, both of you. So anyways, that's some framing. Yep. So, but you don't find your answer in that. No. Okay. Fire away. Because what I, what I, the use case that I have is let's go back to like a business kind of a thing. I'm doing like a proof of concept for someone right now, which is cool. So here's, here's a thousand clients and they have a huge number of products associated with them of, um, uh, machines that they service in the field.
And they want to be able to say something like, um, what are the, what are the top five products and the bottom 10 products in, uh, in 2024, make a chart and then get the results in a web viewer, which by the way, if you ask for results in HTML and put it in a web viewer. It's like beautiful. Um, anyway, no, it's just a, you don't need JavaScript. You just send it into the request and it gives you the results in the JavaScript code. Yeah. Yeah. Okay. So anyway, that's my case. So right. It's reasoning. It's, it's getting the huge amount of data. It has to be able to read everything. You cannot summarize that. And so it has to be able to read, it has to be able to look at, you know, the thousand customers.
The 10,000 products and summarize it on the fly. Well, first of all, so let me see if I get, if I, so let me see if I understand. So, um, the system would have to first determine that it needs to go look for data. It needs to understand where to look for the data, how to retrieve it and much of it. Yeah. Yeah. Uh, maybe it has to go, maybe it has to do multiple rounds of searching because it might be, my data might live in two different things. It might be two different types of retrievals to, to get the data, then synthesize the data and then actually determine, determine, reflect upon whether or not it achieved its task. Wait, wait, wait, not there yet, but once it's retrieved the data, then determine what's the best way to present the answer to the user, whether to have just text, create a document with a summary or create a chart might be a better, yeah, do that.
So is that, is that kind of what you're trying to achieve? That's what I'm, that's what I'm trying to achieve. I have several different use cases for that. There are, there are names for that, Chris, you want to? Yeah. What you just described was an agentic workflow or an agent, um, where, where, although I'm not entirely sure your specific use case, I think this is a great opportunity. First of all, this is just a punchline with Ronnie and I like literally like a punch line. He's so happy that you said that, but so, so well, so what we're describing was solving our problem by letting, first of all, you know, I mentioned the word deterministic before taking anything deterministic out of it.
You just say, this is what my outcome needs to be. Yeah. The, the, the system will then know everything about your tools. You will have described, described like your databases, whether they're semantic or file maker databases and what, what data can find there and describe even the columns and really explain to it where it can find information. Yeah. And by the way, all that report isn't something I type in a query, so I have a table of queries prompts that I've previously curated. And then the user can just specify what data they want, how much data they want, where it is. But the prompts that I built basically kind of do all that. So it's like a report table, if you will.
Anyway, that's kind of further. Yeah. I mean, you're, you're, your thing might not need an agent because you actually know, well, you're telling it where to go find these things in your question. And it also sounds to me like you're at, you know, you mentioned what are my top whatever's, but that does sound like a database search to me. Like, I think you can answer that problem with a search of your database, right? Like you said best, but if you qualified what best means, like who are my best customers from 2024? Like if you mean. List the top list, the product. So basically this is a, if you did this in FileMaker, like with a regular way. Yeah. Talk about how you would do it this in FileMaker.
Exactly. So you would, and why, and why it's not a particularly efficient way to do it. Right. Yeah. So what you have to say, okay, go search all of my products. line item table of how the products connect to the companies through the invoice, let's say, let's just use this typical example of company, invoice, invoice, line item product, right? And then you'd say, so really you're looking at the line item table and you're somehow sorting it, right? So you're saying, look for a period of time, sort them by how many products you have. And so like, you know, one of them will have sold a thousand times, one of them, three. So you take the top ones by number. If I ask for five, that's going to be a thousand, 800, 600, whatever, you know, and then, and then summarize the dollar amount of those. And that kind of gives you that, but another thing you might have to do, and there's, okay, so take all of those, do a huge, huge many to many relation, right? So you've got like go to related record of the whole entire set to there, which we know is really fast.
Um, which then gives me the customers that have ordered those particular products for that particular time, right? This is hard in FileMaker, right? So are you saying if you haven't set up all the relations already, even if you have figure out where the difficulty is, even if you have, this is a report that you, that would take some time to write. Sure. And it, and it wouldn't, and you wouldn't be able to do a lot of other constraints. You wouldn't be able to, it wouldn't be very flexible. You'd have to, you know, like if, if you, if you tilted it differently, and you say, oh yeah, I want that, but compare the first quarter of last year to the first quarter of this year, like, no, forget it. Your, whatever way you wrote it, it wouldn't work like that anymore because you'd have to do a whole different type of a search or. Okay. So you're talking about a couple of things here, Ronnie, I'm going to tee you up here. So first, you know, we, we talked about RAG and, um, context stuffing. This is, you know, this would actually be RAG is what you're describing because you're actually retrieving data to be able to answer your out of the format of RAG versus context stuffing. Cause this is not that Ronnie and I talked about agents, which just because it's another day of the week. So that's why we're talking about agents, but, but it's really, it's, there is a part here and I'm going to toss this over to Ronnie a second,
but the point I want to talk about is while you do know maybe where that data is, you don't necessarily know how you want to present it to the user. Like you want the language model to, to make some of those decisions for you. So Ronnie, talk about that important aspect of the, these agentic workflows. Yeah. And when you do also recognize that what I care about is the meaning, not, I care about the correctness of the data and the meaning of the data much more than how it's presented. Like if I say, you know, which, which of these, which of these products in the last five years, what is the trend of all the products from Acme co in the last five years?
Forget about writing that in FileMaker, right? This is something I can do, right? It's quickly, ridiculously quickly. Yeah. I think, I think the big thing here is that your, the objective is, is squishy, right? It's not, you don't have a deterministic, you don't have from the outgo, you don't have a predefined objective every single time. It's whatever the user comes up with this, you know, today I'm a manager. My, my, my needs change every day. Today I'm looking at production tomorrow. I'm going to be looking at profit inventory and whatever it, whatever it is. The data lives there. I know the, the answer is hidden there somewhere. Yes. I just need to be able to be able to retrieve it. Yeah.
Now I, I've got ways of retrieving that, right? I can, I can query the data in different ways. We've got OData, we've got SQL, we've got FileMaker Finds. There's a bunch of ways for me to query the data. I can pull the data out. I can, um, if it were me, just human, I can pull those things out and put it in a spreadsheet. And there's a lot of things that I could do this, uh, in. And then once I have it, I have to analyze it, you know, aggregate it, analyze it and present it in a way that not just for me to answer the question, but I have to communicate it to others. And by the way, the, the have to analyze it. That's also deterministic because some developer sometimes will maybe with the influence of a customer will have to define how to analyze it and what outcomes they're looking for. So the, the, the, the fact that there is no set outcome from the very beginning. So you're setting up a system that can do multiple things that has to plan how to use the available tools that it has, use them in whichever order makes sense to accomplish the task over there. Um, use any type of saved, um, uh, occurrences of this in the past to make it easier and be able to maybe, uh, call another, yep. And maybe customize this, customize the output to the current user or right for the current objective, all of those things and
do it as for as long as needed and be able to realize at some point, yes, I'm done. This is, this is, this is, which, which ties to was the fact that it made the plan. It also needs to evaluate what's accomplished and it can also, and it could also realize, uh, this is not doable. Right. I will abort be able to, to pay. Yeah, exactly. That is, that is, yeah, exactly. Or, or your data does not have that date. You know, your, your, your database doesn't, doesn't have that right. You're, you're asking about how to make coffee and you have a CRM. So you have a nest cafe. That's an agentic system literally just described it. So I think, so you're not making this sound like it's going to be something that a regular user can do in an hour of training. Well, okay. First of all, how is it even possible? How would you approach this? Well, so yeah, in FileMaker, um, if we had time, I'd show you, um, but I would say I will take time. If you have, if you have not, if you do not have tickets to engage yet, this is a great opportunity. Yeah. I was just going to say, here's, here's, here's one of the plugs. Yeah. Um, so one of the, one of the things that I I've been doing for the past, I don't know,
year and a half is building my own, my own, uh, assistant, right. This is, this is my little playground, my little laboratory. And it's grown over, uh, over time to incorporate a lot of these elements. And it's gotten to a point where, yeah, it is, it is an agentic system in the, in the sense of all the different components that you've just described. It's able to, um, figure out, you know, how to, you know, plan out a course of action, be able to use existing tools, stitch them together in a particular workflow that helps it accomplish the goal that it needs, and then be able to present that in a way that is, um, complies with the objective, but it also has the ability to say, Hey, I can't do this right. And stop and stop right there. So all of those things are absolutely doable within the platform.
Today. So it, it takes, it takes some planning. It takes a lot, you know, it takes some, it takes some work, but with the, with the systems that we have today with all the building blocks that, that the platform has today, it's absolutely achievable. Cool. It just needs a little planning. So, so basically go to engage and I can see how you'll have a, you'll have a kind of on something like this. Yes. I don't want to go too deep, but yes, but yes, I'll show you something. And is this, I mean, that's sort of the end goal, but what I'm my, my real kind of drive with this is which parts of that, that are really meaningful use cases. Can I train people to do in two hours or four hours or whatever? Well, this isn't novel. There's a lot of discussion about this. Actually, this is the number one topic in all of AI in 2025, the whole concept of agents, the resources for understanding agents are literally ridiculously flooded within the market. And what we're really talking about is, you know, breaking down agents as plan action and observe. There's a lot going on in an action, but you can't do plan. I mean, non-deterministically you can't plan without a language model. And then
if you have a language model doing the planning, it's also gotta be the one that's doing the observing that concept right there is how people should really be thinking about how they're building systems and applications. And this kind of ties back to, um, some conversations you and I have had, Matt, in the past where we've said, or I, well, I've said things like, uh, the hubris of us developers thinking that we knew what this user three years from now was gonna want from this one report. That is really a statement about how deterministic that is. And so much so that a lot of times developers are the ones that are applying that deterministic approach to solving a problem. I mean, really at the core of it, that we hear a user's problem and we go, oh, I know the technology. This is how I'm gonna implement that. And it might be the perfect way to do it. And it might accomplish the goal, but what is changing now is that maybe the best tool for the job to figure out what the plan is might be these super powerful models.
Yeah. And therefore, since it came up with the plan, it should probably check itself to see whether or not that plan occurred and then give it some autonomy, some recursion, all the tools that it needs to do all this as well. And then also access to other potential workflows that are similar. This is a, uh, a, a process. Um, and Ronnie, the, I think we should tell everybody about that. Um, anthropic paper that both you and I are so obsessed with. Um, we'll try to put it in the show notes, but it's, um, oh shoot. I should just look it up. Um, uh, go ahead, Ronnie, talk about it while I find it. Yeah. Anthropic, uh, um, posted a blog where, cause there's so much.
Talk about, about agents and, you know, system and, and everybody has a different definition. Uh, it seems nowadays of what, what an agent is. And so they, they decided to just kind of, uh, just draw a line on, on set and say, okay, everybody's talking about this, but just so you guys know where we're coming from and here's how we view it. And here's our definition of it. And it's such a really, uh, thought out and well presented, uh, blog posts that walks through kind of the, you know, everything from what look what's, you know, your typical advanced LLM call nowadays to all the way up to what a full agent is from their point of view. And I think it's just really well laid out. Um, I agree with most of what's in there. It's just really, really good. I think it's a fantastic read for, for anybody, regardless of what you, you know, how you define agents or not. And for our YouTube folks, uh, I'm scrolling through it, but just to give you high level and it's, um, if you, if you're don't have access to the show notes, you could just Google anthropic and agents. That's the first thing that pops up.
It gives its definition of agents by showing what they believe is a, a difference between workflows that employ LLMs tools and that are orchestrated through predefined code paths. That's the deterministic predefined deterministic, same thing or agents on the other hand, that are systems where LLMs dynamically. So if you wanna just break these two things down, it's either predefined or dynamic. That's really, it's the same, could be the same tools, same potential outcomes. And so then it goes when not to use agents. And it, it does a really good job of, uh, and it talks about some of the frameworks out here, uh, that are very popular. And then it talks about like, for example, you know, this is really the building block of, um, this is actually what you were talking about before. This is, I have something that I need to accomplish. I go to the LLM. It knows what my tools are. It knows where the data is. It can use memory if need be, uh, to be able to accomplish this task much like we use, uh, variables and application memory when we're creating scripts, for example, and it comes up with the output.
So this, they do a good job of talking about the augmented LLM. They go a little bit into chaining, which is like a lane chain. If, if anyone's familiar, even routing, and then this other parallelization, which is actually a pretty popular approach as well. And then it, it talks about what the, uh, differences are in a, um, an agentic, uh, system. So then it goes into talking about agents and essentially I, as much as I don't love the graph, cuz they could have an opportunity to talk about how the action and feedback are actually done by the, the, uh, LLM, which I think is the, the, the most profound part of it. Mm-hmm , uh, this is really what it talks about. So if anybody just wants to break it down, it's either predetermined workflows or dynamic workflows. And it, it depends on which one you want to use for your task.
Yeah. So Matt, what you're describing is completely dynamic. You have no idea what the user's gonna ask for. Um, there's a bunch of things that, that the, that the AI will have a way to tackle that task. And sometimes it's just not even answering questions. It might be something else. It might be, you know, uh, you know, submit an invoice or, you know, th th it, it depends on what your system allows the AI to be able to, to accomplish to the tools that it has there and all the capability and how it's positioned and what the objective of, you know, in general of, of the AI is. So there's, it's completely dynamic. There's no, it's not a determinant. There's no determined path for it to, to do that. As opposed to like, say, for example, if you were, and, and, and, and as opposed to like, if you had something that was very deterministic, say you're, you are, you know, I don't know, you're writing, you know, copy and marketing. You probably have you, the first thing you do is you write a plan, right? Kind of quick, quick, quick outline. You do a draft and it gets reviewed, edited, and then the output, like it maybe, you know, in your agency, it runs exactly the same path. They always do that. And you might have an AI do each one of those steps and the output of one goes and becomes the input of the, of the next step.
That's very deterministic. You know, exactly what's gonna happen. You know, each step, you know, what the definition is gonna be for of, of done running for each one of those, those steps. It's gonna go through a path and you know what the output's gonna be. It's always the same thing. And so that's workflow. What you are describing is completely dynamic, right? And that is more a keen of what an, uh, an agentic system or an agent, uh, I think it's an important way to look at how we're evolving workflows in general, because 100% of everything that we've built in FileMaker to date and shipped has been a deterministic workflow and there's something wrong with it. It accomplishes the task. We don't need to apologize for that or whatever. We, those are the tools we had available now in the agentic era.
We know now that some, something else can, can do this. So, so just, and just a, just a word of caution. FileMaker databases exist so that the database administrators and users, they're the ones who, who are dynamic. They're using the data, not in a deterministic way. The data exists so that they can get whatever answer they want, which is different from day to day. But hold on. We're not talking about the user. We're the user still using in these other workflows. We're talking about who created the workflows. Yes. Uh, this is, I think a really interesting, you know, a, a distinction and in all those things before present day, a human developer interpreted in their human brain, what the outcome needed to be and what the tool set is they had available and created that just look no further than a FileMaker script. A human actually created that succession of, of actions and gave it tools and everything it needed to accomplish the task determined whether or not the task, like literally tested it to see if the outcome was correct and then deployed it into production. There's nothing wrong with that. There's nothing to apologize there by any means, but you have a problem that you could easily solve through hours and hours and hours of deterministic workflow creation, or you could make it dynamic and let the LLM know everything it needs to know about your
tools and very importantly about your users. This is a conversation that a lot of people aren't talking about. And I think there's a lot of really interesting work that still needs to be done there and then combine those two things together, make it actionable. Like you're, you're not even developers shouldn't define actionable and actions. Like the thing with the most intelligence should be able to do that now. Yeah. So really to think about this, it's deterministic predefined versus dynamic. And some, and most of the time predefined workflows are fine. As a matter of fact, I can even tell you in the AI space, most of what we've delivered to our clients has been LLM workflows, deterministic workflows where we're absolutely heavily leveraging, uh, uh, LLMs for a certain functionality, semantic search, all the different techniques, but we are in charge of what the outcome is much like what Ronnie just described there. The agentic version is, it's not just, you get rid of all your old, your pre predetermined or deterministic workflows and all of a sudden make them agents. It doesn't make sense to do that in every case. Now here's, here's something to be cautious about though. There's also, there's another definition of agents. This is not something Ronnie and I subscribe to that's floating around out there. When you see like Microsoft talk about agents or meta talk about agents or actually Salesforce is a great one. They're like, we've got a bunch of agents.
They're actually, those are like mini apps that they've defined a very specific, like a customer service agent or a sales agent. Like they have almost interestingly tuned. Yeah, there we go. You just nailed the irony for me to me. I don't personally, like they're using the term agent, but I don't. It's nothing we would ever run into in our community, but it's really not because it's a little bit deterministic and a little bit agent. So the one, the thing we're talking about is literally a technique is the important framing there. Yes. Thinking about AI as a thousand interns, I think has been really useful and actually fits into this. I think maybe if I'm getting it right, there's a couple of ways you could think about this, right? So agentic, the way we first took talked about it would be like having a thousand interns and, you know, a couple of really smart PhD managers that know exactly how to ask which interns to do what to get into other teams and to have another set of interns check to see if they got this stuff right. And then someone really can understand it. And then the second way you have someone who has a ton of experience in marketing with their team, as opposed to like a team of people, you know, well, when you say experience, you mean knowledge just because that'll, that'll map better to what
you're talking about here. And so, so what you were just describing there is what are very commonly discussed as a multi-agent systems. So in that case, you have like a, you said a manager or something like an orchestrator who knows the skill, the, the knowledge that the various different other interns or people on the team know, and can delegate to the appropriate individuals who have access to that knowledge. Right. That is also a very common approach. Like if you go to Ronnie and I both like a crew AI, you can literally just go to, I actually, it's not true. Whatever crew.ai I'll get you there. I forget there's some weird thing with their website there, but that is a perfect example of literally they call them crews, like a project manager, the crew manager, and then the little sub agents. And then essentially the crew manager says, here's what the plan is. And instead of just having tools at its disposal, it also has other agents at its disposal.
And you can just keep going and going and going. And it's just like, uh, in managing humans that have specific knowledge. Uh, can I, can I share one, one, uh, um, kind of epiphany that I've had in, in the last few months? Um, and it's, so I was, you know, like you, I was a consultant for a very long time. I created applications, uh, using FileMaker most of the time, right. For companies of all, of all sizes. And what I've, what I've learned in the last, um, in the last, I guess, year or so, what I've come to realize is that for all those years, right. I was creating my job was to automate everything that was automatable, right. Deterministically, I was able everything from, you know, managing inventory, to posting invoices and calculating profits and, and all that fun stuff, everything that was could be automated in a deterministic way. That was my job to do. And my job ended where a human was required to make a decision. That person had to click on a button that had to approve a PO that do something right. Somebody had to reason to get an automation and deterministic automation to go.
Yeah. With, with the, with the admin now of this new, uh, AI era, I'm able to go higher. I am now able because AI is able to make a lot of those reasoning decisions. And so I find myself now creating applications that are creating code, um, for that AI to make that decision and do a lot of that automation. I can automate things that were not automatable before, right. That required a reason to be crying a, a more, uh, squishy or fuzzy decision-making, right. That could not, that could not be done before. And now the human is at a higher level. And so I'm creating a lot of still creating a lot of deterministic code, but my persona, the, the end user, right. Is not really a human. A lot of times it's going to be the AI who's then going to be right. Doing things for, for that, for that human. And it's just changed the completely the way, the way I'm building, I'm building solutions down, building applications. Um, because I'm, I'm building in a completely different way. I'm, I'm coding in, in a higher level. My users, right. Are going to end up being at a, at a higher level as, as well, they don't have to really go down, uh, down, down to the details. They're at a much higher level.
I think it's, it's, you still have to have the same data models, right? You still have to be deterministic code. That's not going away. Right. That all that stuff is going away. I think just the way that we think about these systems and how they interact and how the end user, like the user, uh, the user experience is going to is, is completely different. A lot of time I'm building, uh, I'm building tools and subroutines that are not going to be, uh, really exposed to the user. They're going to be exposed to an AI who then is going to make sense for it. Who's then is going to expose it to, to, to the user. Like in your example, Matt, it's, it's clearly the example. I'm, I'm creating a lot of things that will be AI will be able to use so that it can determine how to use those and then be able to answer to the user. It's just, it's really amazing how, you know, the, the user experience of these, of applications going forward, uh, is really changing. I'm having a blast by the way, because it's, uh, it's a lot of it's a dream for me, uh, being able to create things like this. There's been a void for many years. We just didn't even really know it was there. Um, I would actually say maybe the user experience might actually be the same in some cases, but what's happening behind the scenes is so much more fluid and dynamic. Like in some cases there might actually be the user just saying, this is the problem I'm trying to solve. Mm-hmm but the way that it gets solved is so much more dynamic and,
and, and fluid and, and LLM driven. Um, the other part too, and Ronnie just mentioned it again is, um, for these to really work, the, the LLM needs to know everything it meet it possibly can know. And a lot of people really focus on, uh, what tools, first of all, you need to give it tools and tools could be everything like math or literally like, you know, Python libraries that are really good at a certain financial task or something. And the LLM would know to go find that or visualization is probably one that people really understand, but also describing your data to the LLM. And I don't mean like field comments or something. I mean, like, literally saying, these are the only columns you care about. And this column here for this is where you would find the answer to these kind of questions or whatever it might be describing it specifically for the LLM. So preparing your data and describing your data in a manner that it can then have the most, like it will have, it takes all the.
Treat it, treat it, treat it like an intern, right? When you have an intern, you would, you literally have to explain it. They're pretty smart, but they have no idea of the context. So you can literally have to sit them down and say, well, here's the schema of our, of our system. Yeah. Here. Our tables, this table stores, this, this, and that this one doesn't that. Yep. And if you want, I would say they're so smart that you have to do that. How's that really? Like, yeah. Oh, I like that. There's multiple right answers is what I like to tell people. And the way you narrow down to the ones that you care about is two ways defining the data. So it knows where to go find it. And what, what, like you could literally be like, what time did the emergency responders show up for this event? And there might be three columns that have times in it. And it's, it's picking the one that it thinks based on the name of the column, but that might not actually be how it works. So when you go define it, it has the.
All the information it needs. So that name of convention I used to use, right? Well, that we all use, right? I look at my code for, I look at my code from 20 years ago. I'm like, oh my God, what was I thinking? Well, and, and, and, and it, it, it, it's for humans to, to, you know, literally naming conventions are for humans, but we have to describe it a little bit better, a lot better now, the other part. Jason LD too, right? I mean, well, and then the other part of the equation is describing the user and I'm very passionate about this because I don't see enough of this. Honestly, I see it as a huge opportunity and even within our space to finding who the user is is so critical because, and this goes, actually, we had a conversation about this, Matt, uh, last year when we had the UI and UX design or whatever conversation. And we were looking at, um, like, uh, uh, a response to a question that had some visualizations that you hated and then, um, the, the actual answer to the data and then very specific actionable insights, uh, listed below it that were only possible because the user.
Was designed, was, was, uh, determined in a, to a very large degree. This is the user. This is what they have to do. This is what they need to accomplish. This is the things that they have access to without what's important to them. And this is, this is what's important to them is essentially what you're doing there. So if you don't know the user and you're just getting an answer back, yeah, you could say it's not reasoning or synthesized or actionable, but if you do know who the user is, absolutely. It's not, it's low hanging. Know your audience, right? Uh, and, and, and that's just something that frankly, we've never had to think about. Sure. Maybe we've created a user table every now and then fine. And, you know, given a couple of fields to describe the user, but literally describing like a pages and pages of who the user is.
That's that's my dream, right? Is you imagine what we, what we do normally, what we're in the business of is creating custom software. But when we create custom software, we're still creating it for a group of people, right? We we're having, we're giving them the exact same experience across the board. If they go into the invoicing layout, it's gonna be the exact same invoicing layout for everybody, regardless if you're the person, you know, in the front desk, or if you're the manager, it's the exact same UI, but they have different objectives. They have different knowledge. They have different perspectives. They have, uh, different things to do. And so I, I think it's, it's, um, uh, what has been constrained of the tools that we've had so far, but well, I have different dashboards for one thing. I mean, even if the screens are there, you can have a large set of screens and it's so narrow, right? So determined. And that meant for deterministic, it's more right. And it's so, it's so much work for us. And we've tried to, to, to avoid that, uh, dashboard stories. Perfect example of that, right? Dashboards are a key part of the development as developers. We have to actually understand the users. This is not a problem. This is a feature anyway, continue. All you're doing is translating what you know about a user story into deterministic workflows to allow it to be done dynamically with
a request. Yeah. But even from, even from the same user perspective, if you have two managers, even two managers at the same level, the same org chart, right. At the same level, they have different objectives. They do different things. Yep. And it is, uh, it's been again, constraints of the tool sets that we've had, but also bit, bit, uh, uh, you know, the, uh, night, I don't know, night was not, not the right word, but you know, hubris hubris of us believing that we can deliver the best possible experience to everybody with the same, with the same, um, you know, with the same tools and well, and let's be clear, we're not talking about Claris being deficient. No, no, no, no. We're talking about literally every tool set available to man and woman. We now at a point, we're not, we're not at a point where we can actually deliver completely customizable, personalized user experience experiences to everybody. And I think it's an extremely powerful thing that we, you know, we haven't talked enough about.
Um, and I can see, I can, you know, starting to touch the future on that, on how amazing it be for every user to have a completely different experience, the best possible experience from their system without having a developer having to tailor every single one of those experiences. Right. Don't you? Well, there's a couple of really important things to unpack there. First of all is that yes, absolutely. And we have those tools available and people should be thinking about it that way, but the role of the developer then. Okay. So now what happens is you don't become less valuable. You, you start leveraging more soft skills, the things that like going through user stories and defining the user, like instead of defining the requirements of your system, you're defining the user. So take all that energy you spent on what would be the most appropriate workflow and apply it to what you know about the user and what you know about the data as well too. And then let the rest of it kind of work in between. And again, just to touch it again, this is not a deficiency in the Claris platform that led to the, this was every platform out there.
Like literally coding and programming has been deterministic since the first line of code was ever written into the first file maker database was ever deployed. It's just now that can get more fluid. And this is what the conversation is that you hear out in the zeitgeist, the tool sets that, that most AI developers are talking about are, um, accomplishing the same task and really leveraging, uh, uh, AI or language models. A lot of times multiple language models to be able to present these dynamic workflows. But at the end of the day, they're the same type of workflows you can create in our platforms as well. So, um, so it's super important though. And, and so that means the role of the developer, the people listening to us talk here today will evolve. And here's why I think it's a big, here's why I think there's a specific opportunity for Claris developers out there. I'm talking to you, Claris developers, and it's this, we talked about the word custom. We have been in the, we're better than SAS because, or we're better than whatever we're better than SAS.
Or whatever, because we have the ability to do things in a more custom fashion or something more specific to these users. Well, we're custom minded. Our users are custom minded. They all see value in that they have data and that custom mindset. Now, the most customizable that software has ever been is now in the agentic era and moving forward, we are perfectly suited for it because of the mindset. There's no unthinking that has to happen. And, and if we're, if we were SAS people, we're, we're screwed. I don't know if that puts us in explicit or not. That's because there's nothing more deterministic than the off and off the shelf system straight up. That is the most deterministic thing ever. As a matter of fact, it's so deterministic that they're imposing those workflows in their developer decisions on their users. And how many times have we said, Hey, look, you take something off the shelf. You're getting 60% of what you need. Well, that percent of what you need changes in an hour by hour basis. It changes on an interaction by interaction basis. So something off the shelf has never been less relevant. SAS is dead. Just ask Satya Nadella.
This is not Chris's blade saying that there's a huge conversation about that, but SAS is dead and it's the in custom is being reborn. And it's because we have these tools available to us in our ecosystem right now within Claris. So how do you leverage that as a custom minded developer, take those same soft skills that you've always had and apply them towards these work. I cannot, I could not agree more. I think, I think that the, the Claris, the Claris developer, the Claris community is one of the best positions right now to take advantage of this new era. Uh, yeah, because you're not, we're not, you know, most out there, they're not, we're not, uh, we're not just pure programmers. I take missions, right? We've got people here. The community are really consultants are have business knowledge, can understand problems, understand the business mindset, understand what's going on, can solve problems.
The problem solver, as we've been saying, you know, for years, that is. The specialty, the specialty of the FileMaker development developer is not writing code. It's actually understanding different business models across everything, because we've got everything from dry cleaners to nuclear bomb, you know, defusing whatever, you know, FileMaker is actually used in like, uh, um, um, nuclear power plants or something in New Mexico. Don't give us an explicit reading on YouTube. I know that. Anyway, there's someone I know, uh, who works in that. Okay. Um, but, but our, but, as a specialty is, um, is understanding all that business being able to be good at that and having the imagination to talk to the customer and then say, oh, I know exactly what the screen is. I know from, from this conversation, I know that you need these tables with these crazy naming conventions.
I'll, I'll regret in 20 years. Um, and, and this type of an interface, you know, and then we'll have meetings and talk about that. So I think it's a huge challenge though, for the development community to start imagining all the crazy things that we can do. So we can, we still have determinate determinations to make about like, how are we going to define what the user is? What kind of information are we going to have? How much information are we going to put in that user table to get information about exactly who that user is and who their user is not. Um, and how many times and, and Ronnie, even when you were in consulting, you get the, or even when you were in SC, you get this, but how many times has this happened to you, Matt, you're talking to a new customer and they're in some specific domain, right? Um, medical supplies or something like that. And then at some point they'll be like, well, have you guys ever done a medical supply company application before? And everyone on the, our side of it goes, who cares? Like literally our skill is learning everything we need to know about your workflows and your systems and your desired outcomes and adapting the software for you. And that's not a SAS.
It's never, it's never been more important to have those skills than right now. Bingo. That's it point. That's a period. People who are listening to this, there's, there's no, there's never been right. A more important time to have those skills than right now. And if that hasn't gotten you motivated to go to engage, we'll just bring it back to where we started to put a nice little bow on this. Um, and again, you know, I'll, I'll admit I've been a bit of a skeptic in the past and I I'm coming from that real skeptical mindset and I'm, I'm doubling down on why, uh, folks in the Claire's community and that custom mindset is so important moving forward. If you want to learn more about that, you can go to, um, you can go to engage where you can sign up for a training day to see Ronnie and I very excited about that. Uh, also I'm doing another session. I just wanted to point that out too, where I actually describe how we pivoted our services to our file maker customers and introduced AI and how much overlap there actually was. So we discovered that we were actually in a really good place for this. Our.
Customers agreed. I'm excited to bring two of my customers into the session to speak for themselves on file maker and AI integrated systems. And then they're going to present ROI because these are things that happened well in the past and show how things that are only possible with AI or, you know, more, uh, dynamic workflows were of value to them. So I look forward to telling those stories, both first person about, um, how I applied or on my team here at I solutions applied our custom mindset to AI and how. It's a very easy, uh, jump. And then of course we'll have the training day. We'll over teach people how to do all this themselves. And so, um, so with that, um, that I think we did it.
We did it. We have a whole bunch more stuff we're going to talk about. We'll just have another podcast about it. Yeah. I think we're going to talk about all the news. It would have taken enough time. Yeah. Yeah. Well, hopefully we just made some news with this conversation. Yeah. Yeah. Yeah. So for those, for those, uh, if you, again, you haven't bought your tickets, bars.com for slash engage, um, hopefully by the time you read this, there's, you still have an opportunity to do so, but this is going to be, this would be a big one. I am never been more exciting than February 19th. This is going to go live on February 19th. So awesome. Yeah. Well, I'm extremely excited.
It's going to be a lot of fun. Ronnie, another banger. Um, you're already in, you're already the best guest of 2025. So we, we challenge people to take that off the mantle from you, but, uh, can't wait to see, uh, can't wait to see Ronnie. Honestly, can't wait to collaborate with you in person and engage everybody who listened to this. I hope you guys got to the end, um, believe with great passion, both Ronnie and I can tell you, this is something that we talk about on a regular basis, but I hope everybody's inspired for this and thirsty for more knowledge on how to bring this to their existing customers today. So with that said, um, another great episode, we need a tradition. We need to say hi to Alexi Folger at the end of every episode.
He's the only one that's listening. So hello, Alexi. Hopefully we inspired you. Um, but, uh, and if there's another person out there that's still listening, well, uh, congrats. Cause I think this is really important. I really strongly believe this is a really important information. So, uh, with that said, uh, thank you again, Ronnie, really appreciate it. Matt sign us off. Thanks everyone.