The Context Podcast: AI Blastoff

‍

Todd Geist (00:03.714)
on the other side, well, then you have a totally different view and it's very binary. It appears to us. So it's sort of like, that's what we call it being AI pill. When you cross the line, you don't want to go back and do things in a pre agentic or pre AI way. You want to stick with that. And so if you haven't crossed that line yet, maybe maybe don't listen to this yet until after you've spent some time with, with cloud desktop or chat GPT or.

Or I don't know, just hang around and trust that we've been through it and people are going through it more and more every day, every week. There's more people who are AI skeptics who have figured out that, this is actually a real thing. It's happening, it's real, and it's not going away. And it has all kinds of positive and negative impacts. And we don't want to pretend that there aren't going to be potentially

dig downsides to it. Like we just heard that Block, the company that has Square laid off 40 % of their company, the 10,000 people company, and they laid off 40 % of them. So I mean, there's issues and we don't want to pretend there aren't, but we also really want to work on the positive upsides of it. So why I just let Ernest jump in and share some thoughts.

Ernest Koe (01:32.972)
Yeah, I think I'll say that I'm having a freaking time of my life. And it reminds me of some of the heady early days of discovering, you know, HTTPD and then hacking on the web and

Todd Geist (01:38.466)
Yeah, so fun.

Ernest Koe (01:51.15)
hacking and file maker and sort of very early days of things where you really feel like you have a kind of control and access to a sort of power that that kind of cuts through all the BS you know and often I feel like I've spent 25 years trying to upskill and learn all the things I need to like be a better engineer be a better developer and a better builder

Todd Geist (02:17.39)
better builder.

Ernest Koe (02:19.874)
But a lot of it is just like trying to figure out like what the right syntax is and what the right construction is and what the right things are and not really about actually building things. I have been, I've shipped way more code than I've had in the last, I don't know how many years. It's crazy. It's a really awesome time.

Todd Geist (02:33.198)
Yeah.

Todd Geist (02:41.218)
It is. Yeah. I think, you know, we've been doing it for, I've been working with AI coding agents for well since.

Ernest Koe (02:49.56)
since Maven, which was like three years ago.

Todd Geist (02:51.234)
Yeah, cursor, super maven. were some others early in the space, but I think cursor was the thing that finally kicked it off for me in a big way because the tab autocomplete was so good. So we went through the sort of autocomplete era of where it just got really good at doing autocomplete for you. And that was really nice. And then we got agents and that was, that was really great. Suddenly we were just having a chat with our code and it was writing the code for us, but we were still doing a lot of

tweaking and fixing and working with it and reading all the lines of code. Everything had to be read very carefully. And then last year, the models got pretty good and that kind of kicked it up again, especially really with, think it was Claude Sonnet and 3.7, that kind of broke open the gates. It got really good at tool calling and it made the coding even better. And then the models got better again. And by the end of last year,

We had Claude Opus 4.6 and 4.5, 4.6 and then also GPT 5.3 is really quite excellent as well. So the models got really good and the harnesses, the products, the cursors or the Claude code that shipped, actually Claude code shipped like six months ago, maybe longer, eight months ago, but it wasn't really used until say December, January.

Ernest Koe (03:54.766)
or five, Ann.

Todd Geist (04:17.998)
which is like ancient times now in AI terms, right? This is March 2nd and December was like, the beginning of December was a different era in terms of programming. And so now we're at this pace where things are crazy.

Ernest Koe (04:19.982)
I don't know.

Ernest Koe (04:31.438)
I remember switching. Yeah, I remember, I mean, it feels like ancient times. I mean, I'm constantly having to qualify that my information is only, it's a week out of date, so don't trust it because a new update has come and it's different now. was somewhat a new thing has landed. But I do distinctively remember November when I switched from cursor to cloud code and really wanted to see how good.

Todd Geist (04:44.74)
Yeah.

Ernest Koe (05:01.198)
think it was Sonnet. It wasn't Opus, was it Opus? It was Four. And it was, mean, that for me was like, okay, this is a substantially different experience. like it's, you know, it's something change. It's the toolings, the tool calls, also the little things like the skills that were starting to roll out. You it had like this wrapper and scaffolding around.

Todd Geist (05:03.716)
was four at that point, I think. Yeah. Yeah.

Ernest Koe (05:30.39)
at least with Claude code at that point, that I thought was like, it felt really different. And Kurser was already doing a lot of this, but you're still in IDE and you're doing a lot of like traditional ways of thinking about code. And he was just starting to do the agent mode and the non-agent mode kind of thing.

Todd Geist (05:47.994)
Yeah, I should say we are going to talk about file maker here. We haven't, we haven't mentioned it yet, but we will. And like we do just to, know, to give away a spoiler is that, um, we have been writing, uh, we've been vibe coding file maker for, well, since, since, since five coding was a thing. And we'll talk a little bit about how we do that, but I mean, there are, there are things you have to do and there's workarounds and issues you have to, you have to deal with. And unfortunately you can't write file maker scripts with, with, um, with.

cursor or Claude, but you can design layouts if you're willing to go with a web viewer. And we'll talk more about that, I think, in a little bit, but maybe let's just continue with this, with what's happened recently. Cause I think it's, don't think you can understate how big a shift has happened in the last three months. So we had, first the idea was the Ralph loop, which seems like so ancient now, where somebody figured out,

that really if you just basically call agents in a loop and give them the ability to like update their context with what they learned in the last iteration and give them tests that they have to pass before they finish a task, that the things just will grind on for hours in fact. And you've seen, you may have heard the stories of cursor rebuilding a browser from scratch with a process that just ran for, I don't know how long it was, 30 days or.

or more where it actually just from scratch wrote a browser. And so that was pretty interesting because the, but at least for me, what I found so exciting about it is that what it really was, was a return to one of the most fundamental programming concepts, the loop. And that was the thing that sort of unlocked this next level was you tell the agent what you want it to do. You give it some code.

or some context and you, you know, point it at the files you want. And then it sends off the context and that question to the model. The model comes back, writes some code, and then the agent would say, well, did this work or not? And if it didn't, then it would take what it learned and loop. So it's just a loop. And it's just retrying over and over again until it finds itself with the conditions that

Todd Geist (08:15.862)
either through like end-to-end testing or some agents as judge or LLMs as judge tests that it completed the task. So fundamentally, that's one of the big breakthroughs that happened is that you just put these things in a loop and suddenly it doesn't matter that they don't get it right the first time because in three or four loops, they'll get it. And that's magic. That is like, wow, we're in a different space now. And that also teaches you like how

You need to start thinking about programming, which is really about how can I get these things to run in a loop for as long as possible with as much, you know, context as they need kind of thing. So that was pretty cool. That was Ralph loops. That was in like, maybe that was over December. That was over the break. I Yeah. Ralph Wiggum, which was a character from the Simpsons. You can Google this if you don't know about it. And there it was the very simple context loop. Did I finish? Nope. Try again.

Ernest Koe (08:59.104)
Yeah, that was Ralph.

Todd Geist (09:14.09)
repeat until done and it burns a lot of tokens but it actually does work and so yeah that was a big one.

Ernest Koe (09:24.354)
kind of chuckling here because I'm looking at your t-shirt and you have the automatic loops and I have the proof loops on here. I'm thinking this is like, this is so weird, but in a way this is exactly like write and meta, right? mean, so we've known for a while now that there's an LLM and then what makes an agent an agent is that it's a loop using the LLM as a...

Todd Geist (09:29.634)
Yes. Proof loops. wow.

Ernest Koe (09:51.662)
as a judge essentially, right? So that's what's one loop. But the cool thing about agentic programming is actually a loop on top of a loop, right? It's actually two loops that are running. It's the agents are, the coding agents are already doing the loop. And then you throw something like your harness on top of that, which is like whether it's a Ralph Wigum or Ralph loop or anything like that. Now you're in this process where you're asking your agent, which is actually already looping to do stuff to then loop through itself as it's looping to do stuff, which is like.

two loops, like, is just what we're doing.

Todd Geist (10:21.39)
Yeah. So I think we just named our next product proof loops. I think we we definitely have to do something like that. Yeah, so they can loop through a single task or they can loop through a task list. And so plan mode was an iteration of this right where all of the major coding agents now do plan mode where they'll they'll create a checklist and they'll work through that checklist. And yeah.

Ernest Koe (10:26.094)
Proof loops.

Ernest Koe (10:45.176)
then you see this. You see this actually if you pay attention to how, if you're using cloud code, things will fail. Things will kick off another process. I'm going to try this. I haven't read the file. I'm going to read the file now. And then I'm to do something else. And that's this internal loop. And then it does that thing. And then it goes through the loop of deciding if you set it up properly, whether or not it's going through the plan correctly and all that. So I think this is a fundamental shift in the way we think about

what this is. I really don't like the phrase vibe coding because I think it's sort of, I'm cool with being, I'm cool with it being cheeky and whatnot, but I think it sort of misses what's actually happening. It's not like you're just asking the LLM to spit out an output. I think that's almost like an old way of thinking things, which is like there's an answer box.

Todd Geist (11:26.874)
Thank you.

Ernest Koe (11:42.358)
and then you're getting answers out of this box and then you apply these answers to whatever product you're doing. I think it's something quite different, which is that you are actually engineering a layer and sort of a way of automating it getting its own answers and delivering a product, right?

Todd Geist (12:01.038)
Right, like I think that's super important. Like people always say, well, they get things wrong. And that's true, they do. But again, if you just empower them to learn from the mistakes that they make, they will get the answer eventually. Especially if it's anything that's that's common out there in all the training data, they're gonna get it. And so your job now as the orchestrator of these agents is to provide them the harness, what we call, what's called the harness sometimes or the...

or the agentic platform that it can use to, yeah, it can use that system to take in feature requests, bug reports, sentry errors, whatever, and then take a stab at fixing, whatever it is, and then run a set of tests. We call them gates that say, yes, you got through, you did all the things, you completed the task and you didn't break anything that we've tested. So therefore you can mark that task done.

Ernest Koe (12:33.486)
or the system, right?

Todd Geist (12:59.802)
And then you go back to the beginning of the loop and you say, there another task for me to do? it, there's another task, great. And it moves on. I think like it's really important to understand that, that, that in computers, like in computer software, in all applications, somewhere there's a loop. Like you basically waiting for the person to click on something and do something else or, you know, there, there's an event loop in, in the browser, in JavaScript, that's very explicit.

Ernest Koe (13:05.006)
Yeah.

Todd Geist (13:27.908)
that there's an event loop waiting for new inputs to come in and then it just runs through its process around again. And that's also the same in all games. All games have a game loop that is basically just looping and waiting for new information to come in that it can respond to. so that's it. So again, we're still in the same place that really all computer activity or everything that we've talked about in terms of like productivity or gaming.

is all based on the same concept, which is loop.

Ernest Koe (14:00.858)
That's, think, really important to know. I it may seem so basic and primitive, but I almost think of it as the reason we give anything in the computing space power so that it goes into a loop and can sustain that until it turns off, right?

Todd Geist (14:15.642)
Right. Right. We've all written this. can file maker anything else. Right. Looping is what we do all the time. We loop through the records. We loop through a found set. Yeah. The operating system. Everybody does this. Right. So the new thing, though, is prior to LLMs, what happened on the loop once you compiled your code was deterministic. Like the same inputs would as long as the state didn't change inside there, like database records or things, you were always going to get the same result. Right. You had if else statements or things like that that

Ernest Koe (14:22.574)
Well, it's like your operating system does this out of the box,

Todd Geist (14:45.796)
When I encounter this, I do that. When I encounter that, I do this, that kind of thing. So it was deterministic in terms of what your program could do. LLMs come along and they change that because LLMs add a little bit of probabilistic behavior to it. They predict what the next behavior is gonna be, what the next behavior should be, and it goes on and does it. But by the way, interestingly, this is also what happens in biology a lot. So for example, your eyeballs, your eyes are not a camera. They're not like projecting

Ernest Koe (14:58.734)
You've got intelligence, leader.

Todd Geist (15:15.928)
what you see into your brain. Your brain is actually predicting what is the next thing that it should see and it's sending it to the eye and the eye is sending back corrections, right? And then the brain repeats that. So it's a loop. How you see is a loop. The brain predicts, the brain is the LLM predicting the next case. The eye is the gates that say, yes, this is correct or it's not. And then it goes to that loop again. So I mean, this is a fundamental concept that's

Ernest Koe (15:17.72)
It's just predicting. Absolutely.

Ernest Koe (15:24.782)
100 percent

Todd Geist (15:44.704)
know, throughout computer programming, but also throughout life, throughout how biology works as well. So it's really core and key and simple. Like I think we all know how powerful loops are. If we've done any programming, any file maker, any other kind of programming, we understand something about loops and all of this knowledge will help you in this new era because that's again, the foundational concept you need to understand.

Ernest Koe (16:07.022)
Getting really comfortable with this idea that you're really basically allowing for errors to happen and then you are creating the conditions for it to be addressed, to fix it. I think that's a sort of a different mentality and a different mindset. I think in some ways what we think we're doing, which is try to get things right the first time.

in predictive, in a deterministic way. So I think the ambiguity is important. Engineering for not specific bulletproof conditions, but engineering for things that can fix itself or can learn and can feed back into itself, I think that's really fundamental to all of this.

Todd Geist (16:56.994)
Yeah, so you hear people kick around these terms, agentic engineering, or they talk about meta programming, or meta systems, systems that build systems. And I think both of these are getting at this similar idea. Agentic engineering is building systems that build systems.

Ernest Koe (17:13.218)
That's right. think in a way, we're also saying that, I think I believe this, like programming as we use now, it's basically dead. But engineering the idea of building systems, to build systems, to write the programs, that's where we are. That's the frontier.

Todd Geist (17:31.362)
Yeah. Yeah. Yeah. And so, I mean, there's just so many second order effects and we'll have more podcasts. We'll get into more and more of these things. There's so many different directions that we could go on this, but yeah, like there's no, there's, you just don't wanna, you don't wanna go back to the keyboard and type functions anymore after you've been doing this new way of coding. I usually, I talk a lot to the agents I don't even like to type.

Do you remember when we were meeting with Simon from Assistant UI and we talked about what LLMs were, were higher bandwidth, right? Because that's another way to think about it is that I don't have to type anymore. I can just talk. I am not a good typist. I can't type 150 words a minute, but I can talk probably 200 words a minute. So it's like getting the ideas faster into the computer and faster to the agent is what makes me feel so productive about it.

Ernest Koe (18:05.293)
Sure, yeah.

Todd Geist (18:30.362)
Let's see.

Ernest Koe (18:30.542)
Where do you think you're spending your time on these days, if not putting...

Todd Geist (18:33.764)
So it's really interesting. like it took it takes a while to like out of so many things, but let me just stick with sort of the theme or what what's. About this agentic engineering or meta programming or or meta systems. So in the old days, but and by that I mean December, January. There's like you go through this loop. have a loop. You're in the loop where you you write some code. You tell the agent to write some code. You run the code and there's an error. There's an error.

Ernest Koe (18:51.81)
months ago.

Todd Geist (19:03.962)
And so you paste that error into the terminal and you say, fix this error and it would do that. So you were in the loop, right? And so you were always the bottleneck because if you didn't give the error back to the agent, it wouldn't know that it needed to go again. so what I've noticed is like, so the first thing is to get it so that it can check its own errors.

And you can do that through various means. The harnesses themselves have gotten pretty good. Like cursor will, and Claude also could just read the terminals that's running in. It knows what logs or what error logs are being kicked out there. But also there's agent browsers. So the agents can actually control a browser and they can connect to Chrome and they can read the console logs out of that. And they can also click around, right? So they can now experiment themselves.

Ernest Koe (19:56.886)
Right?

Todd Geist (20:00.834)
did I move the button? Is the button in the right place? Or is, the button's not blue, it's green, I need to make it blue. So again, you're closing the loop on that with these harnesses, these tooling layer that you can put on top of it so the agent can do more and more of the work on its own. That's pretty important. So what I'm doing now is I'm almost like another step up from that where if there's a feature that we need to add to the system,

I wanna make sure the system can just make a feature. you know, don't actually wanna write the feature. I want the system to be able to take my feature or whatever it is and write it. And that shows up in a lot of different ways where you might have to add more kinds of testing. We do way more testing than we used to because that's the unlock, right? You want the agent to be able to prove to you that it did the thing that it's supposed to do.

Ernest Koe (20:48.844)
Away my testing.

Todd Geist (20:57.122)
And so you do more testing, you give it more tools like agent browsers, and then you have to change the way you think about like.

how can I make it easier for the agent to do whatever it needs to do? And that shows up in a lot of interesting and weird and subtle ways, but it's mostly looking for that. So one example that's maybe subtle is I use less libraries than I used to because a lot of libraries were just sort of low level or DX layers, developer experience layers on top of code that

could be written fine, but you just didn't want to because it took too long to do it or it was too challenging. You didn't understand the domain very well or whatever. So you would choose to use a library and that's, still use them. So this is not like a hard line, but if you use a library and the library has a bug in it, your agent isn't going to look for to fix the library. But if your code,

has a bug in it, it notices, this routine I wrote, this routine we're using to like, I don't know, calculate the sales tax or something has a bug in it, and I can just fix it. So it was just gonna fix it. Whereas if I had used a library that calculated sales tax, the agent would just be like, there's a bug in your library, you gotta go tell the people who wrote it to go fix it. So it's lessening the number of dependencies that we bring into a system. It's just writing more, you know.

whether it's Python or TypeScript, just regular TypeScript. So it's like that kind of thing, know? Tailwind, Tailwind was another way in which this showed up. Agents are really good at writing Tailwind as opposed to like other kinds of CS and JS libraries. It was much more difficult for it to do that.

Ernest Koe (22:42.328)
Yeah,

Ernest Koe (22:54.126)
They are. Although I'm even finding that, you know, where I would have instinctively reached for a Tailwind or a JavaScript library. Like, for example, you know, when we're retooling our own website, it just turns out that HTML and CSS is just really, really great out of the box. Like, all the animations already work. mean, this is the kind of crazy thing, right? I mean, like, the...

Todd Geist (23:14.264)
Yeah. Yeah.

Ernest Koe (23:23.586)
The protocols and the base definitions for HTML and CSS, they're actually really quite deep already. We just think we need all these layers to make our lives as developers easier, but AI doesn't care. You can do CSS animations without me hand-tooling it. It doesn't care that it needs a layer of abstraction to understand it in human terms. It's quite good at that.

things become leaner and meaner and faster and all those things.

Todd Geist (23:54.084)
Yeah, yeah, for sure. I don't think libraries are going away, but I think we will use less of them. There's a whole class of things that really only provided sort of a light layer that made it easier for you to type on a keyboard. I mean, if that's what your library was doing, like it reduced the number of keystrokes you had to do to describe something, I don't think those survive. They just won't be used because the agent type it, you know, 300 words a minute. doesn't care. just, it can apply a diff in one shot. Like it just doesn't need to.

Ernest Koe (24:15.99)
Right.

Todd Geist (24:23.918)
worry about that kind of thing.

Ernest Koe (24:24.064)
If it's purely for human, it's purely abstraction for human purposes, you know, it may not be as necessary as what, as before, you fundamentally.

Todd Geist (24:41.621)
Hang on one second.

Todd Geist (24:49.146)
We just got a something went wrong.

Ernest Koe (24:55.087)
recording.

Todd Geist (25:03.188)
I'm gonna guess everything's okay, we'll just keep going.

Ernest Koe (25:06.094)
It's still recording here, so.

Todd Geist (25:08.432)
Yeah.

Seems alright. Hopefully it'll be alright.

All right, so from there, let's see, where are we going next? this is sort of, I mean, there's a long way we can go with this. don't know, what else we want to talk about here?

Ernest Koe (25:29.08)
But this is, yeah, go ahead.

Ernest Koe (25:34.616)
Well, mean, I think just riffing off of that thesis, is, yeah, things have changed fundamentally, I think. Things have changed really dramatically since two months ago. In 2006, think it's a, we are definitely past this inflection point where, I think it's safe to say that maybe there's still some skepticism whether AI is actually really useful in, you know.

in sort of a general case, but in the domain of developing software and programming, I don't think there's any question that AI is winning.

Todd Geist (26:16.692)
There isn't, yeah. And everything that's, you if you're seeing skepticism out there now, it's every time something like this comes along, there are people who hang on for a long time for various reasons, good reasons and bad. And there certainly will be people who specialize in very esoteric versions of software that might still be writing code by hand. the amount of, the percentage of code that will be written by agents as opposed to humans,

is going to approach 100 % very rapidly. And it's kind of like, what's that paradox called? Jevons paradox, right? When something like this happens, it's not that the old way necessarily goes all the way, it goes away, but there's just so much of the new way that it swamps it out in terms of a signal. it just, all the code is gonna be written by AI agents very soon. Yeah.

Ernest Koe (26:52.767)
Jevons paradox, you

Ernest Koe (27:07.662)
All right. So this has sort of profound implications for our community and the work we do and the business that we've been in, which is arguably you can say we've been working in a low code space for many, years and trying to apply a practice of bringing problem solving to a variety of constituents, including businesses and

Todd Geist (27:17.46)
Yeah, for sure.

Ernest Koe (27:37.452)
for ourselves. And for many of us who have built a career on this, where we have gotten good at not just the low code part, but some of the coding part and some of the architectural and systems level things that you need to know to really be good at your craft, that this has a profound impact on what we do.

Todd Geist (28:01.278)
Definitely, like it's a pretty big deal. Like in terms of like, we just talked about how libraries that just kind of made it easier for you to bring about some kind of, you know, programmatic effect are gonna go away. And I think that's pretty safe to say about low code and no code in particular. Like it's the same thing. Those concepts, those things were valuable because...

Ernest Koe (28:21.166)
That's the same thing.

Todd Geist (28:26.76)
They provided that layer that was good for the humans to be able to do without having to go get a computer science degree or learn how to do text-based code editing. And that whole category is completely unnecessary today because AIs can do it.

Ernest Koe (28:45.132)
Right, because all that stuff was the abstraction layer for us to bring our problem solving and our creativity to actually deal with the problems. now the abstraction layer, AI is the abstraction layer. It is the one abstraction layer to do it all. It's the last one.

Todd Geist (29:04.756)
It's the last one, right? It's the last abstraction layer. Maybe, I don't know. Well, there's a lot of interesting discussions about what might happen in terms of programming going forward. But I think that's just the key thing for right now is that it's the last abstraction layer. You're not going to be, for the most part, writing code. For the time being, you're going to need to be able to do some things around it. And you're going to have to do

Ernest Koe (29:13.294)
We'll see.

Todd Geist (29:34.196)
You don't have to learn this agentic engineering, this loop and how to apply that. And that may persist. That may actually persist for a long time. It's not clear how far down this goes, you know? But I think you're gonna have to be a mindset and wanna build and wanna be curious about that, right? That whole thing is still gonna be really helpful for you to be able to do the things you wanna do in the future.

Ernest Koe (29:48.152)
But to be, yeah, right.

Ernest Koe (29:59.822)
be clear, I don't think we're being dramatic about this. I think this is our experience today. We've already moved squarely beyond the place where we are using abstraction layers to deliver quote unquote value to people.

Todd Geist (30:09.436)
Yes.

Todd Geist (30:19.314)
Yeah, yeah. And we're seeing some interesting things are happening. Like, for folks out there that haven't made this leap yet and you're worried about how you're gonna be able to do it, there's some good news and bad news scenarios to this. I mean, one, you are gonna have to learn some new things. Like that's the bad news. You're have to learn, at least for the time being, you're have to learn all the impacts around security and deployment. And we'll have more to say about that down the road. But there's stuff you gotta learn. But...

But on the good side, there's a whole bunch of stuff you never have to learn. You just skip right over. It's kind of like people who went right to mobile phones, They never had landlines. And so there's a whole class of things that you'll never have to worry about. And that was kind of interesting as we started pushing this stuff out to members of our team that aren't programmers, like people who key positions in the company, but aren't programming day to day. And I was thinking we're gonna have to explain all this like heavy stuff to them to get them to the...

Ernest Koe (30:55.822)
All right.

Todd Geist (31:17.428)
to the place where they could do this. And it turns out, no. In fact, they can just jump right in and the AIs with some of these ideas and concepts are good enough that they can build stuff and they don't have to learn. Like they don't have to learn about enums or generic types or any of that stuff. They can just happily skip over that entire curriculum and just keep building. And that's pretty exciting.

Ernest Koe (31:47.47)
Yeah, I certainly think that. I thought that experience of watching people do this and going, you know, they don't really need us to explain all this stuff. Yeah, they can do it.

Todd Geist (32:03.624)
No, they're just like, the AI can do it. They were like, yeah, the AI can, but see, didn't, some of them, they didn't use the AI before it couldn't do it. So to them, it's just like, works, right? We went through a period where it didn't work, and so we had a lot of skepticism and stuff like that. But now it's to the point where that, yeah, you can really get stuff going and it works pretty well. And it still issues, but it's phenomenal how much easier it is. So you won't have to learn.

Ernest Koe (32:18.221)
Right.

Ernest Koe (32:28.834)
It is phenomenal.

Todd Geist (32:32.732)
all this stuff, you won't have to learn React, won't have to learn Tailwind or any of these things. The AI is gonna be able to do it for you. You have to learn how to use the AI. You have to learn about agentic engineering and meta systems and things like that, but you don't have to learn TypeScript, which is good news, you know? So that's pretty cool. So I think that probably takes us to what's up with FileMaker in this space. Does that seem right?

Ernest Koe (33:01.378)
Yeah, I think just real quick on that, mean, I do think, I mean, I don't think that this is the case where.

you can be absolutely ignorant of all things in this domain. At some point, when we are delivering to production and putting things out there and it has to be maintainable, it to be tested, there are real things that somebody needs to know and they need to be built into the way we build things and whatnot. But I certainly think that the inversion has happened.

Todd Geist (33:16.382)
Yeah.

Ernest Koe (33:42.936)
Right? And it's like before you had to have this huge body of domain expertise and knowledge to even get started. And that's the opposite. Now you kind of come with the problem and you need very little of the underlying mechanics of how to do things, which you can get later and gain later and built on top later and get good at and improve at. But the starting point has changed a hundred, know, 180 degrees.

Todd Geist (34:06.184)
Yeah, I mean, just things like Git, which if you didn't know Git was a major roadblock to any of this stuff, And now it just doesn't matter. Like the AIs can do Git really well. So you don't have to worry about merge conflicts or any of that stuff, really. The AI will figure it out. So.

So in terms of FileMaker, obviously there's problems. We can't really vibe code scripts and database tables and fields and those things. So that's a little bit of a bummer. We've done some experiments with generating code, generating scripts. And I think we could actually improve them. And we may try to do that in the coming months. But I think we've just...

The main thing that we do now is whenever we can, we just, and if it's a file maker project, basically there's two ways we go at it. Either we build a web app, which is all in this space of agentic engineering, or we build web viewer apps, which are really cool and have a lot of really great simplifying

features. And so basically we just don't use layout mode. We have another phrase that we say instead of don't use layout mode, but I'll maybe and with respect, we'll leave that alone. What'd you say? think you might be muted there Ernest.

Ernest Koe (35:40.846)
conveniently my AI is watching out for me. I'll revise that to say, I think we say F layout mode these days. I mean, it's a little bit of a bombastic and tongue in cheek, but I think this day and age, I think it's important to have things that cut through the noise. Like we need a really strong signal for why we think this is important. And for me, it goes something like this, right?

Todd Geist (35:48.136)
Yeah, we say F layout mode. That's it.

tongue and cheek a little bit, but not really also.

Ernest Koe (36:09.742)
In order to participate in this agentic process and economy, things have to be accessible agentically. Meaning AI has to able to talk to it and do it. So you can do AI one of two ways. You can say, OK, I'm going to use AI. We've talked about this before. We can use AI and chat GPT, or some conversational experience, to kind of create this bits and text that you want. And you go plop it into the thing that you want to change.

That's actually, I think, a wrong way of thinking about the space that we're in. I mean, think lots of things are trying to do that. They're trying to bring quote unquote, you know, like this conversational layer into the tooling section, but you still have to manually go in and do things and, you know, and change things. I think that is the basic thing that we're arguing against. Like the test is, if it's not agentically accessible.

Todd Geist (37:06.292)
To the loop is the thing. Yeah. That's the looping process. That's what allows the software to learn and grow on its own. Yeah. If you're always in the middle of having to copy and paste stuff, it's a problem, right? I mean, we did this with ProofChat. If you have ProofChat, you can actually generate scripts and they can make the XML and you can paste it in your solution. It's great. In fact, I often do use AI when I want to write a complex file maker calculation because it's pretty good at complex calcs.

Ernest Koe (37:07.192)
to the loop is the thing exactly, then it's broken. Right. Right.

Ernest Koe (37:34.102)
Right. It's way faster.

Todd Geist (37:34.548)
And it's way faster. can just tell it, I just tell it what's in there and it makes a nice, I can copy that, paste it in. It's great. You know, that kind of thing works pretty well, but that's not accessible to the loop, right? You, it's gotta be headless. Somebody else, the agent's gotta be responsible for, you know, evaluating doneness and going through the next process and all that. So as long as somebody is in there copying and pasting, you're sort of, you're in a, you're in like, you know, the ancient days of early 2025 or something like that.

Ernest Koe (37:44.702)
It's got to be headless. Right.

Todd Geist (38:04.372)
2020, you know, maybe 2024. that's it. So, but there's one area where we can do this, right? And that's the thing. There's one area where we can actually expose, a whole big chunk of our app to this agentic loop. And that is web viewer. So you put a web viewer on a layout. That's an app. You can get your data in and out of there using the stuff that's built into the web viewer.

Ernest Koe (38:09.518)
That's right.

Todd Geist (38:32.0)
And with the proper setup, you can get a very nice agentic loop going with all of that. And so you basically say F layout mode and you build everything. build one, basically we have several apps that we ship this way. And we've been shipping, I think the first thing I did that was actually a product in the FileMaker space, that was a web viewer app. It ran in FileMaker Go and it was Go sign before.

FileMaker Go had signatures. You could do it with this little web viewer app I made. And that was like 2011, 2012. So we've been doing web viewer apps for a long time. And it's always been one of our favorite ways to do it, to build FileMaker UIs. just, since really it became like my favorite in 1906 when we got all the good JavaScript stuff that came with that release. Yeah. So since 1906,

Ernest Koe (39:22.914)
And that's important.

Todd Geist (39:26.824)
This is the way that I prefer to write file maker apps. I still have to write my scripts and my tables in the old way for the most part, but all of the UI is built inside of a web viewer. So if you look at AutoDeploy or if you look at ProofChat, basically they have a single layout that is the UI and there's a web viewer on it and it covers the entire screen. And then now you can, with again, the proper setup and tooling,

You can actually get the full agentic loop going and that's the magic and it works quite well. we have, I don't know, thousands of apps that are thousands of people, thousands of files, I would say at this point that are shipped that way that people use all the time. Think of auto deploy, for example, is that, and that is running on, I don't know, there's well over 3000 servers running auto FMS. So there's a lot and

It works great. Like we have zero problems with it. It's fantastic. And you get a lot of freedom from doing that. Of course you can make the UI look like anything you want. But you also get the benefits of FileMaker, which I think are very useful if that's the place you're in, right? So you talked about security there. What's the issue with security?

Ernest Koe (40:46.35)
Well, I mean, I think there's still some hard edges in this agentic world, right? I I think I don't want to overstate it. I also don't want to understate it. And I think that, you know, I think at some point, all these will be sort of solved problems. But I think right now, especially in an enterprise situation and environment, coordinating security, getting it to be

sort of a solve problem already rather than having to reinvent the whole thing yourself, getting to work nicely with other apps, to work nicely with an enterprise security setup, with SSO, with auth, with users, with groups. I mean, those are still kind of gnarly things that you wind up writing and rewriting over and over again. And the beautiful thing about using FileMaker, especially with an existing solution that you're trying to improve rather than...

You know greenfield per se is that all that stuff including business logic? That's wired deeply and tied to security is already potentially Done for you, and you don't have to figure that out you know and so I think that's a really nice layer It's like you know it's like you know if you can deploy and all the security and stuff is not your primary domain of concern Right like you know as a developer can contribute to that rather than having to solve all that yourself from day one

I think that's an awesome story to have, which is conveniently, think, what Follink is actually pretty good at today, at that level.

Todd Geist (42:20.446)
Yeah, like it's a nice security layer on your data and on your app, right? So if somebody can open the file, they have a set of privileges that allow them access to whatever data they've given access to by the developer. And if you're building web viewer apps, those web viewer apps, if they're connecting straight through to FileMakers, through the JavaScript bridge, then they're inheriting those privileges. So the person who's logged in cannot get at more data than...

Ernest Koe (42:29.676)
Right.

Todd Geist (42:49.128)
than they have been given access to. Whereas if you step outside, you gotta build a web app. You're gonna have to deal with that somehow. You're gonna have to, when they log in, you're have to give them the same account somehow. There ways to do this, but again, you now have to do it. You have to make sure that that website is secure and there's no cross-site scripting or cookie hacking or anything like that that's going on there. You have to do all that and there's libraries to do it. It's all doable.

But with FileMaker, that's not even a question. When that app boots up inside the web viewer, it only can access what that logged in person can access. So that whole problem goes away. So one of the we like to think about FileMaker in this new world is it's an app container. It's a deployment target. I can deploy an application into my secure environment. And I don't have to think about re-figuring all that stuff. And I just need to figure out how to connect it to the

Ernest Koe (43:34.402)
Right.

Todd Geist (43:46.42)
to the bits that I need in the backend and that's very doable. And also very susceptible or susceptible is maybe the wrong word, very compliant or it fits in with this agentic loop because the agents can do this stuff. They know how to do it.

Ernest Koe (43:59.278)
Right.

And it fits in with a lot of enterprise and business security requirements and models. mean, is in an app container, just to spell it out. It is an application that lives, that you can launch on your desktop. So it has the connections to your local space, your local environment. And that's not a small thing. And it's going to secure a connection to it. And it also has a local data persistence.

Todd Geist (44:07.123)
Yeah.

Ernest Koe (44:28.462)
engine already built in. So it's got a runtime. It's got a web runtime and a local database layer. All those are hard problems still today. We still have to deal with those things if you want. it's kind of nice. If you want to build apps, especially desktop or maybe even iOS apps, FileMaker is an app container or a hybrid kind of solution where some things, some administrative layers or crowd layers are still basic.

know, phonic layouts, but all your rich UI now doesn't have to be stuck behind, you know, the old phonic layout. I think that's a compelling reason to think about how we extend what's already there today.

Todd Geist (45:13.14)
Yeah, yeah, I think people don't realize that there's a lot of like really besides security, it's other things, right? You can think of your file maker scripts as the backend server in a web app, right? The API layer. So your scripts can be, you can think of them as APIs, endpoints that you can hit to make certain things happen within the database, update data, delete records, you know, what have you. And also like if you're building in the web,

One of the things about a browser is that control click and you can get the dev tools in the browser. Even in FileMaker, you can do that if it's been turned on in your browser. And then you can look inside it and there's actually the data that's in your view is there, right? And so you have to know that and you have to protect things that shouldn't be visible to the user.

And the beauty of it is you just don't have to pass that to the browser and then they can't see it. But you still can have access to it through scripts. So this is like in web development terms, you've got a backend API server. And that API server has access to environmental variables. It might have things like API keys and stuff like that that you would not want to expose to the browser because you don't want those to be visible within the DevTools. The FileMaker app is that thing. It is your backend server.

you can put your secrets in there. It's encrypted at rest. So whether you put them in as script steps or in custom functions or even in fields, you can encrypt them. You can store them quite securely. So it has that backend secure layer that a web app requires. And again, it's built in. It's stuff you already know how to do. It's right there. So again, it's one of these really nice things that you get to use for free without having to figure it all out, developing a web app.

Todd Geist (47:11.124)
Yeah. So layouts building apps in file maker, um, without layout mode is my favorite way to do it. Like that's the way we have ever built anything in file maker. And, know, going back to the nineties, this is the most fun. If you can get an agentic loop set up on, on file maker in this way, and we can do it and we'll be shipping things and talking about this more. don't have time to get into it today, but we can do a full agentic loop.

Ernest Koe (47:26.413)
Yeah.

Todd Geist (47:40.04)
with file maker based web viewer apps that work really well with things like cursor and cloud code and all that stuff. It's pretty sweet.

Todd Geist (47:51.899)
Anything more to say about the file maker thing there?

Ernest Koe (48:02.51)
I think we have...

I think in not so many words, we're actually arguing for a different way of thinking about what FileMaker is. I think we spent most of our careers shifting FileMaker from a database to an app builder and application builder concept. And we've seen our jobs partly as building apps in FileMaker that uses its inherent built-in database.

backend, if you will. But I think that paradigm is not that useful anymore. I think it's more useful to almost revert in some ways to say that file makers really now that container and that database layer, that backend, and your apps. I'm not even saying your apps are UI or a beautiful web interface or web experience. It's that that interface could just be purely agenda. It could be headless. It could be something else.

But that's not the target that we care as much about these days. And it's true. You might question, well, if you can do all this, mean, super bass is pretty easy, neon's pretty easy, in a genetic world, why would you do that in FileMaker? I think it would be disingenuous to sit here and say, no.

It's absolutely easier in file maker still for backend stuff. I don't think that's what we're saying. think just absolutely for greenfield, newfield things, you know, it's hard to beat the flexibility and scalability of some of these other platforms. mean, most of my work these days are done in either SQLite for very lightweight things or SuperBase for some heavier things or for other applications of Neon or Convex or something like that. Right? I mean, this is our experience.

Ernest Koe (50:09.314)
But I certainly think that there's still a lot of value in extending the runway and lifespan of existing apps, especially those that have been built in FileMaker, to take advantage of this paradigm of how to change the development mindset and the process.

Todd Geist (50:27.731)
think that's true. think we should talk about like, you know, the mild form of creative psychosis that can take over when you're, when you get into this and what you realize is like, you can do so much with these new tools that you just start. mean, I think it, I've heard people say, and not just me, but it's a good two months of just like doing everything I ever thought of ever doing.

Ernest Koe (50:35.349)
100%.

Todd Geist (50:55.411)
And it's still some of that going on, but eventually you start to recognize that now that you can build anything. It used to be that if you built something that was had that did something of any value, then you had a thing and you could, you could actually sell that thing probably, right? Like it was the building that was hard. So if you got it over the finish line and it could do ask, then you had a shot at selling that. The problem is that now.

Ernest Koe (51:17.324)
It was a big deal. Right.

Todd Geist (51:24.583)
building the thing is easy. It's like, it's approaching really easy. There's still things to learn as we've been saying, but it's like, there's no moat in having working software anymore. And this is why you're seeing all of the stocks in the SaaS ecosystem crash because again, the hard part was getting the thing that they could sell to people. Now that's easy. what are they gonna, it, know, prices, something's gonna change. Like the cost is no longer the same. So you can't support those.

those dollar values that you're getting for proceeds and things like that. So the question is no longer how, the question is what. And this gets to like the really interesting question of taste, which maybe we can touch on a bit. But before that, I want to get to like what's valuable for your business. And because you still need to generate some return on investment for the work that you're doing. Now, I think at first, if you haven't crossed this bridge yet,

do whatever is fun, just do it. Get through it by doing anything and everything that you can think of doing. But at some point you're gonna return back to, I need to actually make some progress. I need to actually, you know, not do all the fun side projects, but the things that actually matter to my company and to my business in some way. And so one of those things might be like, yes, we could rebuild our entire system in something other than FileMaker. We could do it.

It looks like it's possible, but we've not done these things. We don't really know how long it will take. It will take some amount of time and you will learn some things the hard way and that ROI will eventually come maybe, but it's gonna be not clear how long it's gonna take. Or you can plug into this agentic stuff right now and start extending the system you're in with a new layer of UI or a new layer of

of using MCPs, another we didn't even talk about is MCPs, which is a whole other thing we should get into. But you can start to extend your current file maker system into this new agentic ecosystem without having to do something which sure you could do and maybe it's the right thing to do, but it's not clear yet. And so this is a good first step. It's a good way to get going. So that's one thing I think seems pretty obvious is that

Todd Geist (53:52.391)
You might have a really good file maker system. You may not want to get rewrite the entire thing. There might be other important things in your company that you need to do that you should get those done instead of rebuilding the whole system and just extend it with this kind of thinking.

Ernest Koe (54:10.19)
I think that feels right to me. think that there's a lot of hubris with some of these ideas that I think we're kind throwing around that I want to sort of remind listeners and people out there to sort of pay attention to a little bit. I think we're going through a little bit of ourselves. Just because we can doesn't mean we should. And that's got to be a factor.

That's got to be balanced by assessment of risk, where your time needs to go, what's more important. And I find that that's increasingly like the most important question.

Todd Geist (54:50.651)
It is, it is for sure. Because you like, what happens once you get AI-pilled is you suddenly everything seems possible. And it certainly is more possible than it was. Very true. But there's a long way between, you know, between it being possible, more possible than it was and like, is this the right thing to do?

Ernest Koe (55:02.466)
I was having a...

Ernest Koe (55:12.014)
I had a funny conversation this morning. I scheduled some time with my electrician, great company, and they do some work for us. And he came over and he was like, you know, we were talking about my little network closet there. And I was telling him, you know, like, I really need some help, like getting some of the wiring here sorted out. Because I've got fiber, I've got like drops that are kind of over the place, know, blah, blah, blah, blah. And I said something like, you know, look,

look, Fred, I know I can do this myself. I've crimped wires and do that. And he, without a beat, just shot back and said, yeah, but your time is more valuable doing what you do. Like, you I can do this part. You know, that's what I'm good at. I was like, yes, this is the agentic world that we're in. Just because I can crimp cables and I can lay out a network closet and I understand how these things work doesn't mean I should be doing it because I can do other things that are, you know.

potentially much more valuable to me and to the business and whatnot. This has always been true in a way, but I think it's become really heightened for us who have a builder mindset because we think our job is to do everything ourselves. I think our job is to, I'm going to build a database, not just a place where you store records, but literally rewrite SQLite from scratch because you can.

Todd Geist (56:21.971)
Yeah. Yeah.

Todd Geist (56:33.267)
SQLite has been rewritten like several times in the last year because people are just doing it now, know, for fun almost. But yeah, it's that's very true. I similar story with my networking. I used Claude to help me figure out how to build my network out, you know, and it's certainly doable. I was able to get it done, but was it the best use of my time? I don't know. So I think that's one thing, you know, it's pretty clear is that if you've got a good file maker system.

Ernest Koe (56:41.279)
That's right.

Ernest Koe (56:51.48)
Right.

Todd Geist (57:03.495)
You should at least consider not checking the whole thing out, but starting with this, with this, you know, F layout mode, getting into it that way, or look into MCP servers and things like that.

Ernest Koe (57:08.76)
Yeah, I mean...

Ernest Koe (57:12.436)
Right. I will go so far as to say that...

FileMaker as a platform and let me see if I get this right. I think in an environment where people are empowered to ship and iterate quickly and deploy and you want to create a space where the foundations, sort of the paved path for how you do that and how you share.

systems and applications and products safely within the organization and at scale, FileMaker is still a really great place to do that kind of work. You can encode a lot of business rules, business practices, security policies, and all that into a battle-tested runtime, basically.

And that's not a small thing. I'd like it to do more, for sure. I'm hopeful that there is a universe and a world in the future where it kind of grows up and supports this at a deeper level. But even today, I think there is still a lot of upside for why you would deploy. And I think I'm very comfortable making a case for why some apps should still continue to be file maker apps, even if it's

built with a agentic process with WebViewer and web front ends and stuff like that. But I don't think I can credibly sit here and say that all cases are like that. I mean, this is one of the dangers. I think there's going to be an explosion of things, and we're already seeing this, and the demand for systems that work agentically and work in this new way.

Ernest Koe (59:06.97)
I think is going to be an order of magnitude bigger than however, whatever we've imagined in the past. I'm hopeful that's the case. I think in that space, there is a lot of technology and a lot of stacks and a lot of systems. I think we have some opinions on what works and what matters that we want to bring to bear all kinds of problems that businesses have and all kinds of problems that people who want software.

have, I think in that space there are still a category of problems that could be really and maybe really, really great file maker problems to solve, right? For cost reasons, for security reasons, for all kinds of other reasons. But I don't think it's like, I think we have a habit of saying that everything's a file maker problem in the past, which I don't think that's happening today, right? It's not like, hey, we've got a hammer, which was what file maker was for a of us.

Todd Geist (59:56.307)
Yeah, for sure.

Ernest Koe (01:00:03.854)
every problem is a nail, I think that paradigm is definitely gone.

Todd Geist (01:00:08.497)
Yeah, for sure.

For sure. So yeah, so we think you can, you can plug FileMaker into this new ecosystem. I think we think you have to like, it's not, I don't think like just to maybe spend a minute on like, again, we, there are very clearly doubt they're going to be negative impacts that AI is going to have on our, on our society. Like I think that's already beginning to happen. And there's all kinds of things we, that we can't predict, but

Ernest Koe (01:00:19.576)
think you have to.

Todd Geist (01:00:39.987)
and it uses a lot of energy, know, whatever the, there's a lot of objections to it. And I don't wanna dismiss all those, cause they're all real and we need to deal with them. It's absolutely true. As a...

civil society, we need to deal with them. However, if you are in the business of making software, building systems for people or organizations and you think you can ignore AI, I just don't think you can. I think it's been very, very, very hard to not use these tools. They are very good. They're getting better.

you're gonna, it's gonna be very tough for you to compete in this era unless you are fully engaged as much as you can be in this new way. And that might hit some people like harshly or maybe make them angry and I can understand that. It's real, but it's also, I just don't think you can avoid it. Like I think this is one of those moments in time where you kind of have to get on board.

Ernest Koe (01:01:53.71)
I would even perhaps frame it as there is never a better time to get into this problem solving space because the things that you can do that were previously cost prohibitive before, the barrier of entry for that has dropped dramatically. The economics are changing, but they're changing for everybody. It's not like, they're changing. And the benefits may be still a little bit uneven on the overall AI space.

But I can say that for sure the economics that we've been operating before, even the simple idea that it used to be for sure for most people much cheaper to rent than to build. mean, that's obviously the floor of that has come down dramatically. I don't think it's zero. There's still a lot of problems and I think still a lot of needs that need to be addressed. But for sure, all that's upturned. So in a way, you can say that the market for

our skills and what we do has just expanded dramatically. Just dramatically.

Todd Geist (01:02:56.999)
Yeah, yeah, that's back to the Jevons paradox. There's gonna be way more custom software written today than there ever was before by an order of magnitude or two. So we're talking a hundred times more. That means that people who have experience building are going to be in demand for some period of time, you know? so the opportunities are huge here. we like to focus on those just because, you know,

I wanna, because it is fun, it's really fun to do all this stuff and we're trying to have a positive impact for our customers and for our employees and for ourselves. So we don't, we're, and we just don't believe you can stop this thing. We think you gotta get on. And also that it's really fun to be on and it's really empowering and feels like there's so many more classes of problems that you could address. Like you may need, for example, like maybe you need a desktop app.

for Mac OS. Well, guess what? You could do that pretty easily now with, you you can write a Swift app pretty easily with these tools. And maybe that was a key part of your thing that you were gonna do and it was really hard. You couldn't do it. Well, now you can, right? If you need some kind of crazy fast, you know, custom data parser, well, Rust is really good. And you don't have to know how to use Rust. You just have to know that that's the language because it's really fast.

Ernest Koe (01:04:12.802)
Yeah. Yeah.

Todd Geist (01:04:26.899)
has certain characteristics that you need and you can probably get it to work. So that kind of stuff is really awesome. We just have so many more opportunities and it's really, really fun. I think I wanna, so one of the sort of inspirations are the things that I...

that I go to in this is.

Todd Geist (01:04:51.739)
is actually the Red Mars, Green Mars, Blue Mars series by Kim Robinson back in the nineties. And it's a great series. I highly recommend it if you're into sci-fi. But one of the things out of all the many cool things that are in there is when they need to make stuff or build stuff, they just talk to AIs. That's what they do. And in that world, the AIs can also like make robots and the robots can make other robots. And so you get this meta programming thing and that's what they do.

And there are still people, there are builders there who are the experts at it. And they, they're the ones who can like, you know, get a lot of stuff done by just talking to AIs. And I find that book very inspiring and we kind of use it as a, as a, yeah, where we hope things go, and at least in terms of how AI is, how AI is useful to us. Yeah.

Ernest Koe (01:05:36.258)
was a road map.

Ernest Koe (01:05:43.968)
Yeah, some people read business self-help books. We read science fiction.

Todd Geist (01:05:46.993)
Yeah, we read sci-fi. It kind of feels like that's the right vibe for this though, right? Like this feels like sci-fi in some ways and I think it is.

Ernest Koe (01:05:52.59)
And now it is. totally is.

It's a great book. I highly recommend it to anybody, mean, our entire leadership team.

Ernest Koe (01:06:03.906)
is reading it. But I think it does start in 2026.

Todd Geist (01:06:05.607)
And it starts in 2026, which may be a coincidence or maybe not, but anyway, yeah, so.

Ernest Koe (01:06:13.42)
Yeah, I think, I mean, there's lots of parables in there and lots of stories, you know, but one thing that stands out is that...

You know, one of the core thesis of the premise of this book is that there's 100 colonists, like the initial 100 that go to Mars, and they set up this whole thing. But they all have different roles. And they overlap. They cross-pollinate. But they all have different skills and strength. And just on the personality side, on the technical ability side, there's the super engineer, and the biologist, and the geologist. They all have this tapestry.

interlocking roles. yeah, but 100 of them with AI and robots, basically stand up this whole terraforming project or aeroforming project. But I think the stand out to me also is that differentiation as an idea used to be really expensive. It's not so much anymore, at least in the domain of technology and software.

Todd Geist (01:07:01.745)
kickstart, the whole thing. Yeah.

Ernest Koe (01:07:22.112)
And I often think that I think sometimes we dismiss this because we think, it's just programming and software. But I'm reminded that programming and software is what runs probably more than 75 % of the economy. So at some level, it touches something that people are doing at some level. And not just in terms of business systems, but everything that we do has something running in it.

And I'm forcing myself to think even more expansively. I think the problems that we're solving is maybe not even just software problems. We should blur the distinction like they do in the book, Mars. They think about not just a software program, they think about robots being the thing that is the interface between the environment and their systems.

Todd Geist (01:08:15.731)
Yeah, they have to figure out how to make giant tent cities. That's their problem. And that's solved with software and with hardware and solved with AIs.

Ernest Koe (01:08:19.598)
Intensities,

material signs and genetics, all that stuff.

Todd Geist (01:08:28.211)
all that stuff. We just wanna keep it positive as much as we can, cause that's fun. That's the way we're trying to approach it without dismissing the real concerns. And we keep an eye on that and try to do the best we can to mitigate whatever we can in terms of its impacts on our people and our customers and stuff like that. But it's a ride, man. It is a ride and it's a lot of fun and it's just getting started.

It also looks like no one company is going to be able to control the whole thing, which was a real big concern. So, you know, this idea, this idea of a loop is so simple. And there are models that are free that are almost as good as the frontier model. So there's, there's no super moat that somebody is going to be able to create around this capability. And.

Ernest Koe (01:09:06.808)
That's good.

Ernest Koe (01:09:23.086)
Yeah, maybe for, right, I think maybe for a file maker centric listeners and our clients and customers who, you either in the product side of things or in the services side of our house or even on the platform automatic, we should maybe highlight a few things we're working on and can I start? No.

Todd Geist (01:09:44.627)
Sure. Yeah.

So we've shipped a couple of things already in the AI space. So we have Proof Chat, which is embedded AI chat in your FileMaker system. And that's pretty cool. We like that a lot. And then we also shipped MCP servers, which we didn't talk about too much here, but these agents need access to tools. And one of the ways you give them tools is MCP servers. And so if you have a FileMaker server running AutoFMS,

you have MCP servers that can control both the server and can call into your file maker systems and interact with them via scripts that you expose to the agent. So that was kind of our first two things we did to bring some of this agentic program and this agentic power that we get to the platform. And then we also did the Vivex thing, which is also in Proof Chat, which you can actually...

Ernest Koe (01:10:35.244)
layer.

Todd Geist (01:10:41.981)
five code, some file maker. I wouldn't say it's great, but it's at least a step in that direction. And so that's, that's what that was sort of really all that was shipped last year actually, it seems so long ago, but we are working on a bunch of new iterations on that as we learn this new stuff that's happened in since Christmas break really, and all the different things we're doing. We didn't really talk about local agents, which is another big giant thing that hit.

Ernest Koe (01:11:09.72)
All right.

Todd Geist (01:11:10.323)
that we're doing a lot of stuff with. But we have new versions of MCP servers and tooling that you can plug your AI coding agent into that will give it access to your FileMaker system so that it can learn what's in there. It can learn the layouts, the tables and fields. And then when it generates the code, will already be matched up for that. it's kind of the context part of that harness is the

getting the context of your application to the coding agent so it knows what to do. And then also the deployment side so that you can actually then deploy that generated artifact back to FileMaker. It's one of the things that we're working on. And looks like it's gonna be very doable and kind of help to close this loop without having to go through a lot of the hard learning that we did over the years to make this possible. And things like MCPUI, where you can actually

If you're in Cloud and you're connected up to your FileMaker app through an MCP server that's got UI support, it can actually project back into Cloud or ChatGPT UI. So it can actually show you a form, like a create contact form or a button or something like that that the user can hit to drive something happening in the app.

Yeah, and then what else? What else we got? What am I forgetting?

Ernest Koe (01:12:39.582)
I think on the applications side in terms of AI applications, I mean, we've done quite a lot of exploration along the lines of the same themes that we just talked about. We have autonomous agents running business processes, even doing things like checking for block posts and broken links and doing some of the automations that we have. I think we're trying to understand

more deeply sort of the intersection between the human and systems, if you will. And I think that's been a really fervent sort of area of, know, college research or exploration. But so far, especially in the last couple of months, it seems like there's just some amazing things, like, you know, being able to write your own personal local Asians that do stuff for you. Like, I will never do time differently ever.

do my time tracking myself anymore because we have, know, it's early automated, it's happening in the background. And this is actually a cool story because I think in that whole exploration, a lot of things that came out from there were like, yeah, you know, building memory systems and, you know, it's important, you know, building persistence and how you architect that and how you allow and then sharing that with the team and being able to scale that out. I mean, those are the kinds of foundations, stuff that you still need to know to...

Todd Geist (01:13:39.879)
Yeah, it's totally automated now.

Ernest Koe (01:14:07.672)
bring something like this so that it can operate at some scale, aside from just a little toy thing that you do for yourself. I think memory is actually really interesting thing. And maybe we'll release this, but we built a tool called Teseract, where basically it tries to be a privacy-first way of capturing all the activities as you work during the day. And so you don't need to know.

what you did anymore. The AI knows what you did. It can pose your time sheets where you, know, sort of mostly automatically and stuff like that.

Todd Geist (01:14:44.627)
It works great. It's fantastic. It's like, you know, the, solves a number of problems. One is time tracking, which we have to do, you know, that's how we get paid. But the other one is like, you know, I'm almost 59 years old. My memory is not what it used to be. And now I can just ask Tesseract, which is connected to Claude. can just say, Hey, will you check my Tesseract for, I was looking at a website. It had something to do with this. I don't remember the URL. Find it.

And it was sometime between these dates, like two weeks ago. was two weeks ago, something like that. And it will find it will find the URL. It's always a here's the website and that right there. my gosh. Like that one feature is such a big deal for me because I, I tend to read a lot of stuff on the internet and I, I'm not capturing it all properly or, and you know, it's hard to do. So this is just capturing that for you. And then you can just, again, right within.

Ernest Koe (01:15:25.582)
You

Todd Geist (01:15:41.267)
You can just say what what did I do today and it will give you a summary based on what it saw you do on your screen And it's pretty darn good And then you just you know, then it's a matter of hooking up to a time to a time tracking app and we have that all set up So that's how we do time tracking now

Ernest Koe (01:15:46.37)
Right.

Ernest Koe (01:15:59.586)
Yeah, I think two things stand out in all that. One is I'm finding that we're writing much more targeted, small things that are agentic AI first. And so I think there's an old way of doing things where it has to be this big app that solves all the problems in once because your users are sitting there. But if you start with the proposition

Todd Geist (01:16:11.837)
first.

Ernest Koe (01:16:28.684)
or the premise that wherever your assistant's going to be, that is the center of the locus of the work. That's your core UI and everything else is just in service of it basically, fundamentally. Then the whole thing kind of changes. But also, these layers that I think we've been doing quite a bit of work on, where it's intersection of memory and persistence.

the infrastructure to allow these things to be deployed and scale correctly and the agentic loops and the harnesses that you have to build so that you have a self-referential learning loop that's happening that can kind of solve problems for you. I think that mixture of capabilities is really sort of the next frontier that we think are going to matter for our businesses and other businesses too, right, other customers.

Software is almost that's the table stakes. And then everything else is how you make the business of software agentic, both in the delivery aspect, but also in the application aspect. mean, so exactly. All right. All right.

Todd Geist (01:17:38.961)
maintenance and upkeep and all that stuff, right? This is where the automatic platform has served us well and that we have a lot of infrastructure that we're running and monitoring and keeping track of. so, you know, we can see how in the future we're going to be deploying fleets of agents to our infrastructure that will be doing all kinds of things. Monitoring file maker servers, for example.

reading our email, know, the local agents that can read our email and tell us what to respond to and what not to, and all this different stuff. Infrastructure is part of it. In fact, one of the key things to really get in this whole thing is how the internet works, right? And that's like, I think if you're gonna do this stuff at scale, that's one of the things you have to come to grips with. And FileMakers protected us from that for a long time, but not entirely. You know, if you did API integrations or things like that, you discovered it.

So yeah, it's, and again, it's like, these are all open standards that were, that are being built on here. So it's worth it. Like I think if there was only open AI and they had the only model that could code, that would be worrisome to me. I'd be very concerned. That's not the case. have four frontier labs that are producing, four or five frontier labs that are producing really powerful models and they are competing like nobody's business.

And then you've got the free stuff, which is actually just about six to 12 months behind in terms of its capabilities. So, you know, that's the kind of stuff that you want to run on. So we have the infrastructure to run a free local model if that's something that you need because of privacy or you want it cheaper. Those kinds of things come in handy.

Ernest Koe (01:19:26.446)
The cost of that is also coming down dramatically. It will take some time still. I really big models are still pretty expensive to run. I'm always, I saw, I think it was yesterday, maybe today, that Apple released its Python bindings for Apple Intelligence, which was really cool because you previously could almost only do that through Swift to get some of the cool things that's closer to metal.

But stuff on device and stuff on the Mac, the fact that your Mac now can basically run a, know, like a Quinn 3.5 model, you know, to which rivals like OPE has 4.5, you know, just a couple months ago. I mean, that stuff is pretty wild. So I think there's still a lot of innovation happening.

that I'm pretty excited about. And I don't think where we are, where things are going to be in six months, I think would be, I don't know, I think it's gonna be kind of wild. I'm kind looking forward.

Todd Geist (01:20:31.195)
It is I want to. I want to just reiterate the edge stuff you're talking about on device stuff. I think that is like for a long time. You know if you wanted to do on device, you're basically playing in the App Store game, which is like it's a thing. It works whatever, but it's also like you know a million billion people in there. So we haven't paid attention much to the device. Cause we had to run on cross platform stuff. But now you've got you've got AI deployed on device like.

Ernest Koe (01:20:36.812)
Yeah,

Todd Geist (01:20:59.411)
Yeah, just the other day from my local agent, I wanted it to be able to take PDFs and parse them and store them so I could search them. And the OS has stuff that you can use to do that. It's just sitting there. The command line tools are stuff you can easily install. And one of the things I got to use, you again, this is like this would have blocked me is I wanted to use Apple Vision from a TypeScript app to

Ernest Koe (01:21:05.176)
Yeah.

Todd Geist (01:21:29.235)
to be able to take a PDF that was scanned images and not text and go image by image and basically turn it into text. Turns out that it's really easy to do with Swift and it's really easy to call a Swift file from TypeScript and I didn't have to do really anything except ask. Like, isn't there something on the OS that, can I use Apple Vision?

Ernest Koe (01:21:42.158)
It's really easy to do with Swift, exactly.

Todd Geist (01:21:54.771)
It was like, yes, here's how you do it. Here's the little Swift wrapper around Apple vision. can call this from TypeScript and boom, now you've got this and it's running on device. Like it's right there. So this whole new, it feels like there's a whole new frontier opening up, which is on device stuff, you know? And that was like, that was one of the other crazes we didn't really touch up. Maybe we can end on it, which is the open claw thing. It started out with Claude, Claude bot, and then they got.

Ernest Koe (01:22:18.433)
Right.

Todd Geist (01:22:22.387)
They had to change the name first to Moltbot and then now it's now, now it's OpenClaw and it is the fastest growing software project in history. And it is now impacting the it's impacting Apple sales because one of the ways that people run OpenClaw, actually probably one of the most popular ways is to run it on Mac minis, which explains why I have four. So you basically give your agent a whole computer and you say here.

Ernest Koe (01:22:45.134)
You

Todd Geist (01:22:52.177)
Right? And this is actually super valuable. I don't actually want Open Cloud running on my main computer for lots of reasons, which I hopefully make sense by now. But like, this is a whole new frontier, another exciting area for us to get into, you know, things that would have been very, hard for us to do before.

clock can figure it out, cursor can figure it out.

Ernest Koe (01:23:12.844)
laughing because I was too chicken shit to do open claw because I was scared that he would just basically wreck my life. So I wound up writing my own agent to do this. you know, it's it's it was it was I think I think the open claw and just just a two second thing on this. There's so much to talk about. Right. But it's sort of affirms like this pattern that we've been seeing. Like, you know, I think

Todd Geist (01:23:19.923)
It has, Rex.

Ernest Koe (01:23:42.03)
I think there was a story that in order to get into this, you have to have a lot of infrastructure, a lot of big models, lot of like LLMs, a lot of this, a lot of that. And it's all inaccessible except for the big players and the frontier people. And OpenClaw sort of kind of painted a way of doing this that I thought was sort of emerging, but they crystallized it in a very...

real way. some of the big themes like command line things just work because agents know how to read docs and they don't need to tokens on loading into concepts. They can just call bash scripts to run things and run command lines and like agent skills. So fundamentally all the command line tools that we've had from day zero of the internet all of a sudden become valuable and you don't have to learn them.

Todd Geist (01:24:38.545)
And you don't have to learn them because the agents already know them. They can figure them out. No problem. So you're seeing things like I watched this video the other day on some social media. I don't know. was a young mom with a baby. She's like holding a baby and she's talking about how she has gotten her open-claw agent to shop on Amazon for her. Now she just literally takes a picture of what she needs, sends it to open-claw and then it's ordered for her and shows up at the house. So yeah, it's pretty wild.

Ernest Koe (01:24:42.305)
Okay. Yeah.

Hmm

Ernest Koe (01:24:58.892)
Right.

Todd Geist (01:25:08.645)
And it also has that feel of the open-claw thing has a feel of like the early web when it wasn't controlled by any like, you know, there were no big, big corporates in it. It was like this weird thing that came out of nowhere and suddenly people were doing it all over the place. And that's kind of what happened with open-claw. just kind of came out of nowhere. And it is now this huge open source project with the whole foundation behind it now and hundreds of thousands of people creating skills and extensions for it.

It's pretty amazing. Pretty cool.

Ernest Koe (01:25:41.45)
It's pretty amazing. I'm not telling people to run out and to do this and put their production systems on open claw.

Todd Geist (01:25:45.683)
Do not don't put your like get a mini or you do it online. You can also like you you can host it online. You can you can get an open claw hosting platform if you want and that's a safe way to do it. You know, do something else, but it's so fun like it is just really, really fun. To have an agent and you know you get to come up with with with fun names for them and give them back stories and you know all kinds of fun stuff so.

Ernest Koe (01:26:15.448)
Well, I know we have lots more to talk about, but...

Todd Geist (01:26:16.211)
I don't think we probably, yeah, we're an hour and a half in, I don't know, this could cut down to something shorter than that, but it'll still be pretty long. But I don't mind, like I think like this time is a time for lots of ideation and rambling and, know, this is like one of those explosion periods where you're lots of exploration going on. so we're gonna try to do this more often and just talk about the things that we've been exploring and things that we found interesting.

and keep encouraging our community, which is FileMaker people for, I don't know, 25, 30 years, something like that, to explore this space and to not sleep on it, to get into it, keep going. You're a builder. If you weren't a builder, you wouldn't be using FileMaker. And this stuff is fundamentally like the best thing that's happened to builders since...

Hypercard or FileMaker or something like that. any other closing thoughts?

Ernest Koe (01:27:18.446)
percent.

If you liked this, know, give us a like throw us a subscribe whatever share it with your friends and I think

some more routine thing as we

and even just share out what stuff that we're working on. So hope to do more of this and hope to see you all next time.

Todd Geist (01:27:41.095)
Sounds great. See ya. Bye.

‍