Tyler (00:00)
In this episode, we talk about harnesses, agents, orchestration layers, and other terms that mostly mean the same thing. Let's go.

Rick (00:26)
What's up Tyler?

Tyler (00:28)
going on. We're doing a Tuesday episode. I can't remember why, but yeah, off cycle. you have yeah.

Rick (00:35)
I think I have a dentist appointment on Thursday morning. Yeah, and by the way, by the way, our time

is no longer gonna work because I take kids to school every day of the

Tyler (00:46)
okay.

Rick (00:46)
week now. So we need to think about that. I'm gonna talk to Sable about it on Sunday and see if if this time will if we can like carve out that time still. But that's

Tyler (00:55)
Mon Monday,

Tuesday, Friday work better for me than Wednesday, Thurs like we we normally record on a Thursday. I I can make whatever

Rick (01:03)
Monday, Tuesday, Friday?

Tyler (01:04)
yeah, those are when I'm at home. and it's just well, especially big news for me that is so in the weeds you couldn't possibly care about it, but I'm getting a dedicated desk out in the open office of less annoying CRM. So I'm gonna be out out in the the public space a lot more. not good podcast recording.

Rick (01:21)
Gick Wha why are

you doing that?

Tyler (01:24)
I'm keeping my private office too. I just I have a leaders eat last mentality, meaning everyone at the company gets a private office. I have the worst one. it's there's no windows. It's just very isolated and not a great environment. it's only two days a week and I'm in meetings half the time anyway, so it's I'm not like in there that much, but you know, I j I just like to get out there a little more and and be around people and also get some actual sunlight.

But I still need a private office for podcast recording. I record a lot of Loom videos, stuff like that. You know, you wouldn't want to do that in an open office.

Rick (02:03)
So basically, you're going to move your primary desk to the open office and then use your dungeon office only when you need it. All right, cool.

Tyler (02:11)
Yeah, yeah, exactly. Also,

I I mean the main thing I use it for is taking naps. Jaboy does like to nap.

Rick (02:17)
Does it have a couch in there for you?

Tyler (02:20)
It's not big enough for that. Th all of the offices are like they're more like call rooms, like they're very small. but I have a little like inflatable hiking mattress that

Rick (02:30)
My goodness gracious.

Tyler (02:31)
it's actually great. You know, I I just put it down and take a take a quick nap. It's it's a perfect.

Rick (02:37)
I've never been able to do that. I admire your napping ability.

Tyler (02:42)
I love it. what's

Rick (02:44)
What's going on?

Tyler (02:46)
yeah, not much. I've you know, I've got a few random little topics, but I feel like overall things are are pretty calm. I'm in that, you know, in the ebb and flow of of things, I'm I don't know what ebb means and what flow means, but I'm I'm in the part where I'm knocking stuff off my backlog and actually kind of getting close to a point where I have to figure out what I need to be working on, at Less Annoying. Like the you know, I it's it's almost

Easier when you have a bunch of kind of unplanned hair on fire work. Not that that stuff's fun, but you never have to sit down and decide what to work on because like it the the work decides itself. I'm I'm approaching a point where it's like, if I have a week free with nothing to do, what sh what should I be doing right now?

Rick (03:29)
Cannot

imagine.

Tyler (03:31)
I probably won't actually get there. Like something else will happen before I get through my backlog. But I'm, you know, every my my workflows every day, like I use Gmail as my kind of task list. I snooze very aggressively. So every morning I get like 20 things unsnoozed. I look through it like first thing in the morning, I go through it and I snooze, you know, 18 of them or something. So I'm I'm down to an inbox of two things. And then I that's what I spend the day on. But I'm getting close enough that.

I'm like, okay, there's only like six things left, you know, if if I get three more uninterrupted days, I'm I'm down to zero.

Rick (04:05)
What inbox tool do you use? Do you have multiple inboxes of your emails coming into one inbox or do you have multiple inboxes that you're snoozing across?

Tyler (04:13)
I probably should change this, but my approach is like years ago when Less Annoying CRM first started, and I had multiple projects going on. I forward all of my email to my personal Gmail. So Leg Up Health goes there, the Yummy Life, which is my mom's blog that I help her out with, less annoying, a few other email addresses, all go to my personal Gmail account, and then I just use that one inbox. It's it's like ninety-five percent less annoying CRM email, but yeah.

Rick (04:41)
So you're using a personal Gmail account to manage multiple accounts. Cool. That's awesome. You don't use front?

Tyler (04:46)
Yes. What do you you just have

different pr we only use front for support emails. front does not have I love front as a collaborative inbox tool, but I don't like it as much as just like a personal I wanna like snooze stuff and set up filters and all that.

I should actually well, sorry, no, I I just I'm trying I the reason I'm hesitating is I'm like, why don't I use front? And there's a better reason for it. I have a an automation setup. Any email that I send from front includes this little hidden code in the footer that if then the person replies to me, it goes into our group inbox. So basically, if I want to email somebody and then let the support team take over, I send it from front. If I send it from Gmail, it stays private.

But I do like front. I'll I'll put some thought into that. Yeah. Anyway.

Rick (05:40)
Yeah. Yeah, I just recently

consolidated a couple of my inboxes on front away from my personal Gmail. So I just did the I moved I had the same setup as you where everything would go into my R Lindquist one one at Gmail and I switched everything. Yeah, well that one is my spam catcher. So it's

Tyler (05:52)
sure you want that to be public, Rick? okay.

Rick (05:57)
it's the anyway, every every it's it's the the one that catches everything. And so then I moved all my non R LindQuest one one at gmail dot com email addresses.

Tyler (06:07)
Gotcha.

Rick (06:07)
to another

to another inbox that has like high signal. And that all cut that all goes through that all

Tyler (06:11)
So you're just not gonna check that one anymore.

Rick (06:13)
goes through front in multiple. Like it catch it catches everything.

Tyler (06:15)
okay.

Rick (06:16)
And then it but it's it's like four or five inboxes.

Or

Tyler (06:19)
Interesting. Yeah, I never

really thought about using front like like Michael, our head of customer service, has front set up for all kinds of fancy features for the the support inbox, the shared inbox. I guess I guess all those features are available to me individually too. I just never really thought to use it that way.

Rick (06:34)
And you can connect as many Gmail accounts as you want. It's kind of nice. well,

Tyler (06:37)
Hmm. Hmm. Okay. What's going on here?

Rick (06:41)
let's see. I I I the big update I have is that I feel like I've hit a milestone with AI, like about this a little before this time last year, maybe 14 months ago, I definitely woke up one day and was like, my god, I'm behind. And I feel like for for for the last 14 months, I've just been consuming information every day, like every week trying to catch up and like.

s at some point in the last month, like I've started just like the the matrix became readable to me a little bit. Does that make sense? as like all of a sudden the c

Tyler (07:11)
Mm. Yeah. Well er

I've observed this of you in our conversations and like the listener hears these podcasts, but like yesterday we had a like a health partner meeting and yeah, like you you are displaying expertise that you definitely were not six months ago.

Rick (07:28)
Yeah. And so I've I've kind of transitioned from like learning mode, like just trying to understand like what is that thing? What is it what does that word mean? I understand what that word means, but now I just I don't understand what it actually does to, I understand the language. I'm now in like application and experimentation mode and

Tyler (07:48)
Mm-hmm.

Rick (07:48)
actively building. And it's just a way more fun place to be. it feels like I'm progressing ideas.

with the tools and learning more as a result of that versus just trying to figure out how to unlock the tool. I don't know. That's the milestone I wanted to share. And I just wanted to if you ever had this moment,

Tyler (08:05)
Yeah. Well I

absolutely and I I wanna like yes and you here, like from what I'm observing from you, which is I it's so easy for people in both of our positions to be good at talking about something, but not really ever do anything with it. like I your job at Windfall, I don't really know all the details about it, but I'm I s like you oversee a lot of people, you're an executive. I bet you could just go in there and sit in meetings all day.

And like more or less do your job. But like for me, that is an unfulfilling life. And you need to be an operator. You need to be like in the trenches. Even if you yes, you are a leader, you're a manager, you oversee people, you still need to like be an expert at doing the thing. And that's kind of what I've observed from you. I feel like is, you know, three months ago, and myself as well. I haven't made the progress you have, but like three months ago, you and I would have a conversation about AI and it'd be like, yeah, we we know these terms.

We can talk about it. We can talk about how one could use this. And I feel like like last night in our meeting, what I saw from you is you're like, click this thing and Claude. Now click that thing. Now go over and and it it's much less conceptual and much more like a person who has experience doing this is showing someone else how to do it.

Rick (09:20)
Yeah,

thank you. Thank you. so yeah, that's fun. kind of translating that into actual like updates. Like I've got multiple different projects going now. some related to leg up help, many related to windfall, some just personal projects. I'm trying to push the limits across all the different ways to do things, cloud being the primary sort of let's just call it model and and and interface. but yeah.

Tyler (09:48)
See you f you fell behind already, you're still using

Claud.

Rick (09:51)
Yeah, there you go. I I I just like to me the the the utility in like switching models and optimizing models is so low relative to like stretching towards like what it can actually any of them can do for me. and unlock like that's where I'm focused right now. So I could care less, like unless unless there is a really good reason to you switch to a different like model. I'm not not not like gonna even worry about that right now. but like

Like a couple of just examples of things I've got running. I think there's some really interesting sort of scheduled tasks and routines that you can build with the connected tools within Cloud now that that are automating aspects of my job that and other people's jobs that you know save a lot of time. and you it's just that's cool. Like examples are, you know, you used to have to go review a report and then

Go through it and say, this report is telling me to do something. Now you can point Claude at the report or the a query that has the effect of the report and tell Claude to action that. And now you're moving from like having to do work to figure out what to do to Claude does the work to tell you what to do and actually can start doing that work for you. the I mean a good example of this is like if you've got a pipe in sales, pipelines exist and you've got you know.

Closed dates to per project forecasting. A big job of RevOps is making sure the forecasting's good. You can now have Claude go poke at the forecast reports and call out people, shame people, or praise people for

Tyler (11:32)
Mm-hmm.

Rick (11:34)
you know how good their forecasting is. another example of that's kind of like within the Claude sort of UI ecosystem. Another example is Claude has hosted infrastructure for

what are called managed agents. So you basically you could just go in and like through chat build a autonomous agent that actually can do stuff. Like it could be in Slack and respond to people in Slack. It have a mind of its own. so I've got some some experiments running with that related to new higher onboarding and doing sort of more like proactive things you know mostly Slack based but like to to help people

w

that are that are new to a company move through the process. on the personal front, I I I every Sunday I do a weekly reflection and I look at like my I pull up my Apple Health, my aid sleep, my with my my withings weight, like make sure everything sinks over to Apple Health. And I just check my weight, my steps, make sure I'm like on track with like what I'm supposed to do. And if I'm not, I'm like I make I reflect on that. I I set up an app we talked about render last episode as a

Tyler (12:45)
Mm-hmm.

Rick (12:46)
sort of way to host host yeah,

Tyler (12:48)
Like a hosting vibe coded stuff. Yeah.

Rick (12:51)
like you don't have to worry about any of the infrastructure underneath. And basically use cloud code to build a a mini app that pulls the the health data down every every day and then sends me an alert on Sunday. So I don't have to go looking for that. I mean it's saves me fifth you know five minutes, but it's kind of a nice little cool trick. those are the kinds of things I'm working on right now, but it's just the I feel like I'm still barely scratching the surface of what

it can do, but as a result of shifting to this, I'm doing stuff, it's gonna compound into more exciting things, I think.

Tyler (13:23)
Yeah, now so the personal stuff is great. I should do more of this. Again, this is something you know I talk about, I think about how many times a day do I have the thought I could vibe code something to do that, but I don't I don't actually do it. one thing I like about like you showed me your app for showing your fitness stuff. And like, yeah, it's very basic. It's not like exactly impressive per se, but it's like you're you're doing the reps so that the next time you want to vibe code something, it'll feel like a 15-minute project as opposed to like a

I don't even know. I have to go figure out how to even like how does it ho how's hosting work and all that? on the work front, do you feel this concern that a lot of people have of like it's you said earlier it's kind of starting to do the work. So what are you here for?

Rick (14:15)
I'm making sure it's doing the right work. I'm I'm managing there's still I think a human sort of workforce AI management task here. we're still very early. Like someone's still gotta be, I don't like saying this, but like the throat to choke, you know, if the AI does something wrong, right? so

Tyler (14:34)
Yeah.

Rick (14:35)
I I kind of look at my my role as like one, trying to create cre like we still have humans doing the lots of most of the work in most companies. And so like if you look at like the work

you know, needed to be done to the human ratio. like my job is to reduce the ratio,

Tyler (14:52)
Yeah.

Rick (14:53)
or sorry, increase the rate rate at what which work is getting done per human, and and and sh while also ensuring quality.

Tyler (15:00)
Well, so okay, but like what it should be is some of the work can be done by AI and some of the work can't, and ideally the work that can't be done by AI is the most important work. We already see some examples where that's just not true. For example, AI cannot pick up a box and put it on top of a shelf, but that's not exactly like high leverage work, you know? what?

Rick (15:23)
It can do that.

I feel like Amazon's got factory

like warehouses that are doing that.

Tyler (15:26)
th there are sorry,

there are very specialized robots that do very specialized things, and there are general humanoid robots that suck and explode every five seconds.

Rick (15:36)
Okay, okay.

Tyler (15:37)
There's not like like hu an Uber drive, I mean, okay, sure, Waymo will replace Uber drivers, but it has not yet. Like there's stuff that humans like a lot of physical what what is a lot of low-paid work cannot be automated with AI yet. And then a lot of previously high-paid knowledge work can be.

So it's there's a there's that disconnect there, but like what you want is it does the unimportant work or the low value or low leverage work for you. And then you, instead of spending 10% of your time doing the high leverage stuff, you spend ninety percent of your time doing the high leverage stuff. Is that actually what we're all feeling? I don't know.

Rick (16:14)
I think it is. I'm I

Tyler (16:15)
Think so.

Rick (16:15)
I'm seeing it. I'm seeing it happen. I I think yeah, and and I also think that the the the other thing that I'm noticing, and I I don't know how long this will take to play out, but people who are previously like gated on being able to do technical things are slowly being ungated, meaning work that they would have to go like, I need to make a request.

And some percentage of the time, like they just don't make the request because of the effort required to make the request. Now they could just go yeah, they could just go do it. and I

Tyler (16:45)
for sure. I do this all the time. Yeah.

Rick (16:48)
feel like that is still like early infancy, but like as that as as people get ungated and actually feel empowered to like act on it, I think that we're gonna see an explosion of let's just say like technical applications of the tool.

Tyler (17:04)
Yeah. No,

I'm not denying the utility in it. One of the things I was getting at is a a thing that I keeps coming up for me is I don't see the use cases for like I obviously see it for AI coding, no question. And I use it all the time as a question and answer brainstorming partner, design partner, whatever. But like I have not really automated or like replaced work I was doing with it. It's more it's augmenting the work I'm doing more so than like, okay, AI, go do this.

You know, every Monday morning. So I never have to do it. But I think that ties to what I was saying of like, I think I had already removed most of my like low leverage work, either by just not doing it in the first place or someone else at the company does it or whatever. So I don't think I've seen a lot of those opportunities. And then when I look at the stuff that I do spend time on, if AI can do that, I'm out of a job. Like I don't want it to replace the stuff I do because the stuff I do.

is core to my identity as you know, as a person at this company. I sure am. I mean, I don't think

Rick (18:03)
you're threatened. yeah.

Tyler (18:06)
it c I it would suck at doing my job right now, but if it got better.

Rick (18:09)
Yeah. Well, I hate that kind of segues into something I wanted to talk to you about. Like, would you be open to like working on a side project with me where we basically vibe code a AI native leg up health platform that actually co like starts to b imp self improve itself that we like a with agents that we monitor?

Tyler (18:29)
I mean, what you just said sounds like a ridiculous pipe dream that I don't believe in, but

Rick (18:34)
Okay.

Tyler (18:34)
I think we can keep talking about something that feel like 'cause we were we started this conversation last night, so I have a little context that the listener here doesn't. Self improves itself like so you just want like the platform to get built by a swarm of agents and we got just sit back and it gets better and better. Like that feels unrealistic to me.

Rick (18:55)
I I I think that's too extreme. So I

Tyler (18:57)
Okay.

Rick (18:58)
I would I would say today something like we can both contribute to improving the product and supervise some like very basic agents that help us improve the product on some regular cadence.

Tyler (19:16)
So, my argument would be that the interesting thing to do is not to work on the product here. And I feel like this has been a recurring theme for the last few years, really, with Leg Up Health, which is what is the role of tech in the business? Because it it's a business that's primarily an insurance advisor, brokerage, whatever. And I'm here, like like pre AI, I'm here as this tech guy who can build custom software. And there's kind of this story we could tell of.

There's lots of insurance brokerages or whatever you want to call it out there. We have software. That's like one of the things that differentiates ourselves. and I think what we've seen time and time again is sorry, let me let me pause. In or let me back up. We always we originally thought that the way to leverage that technical expertise is to build a consumer-facing SaaS app that the the consumer logs into.

To whatever, get their questions answered answered, manage their insurance policies, whatever. I think we have found pretty resoundingly that that is not what and and no one wants a a platform related to insurance. I think the value that can be brought by technology in this business is streamlining the work done by humans as brokers. Now, maybe that's what you meant, but like when I hear product, I think

Rick (20:36)
No, I think you

Tyler (20:36)
like customer facing.

Rick (20:38)
that's exactly what I'm talking about. So and and for this idea at this app two point what the our f our our our beta customer is Legapell.

Tyler (20:52)
Okay, so you're saying build software that could be in theory sold to insurance brokers to make them more efficient. Gotcha, gotcha. Yeah, that's that's more interesting to me. And like my guess is at first there is no app and there is no UI. It's a command you can run in the terminal or something. I don't know how you in initiate it, but it's like I have so well, actually, since our conversation yesterday, I did a little looking into terminology here. I feel like terminology.

Is so hard with AI because we're all saying different things and meaning different things. Can I do a quick little deep dive on what I found?

Rick (21:23)
Mm-hmm. Yes. I would love it.

Tyler (21:27)
Okay. So this was just like talking to ChatGPT and asking it for some definitions. I don't want to act like I'm bringing some novel insight here. But in our conversation last night, just to catch the listener up, we were talking about I I kind of described it as I think tier one of AI usage is.

It's a better Google. I have a question, I want an answer, I'm gonna ask ChatGPT. Tier two is what I'm describing as like what you're doing right now, which is using tools made by somebody else to automate and systematize those prompts that it's not just ask a question, get an answer. It's automatically run this every Monday, connect to an MCP server from somebody, pull data down from one thing, push it to another. It's kind of creating, I'm gonna

I'm avoiding using the word automation. No, no, I can use the word automation. There's a d there's a different word I'm avoiding because it's gonna mean something else in a second. It's an automation that actually does some kind of pre described process. It doesn't have to be scheduled, but scheduling it is just another layer. That's tier two. Before I move on, do you like any wanna clarify any terminology I used or anything?

Rick (22:40)
Can you

c can you go can you go back to tier one and just define tier one, tier two one more time cleanly?

Tyler (22:45)
Yeah,

t t tier one is I'm asking a question to ChatGPT, I'm getting an answer.

Rick (22:50)
Okay, tier two.

Tyler (22:52)
tier two is I'm using MCP servers, I'm using pre built skills, I'm running them on a schedule, some or all of those different tools. to basically get it's going from question and answer to I'm building actual automations and

Rick (23:07)
multiple systems chained together and some scheduled or automated work being done. Okay. Cool.

Tyler (23:12)
Yeah, some combination of those more advanced features. Yeah.

here's my hesitation with tier two is it just feels like hum you have nothing to offer here. Like temporarily you have something to offer. Rick Lindquist has found out how to click the right buttons in Claude to make it do what he wants. But like long term, are you really providing a whole lot of value here if you're just

You're you're basically just giving a prompt and getting a response back still. You it's just like

Rick (23:43)
You're you're still managing automated root prompts.

Tyler (23:46)
Yeah. Yeah. Which it just seems like a year from now the AI will do that. Like you'll just say to the AI, hey, do this thing, and it'll you don't have to click any of these buttons anymore. And your knowledge that's an advantage right now will not be an advantage in the future. so I can't help but the ambitious person in me is thinking there must be a tier three, there must be a way to be like good at this in a defensible way. What does that look like? That's how you create long-term value, not just like short term time savings from AI.

And so I think tier three I will describe as creating your own harness. and this is where I got into the definitions, and I want to know I want to define this as I understand it. A harness is code that combines like deterministic code, meaning you give the same input, it gives the same output. All coding throughout history is deterministic, whether you're using PHP or Python or JavaScript or C. That's code.

And then the LLM is non-deterministic. It's like thinking. There's like this random, unpredictable element to it, but it can solve problems that it's never seen before, whereas your deterministic code can't. So basically a harness is the code that orchestrates so the two two words now that people use a lot. The harness orchestrates the work between your code and the LLM. I have more to say, but let me pause there. Any thoughts, questions?

Rick (25:08)
You can't the listener cannot see me, but I am nodding very

aggressively. Like

Tyler (25:11)
Ha ha.

Rick (25:12)
I totally agree with where you're going here.

Tyler (25:14)
Okay. The final thing I want to add here. So everything I just said, I think I already knew. Where it introduced a concept this morning when I was talking to ChatGPT that I had not thought of before. I said, so maybe Rick, okay, Rick and I should build a harness. Like that's an agent, right? Should we be building an agent? What's the difference between an agent and harness? It defined, it said, no, what you want to build is a workflow. This is the word I was avoiding using earlier. An agent and a workflow are both.

Two things that a harness can orchestrate. The difference is what is driving the decision of what to do next. You've got two parts, two, two modules within the harness. You have the code and you have the LLM. An agent, the LLM is driving the decisions. So for example, coding agent.

The code says to the agent, here are all the things we can do. I can run Linux commands. I'm connected to these MCP servers, like yada, yada, yada. The agent thinks about this and it takes a prompt from a user, like, what do you actually want to do? The agent thinks about it and it says back to the code, hey, you should run, like the user is trying to fix this bug. We need to figure out what file the bug is on. Can you grep the code base for this terminal, this keyword? And the code says, okay, I'm gonna run that command. I'm gonna get the output, I'm gonna give it back to the LLM.

The LLM looks at that output and says, okay, well, what should be next? Let's look at that file. Can you open up that file and send me anyway? Back and forth and back and forth and back and forth. But the LLM is telling the code what to do. A workflow, according to ChatGPT, is the opposite. The code has defined what needs to happen, and it's telling the LLM, here's an input, here's how I need you to process it, and you're gonna give me these results back. And then with those results, I'm gonna decide what to do next. And for what leg up health needs help with.

Which is things like we need to when we're qu giving a group insurance quote to a business, you know, first we need to pull down rates. Like or first we need to collect demographic information, then we need to pull down rates, then we need to create a document with all the rates, then we need to compare it to what they were on before and write an email up with that. But like that is a workflow, not an agent. And so it was saying what you and I should do is start building a harness to run workflows.

Rick (27:33)
I

a hundred percent agree. there are multiple examples of this playing out in the marketplace right now in different ecosystems and industries. one very like relevant one for me right now is Pylon, which is a B2B ticketing platform that is taking on intercom, Zendesk, and Salesforce Service Cloud, and winning for this exact reason.

They basically have built a harness on top of a very some very important workflows to basically own like make make the power of what Zendesk has had historically and Claude has in the f you know, going forward happen within Pylon. and it is amazing.

Tyler (28:22)
You're sorry, you're using this? You have used this?

Rick (28:25)
What

we are we are I can't speak to like the specifics, but like we're looking at it as an option. And it's one of those things where it's like the tool from a demoing and to utility perspective is so freaking powerful. You it's so much more powerful than Claude by itself. It is so much more powerful than the non AI native tool that pre precedes it, like a Zendesk. It's like a no brainer. Now, I do think like which one do you go with? Like if there's multiple pylons, like

Right now it's pretty obvious they're the leader, maybe, but there's gonna be fast followers.

Tyler (28:56)
Well, you think more th more

than Finn? I would have said Finn was clearly the leader.

Rick (29:00)
For B to C.

Tyler (29:02)
Really?

Rick (29:03)
Yeah, they're much more foc they they Finn w was called Finn because they were focused on financial services originally.

Tyler (29:09)
Finn? No no Finn is intercom.

Rick (29:11)
Finn is intercom. Finn was, I believe, named for the financial services vertical. Banks and

Tyler (29:18)
I thought it was saying

end. I thought it was saying we're done. Like we have concluded the

Rick (29:21)
finished.

Tyler (29:22)
conversation. Yeah, th that's how I always interpreted it. I don't I could be wrong about that.

Rick (29:25)
interesting.

Maybe I just totally misinterpreted that. but yeah, like I would put like we we did I've done evaluation across fin, Zendesk and Pylon. And it's very clear for B2B specifically, Pylon is the the winner from a feature and capability perspective. But it's for the reason that you're describing, which is they've basically combined a harness for a workflow and basically giving you all the power of Claude, but for a very specific

Tyler (29:54)
Yeah.

Rick (29:56)
set of conditions and like outcomes

Tyler (29:58)
Mm-hmm.

Rick (29:58)
that you want. Is that that's what you're saying, right?

Tyler (30:03)
yeah, I I wouldn't I I don't want to get too caught up in the terminology here, even though that's what I started with, but like I wouldn't be surprised if by the definition I gave what they're doing has some agentic elements to it, like where it's kind of t telling the LLM, here's the situation, what do you think we should do next?

Like like a workflow has to be super, super well defined. Again, like enrolling a person in in in health insurance feels like, okay, it's step one, step two, step three, step four. We only need the LLM for okay, we have 50 health insurance policies. Can you narrow it down to four for me? And I'll pitch those to it. Like you only need the LLM for very specific parts of it. I do feel like Pylon is doing customer service, right? It there might be more the LLM driving things. I

It doesn't matter. Yes. Like that's what I'm saying is if you control the harness, whether the harness is a workflow or an agent, probably doesn't matter much. If you control the harness and it's custom and proprietary, that is a business. That's a real defensible, valuable thing. Whereas if you are just building things in Claude or ChatGPT or whatever that they're running for you, that strikes me as a very temporary productivity improvement.

Not an actual defensible valuable thing you can sell to people.

Rick (31:29)
So

do want to do this?

Tyler (31:31)
I mean s I am interested in it primarily because I want to know how to do this, not like I think it would be valuable for like a health to have it. that's that's a number two for me, I think, on the on the list. Because I do like we we had our conversation with JD yesterday. He brought so that the the agenda of the meeting is JD bring like open enrollments coming. He brought all the stuff he spends time on, and he was like, Here are the top

five or top ten things that if we could make them more efficient, I would be more efficient and I and open enrollment will go better. Every single one of them can get solved with just the tier two stuff that you already know how to do. I don't I don't think we really need any of this for leg up health, but I do think it would be a a valuable experience for you and I to I like what does it even mean to build a harness? I d I don't know.

Rick (32:24)
Yeah. And how do we I like what I'm interested in is like, one, if if if you can help, I would like to learn how to collaborate on code. Like that, that's like

Tyler (32:32)
Uh-huh.

Rick (32:32)
my outcome. And the secondary goal is then I want to learn how to develop with AI. And then the second, the third goal is I want to help like up health grow. And like if all three of those things are true, like it's a no-brainer time for me to spend time on. so anyway, that's the idea. Think

about it.

Tyler (32:53)
Okay, so

you no no, I like it. Let's just assume the answer's yes. What I think what we should do is pick a specific workflow, ideally a relatively simple one, like let's not start too ambitious here. Like what's something that we could build a little tool, an internal tool for JD to use to streamline something he's doing.

We j we just need to pick something and go go build it. but it's har like I think we've hit the end of this conversation. As we've gone as far as we can go without having a specific idea. Do you know what that might be?

Rick (33:29)
Yeah, we need to we need to

we need to get something super narrow but but valuable.

Tyler (33:35)
Or n I mean, or not valuable. Like if the point is for us to learn, we gotta start somewhere. Yeah. Yeah, it's hard to know if we did a good job, yeah, if he's not using it. D

Rick (33:39)
Yeah, yeah. Something that JD will use. I guess is what I'm trying to say. Valuable enough for him to like use. Yeah. Yep.

Tyler (33:46)
anything come to mind for you?

Rick (33:47)
Yeah. It's probably some version of what we already started on with that comparison. it's a, you know, the I'm an agent. I have a business owner who wants to evaluate, you know, individual health insurance versus group health insurance versus, you know, all the different options, you know, and then you know, pick a pick a chassis and then

understand like how to move into quoting. and then actually fulfill the quote. But like I would just the first thing would be where where I would start.

Tyler (34:14)
So so what's the actual like you push a button, run the workflow, and you get some kind of deliverable at the end. Is the deliverable a PDF file summarizing their options for them?

Rick (34:26)
I'm trying to stay focused on like JD being the customer here versus the Yeah. Yeah, yeah. Yeah.

Tyler (34:30)
Right, then J D sends the the PDF to the customer.

Rick (34:33)
So so typically like when you're doing this, there would be like let's just say publicly available information and then private information. Private information would be like the names of all the employees, their family size, incomes, like social security numbers, like that kind of stuff. That's like a pure census. And so like phase one would be like

Soft intake. Phase two would be hard intake. And I think the first tool would be like send a customer a soft intake. Receive the soft intake, make a recommendation on a set of options to do a hard intake on. And then the the next step would be facilitate the hard intake. like send them another form

Tyler (35:22)
Okay.

Rick (35:22)
that gets them that gets the data in order to.

get the data on that. And then J D then goes and gets goes out like has the data. Now he goes into the these these other systems to get the data, brings it back, and then we package it. That would be the third the third thing.

Tyler (35:36)
Okay. So again, back to terminology. I think that's multiple workflows. Like, I think the definition of a harness here is the harness is what runs when a user a human gives one prompt or triggers triggers the workflow in some way. the the harness like it's what happens to get the response. So

It might be and you can maybe chain multiple workflows together, but like a single workflow, like like the reason I said what's the deliverables at a PDF. I was imagining you say the input is here is a company name, or here's whatever information we can get from Apollo, like company name, address, whatever. And the the harness is saying, okay, step one, let's make a guess as to this company's census information. Go online, look at their website, look at LinkedIn.

Like find a few and and our code is saying this. So our code is prompting the LLM saying, here are all the sources you should go look at. Go do whatever research you can on this, put together some estimates for number of employees, average age of employees, what like y you know, et cetera. And then it goes back to the code and says, okay, here's a JSON blob of the inf like my best guess as to this company's demographic information. And then the the deterministic code looks at that and says, Okay, well, we have a bunch of if statements that say probably this is we're gonna lead with paid.

leg up benefits or we're gonna lead with group insurance or whatever, goes back to the LLM and says, okay, look at all the available group insurance plans. Give me three that fit into this, you know, these criteria or something like that. And then it but but then the output is here's a PDF that gives you your options. And so for now, JD just goes in and types the name of the company, hits enter, gets a PDF and emails it to him. But like to me that's like

Rick (37:25)
Beautiful.

Tyler (37:25)
a pure workflow.

Rick (37:26)
Yep. So like it first workflow is you have a lead, they're you you have an initial call with them or an you have you want to send them with something initially, and all you have is their name and their website, and you want to put your best foot forward with lots of value. And that's it. And the goal is to get to to a hard quote request. That's the outcome. And we don't even worry about solving that problem yet. We just focus on that one. I think that's a brilliant idea.

Tyler (37:54)
Okay. Will JD actually use it? 'Cause like a while back I made this online quoting tool that kinda does something similar to this but without the AI in it, and I don't get the impression he needs it.

Rick (38:09)
That's a good question. Yeah, we should.

Tyler (38:09)
Like I feel like we should talk to him.

maybe I mean he s like I said last in the meeting yesterday, he gave us the list of five things that he spends time on. Should we just like pick number one and be like, We'll handle this one, you you handle the other four with tier two AI?

Rick (38:24)
Yeah, I think I think so. Like let's I think next step is I need to go I need to do an intake with him on like let's like actually rank these. Tell me like where you're like actually like quantify where you're spending time, what is most valuable, what like is very solvable tier two wise, what actually requires tier three. and then like let's I'll I'll have like some sort of doc that like specs this out in terms of this is we should it should be very clear like which one of the things is obvious, like

has a need for tier three and also has enough repetition that he will use it.

Tyler (38:58)
Yeah. Yeah.

Rick (39:00)
But I can imagine from a prospecting perspective, like if I could go into this tool and say, I want to prospect to the CEO of this company, can you build me a deck that I can email them and say, hey, like I would like to be your broker. Here's the deck I put together for you. Like I would use that.

Tyler (39:14)
Yeah. I would think so. What what he's gonna say i cause like J D is not that busy most of the year and then he's extremely busy during open enrollment, and that is less about prospecting and more about like renewing current clients, I think. So he's probably gonna say something in that ballpark, but yeah, okay, you you'll go talk to him, figure figure out what the example is. It doesn't even have to be that good. It th it's just a playground for us, but

Rick (39:39)
Yeah.

I like this one because it's simple and it's it there's no proprietary data involved. Like it's it it's like except ours. and to me like it's I like I like like if I were gonna go through the list, like one thing I would say like really high score on for this one is like there's no pro like confidential data being shared. And it's like, okay, that

Tyler (40:00)
Mm-hmm.

Rick (40:01)
makes this simpler right off the bat.

Tyler (40:03)
Yeah, it doesn't need to even con you said except ours. It doesn't even need our proprietary data. Like it's what

Rick (40:08)
Yeah, exactly.

Tyler (40:09)
are the what are the group insurance rates and what public information can we find about this lead? Yeah. Okay. Well, I'm gonna be in standby mode until you come back with some conviction on what exactly to build, but I'm game to you know, put a a low amount of effort into trying something out and seeing where it leads.

Rick (40:30)
Yep,

cool. And we and and just to be clear, like this will not be constrained by anything we built at Leg Up Health previously. It's gonna be completely independent. All right, cool.

Tyler (40:37)
Yes. This

is I I was just in Bend, Oregon for a friend's fortieth birthday, and th this is a San Francisco friend, so all the people there except me basically were kind of Bay Area tech workers. and some of them are, I mean, just vibe coding up incredible apps. Like one in particular person was showing me all his different apps, they're all mobile apps. I'm using them. I'm like, I'm like, this is this is better than anything I use.

Like any every app installed on my phone is worse than than any number that and he's purely vibe coding. He's not a coder, he's a project manager by background. It's incredible the quality he's making, but it's all brand new stuff, right? And he's shipping a new app every week and then like basically abandoning it more or less. it's incredible how good you can make a kind of quick prototype like that. And I still don't know how that scales as the like.

Like sometimes I look at less annoying CRM and we are using AI, it's helping us, but we're it's not like that. We're not like just shipping 40 different things every day and incredible quality. And I'm I'm often wondering like, are we doing it wrong? Or does this just fundamentally not work on this big of a code base with this many users and this many developers and all that? But anyway, yes, starting something new is definitely gonna be a lot better than trying to modify our current code base.

Rick (42:02)
Correct. Cool. What else is on on your list? Did you want to cover? I know I just stole the whole episode. I appreciate it.

Tyler (42:07)
no, that's that's I I was expecting that. I had that as one of my topics as well. I wanna give a a shout out to listener Tim. we were talking about marketing channels, I think last episode, two episodes ago,

Rick (42:19)
Mm-hmm.

Tyler (42:20)
for Less Annoying Serum and how nothing's working, but like one of things I said was, basically anything that's purely online is we're just gonna lose. Less annoying serum's gonna lose the competition. It's a zero sum game. HubSpot, Salesforce, Pipe Drive, they're gonna destroy us.

So we need to find things that are either niche enough that none of those guys are competing, or ideally something with like an offline, non-globally scalable element to it. and so Tim reaches out and is like, why don't you send physical mail to people? And I'll admit when I first read this email from him, I was like, I didn't have a good reaction to I was like, this is crazy. No, no way. And then I thought more about it and and he sent links like

You can use this service and here's the data you pull from, specifically to go after health insurance agents. And the more I thought about it, the more I was like, yeah, like this feels absolutely worth trying out. So I sent that over to Eunice. She was really excited about it. nothing to report yet, but she is she's going after it.

Rick (43:24)
Cool. I can't wait to hear about that.

Tyler (43:26)
Yeah. a thing that would have scared me off about something like this, as well as display advertising, these are two things that I've avoided in the past because creating the creative is so much work. Like you have to go into Photoshop or Illustrator or wherever and design a page and it is hard to make something look good. Or you can pay a graphic designer like five thousand dollars for what is supposed to be just a little experiment. I mean AI being able to or or even pre AI, just being able to use Canva,

To whip together like a brochure or a quick like display ad or whatever, really unlocks this type of thing in a way that I I think in the past it would have scared me off.

Rick (44:07)
Yeah, I agree. I was just thinking about this in the context of our previous discussion. It's like that's a you we just created a PDF. We're basically doing the same thing. We're creating

Tyler (44:16)
Yeah.

Rick (44:17)
a well-designed asset for JD to email to a prospect. but we could just as easily create a well-designed brochure to mail to a prospect. it's it's like a very interesting. Like I I like the realm in our app of like

If I could go back in time and spend my last twenty years working on businesses, I would not work on saving people time. I would not work on saving people money. I would work on making people money. and like

Tyler (44:45)
Mm-hmm.

Rick (44:46)
I I I feel like that's kind of what we're focused on. Like maybe that's the box we draw all something around. It's like, let's focus on helping JD leg up health make money. And if yes.

Tyler (44:56)
Yeah, J JD is the customer.

Leg up health is a less of a tech business than we've been saying. Leg Up Health is just an insurance advisor with a really, really good insurance advisor, but like nothing proprietary about it. And then we sell JD tech to make him better at his job. Yeah, yeah, yeah. I see it. That would be quite a pivot away from from what Leg Up Health is right now, though.

I'm not saying you have to commit

Rick (45:24)
Or an accept

Tyler (45:25)
to the pivot.

Rick (45:26)
Yeah, or an acceptance of reality. Like I mean th there is like a

Tyler (45:30)
Sorry,

if the idea is that this starts getting sold to other insurance agents, that would be quite a pivot.

Rick (45:33)
Yeah. It's not it's something

different. It's not Leg up health, right? Like Leg up health is customer one.

Tyler (45:40)
I actually like I'm gonna go down a weird rabbit hole. D you have a little more time? You have you have hard stop at 130?

Rick (45:44)
Yep. I do, I do, yeah.

Tyler (45:47)
Okay. Well, I've got five minutes then. in this trip to Bend, Oregon, talking to all these tech guys, several of whom work at OpenAI, like one of the things that like I I just don't see any way that OpenAI or Anthropic make money selling generic models, even if they're ten percent better than the other models. Like, I I don't see any way that that's gonna be a two trillion dollar business.

I'm not sure I have conviction behind this, but one of the things I said was like, if I'm open AI or anthropic, rather than saying I'm gonna sell a model to researchers who will then cure cancer, I'm gonna hire the researchers and cure cancer and sell the cure. Like, I'm saying that a bit tongue-in-cheek, but like owning something proprietary and using it yourself rather than like the the Silicon Valley approach for the last, you know, 15

Plus years has been build a product, a mass market business B2B SaaS tool, sell it to every b you know business possible and make between tens and hundreds of dollars per user per month, and hope that that turns into a multi-billion dollar business. But if instead it's build the product and then use the product and you're the only one with it, I think that's potentially what AI opens up for a lot of these businesses.

Rick (47:09)
applying that to leg up, like if we built the product for leg up and said, No, no, this is our advantage. We built the product. JD is the only customer. Yeah. Yeah. Yeah.

Tyler (47:14)
Yeah, try being a better health yeah, try being a better insurance broker than we are. You don't have the tech we do. Yeah.

Is it is it more profitable to sell a fifty dollar a month thing to every insurance agent in America, or is it more profitable to be the only insurance agent in America?

Rick (47:29)
Yeah. Or Utah. Yeah. Yeah. Yeah. Yeah. Yeah. No, but you can you

Tyler (47:32)
Or you talk yeah, yeah. I'm talking obviously too big, like this is grandiose. I I don't really mean it.

Rick (47:36)
can you can scale this up and down. Like it's it's frameworks. that's I like that.

Tyler (47:38)
Mm-hmm. Yeah.

All right. Well, it seems like that's a good stopping point for us.

Rick (47:47)
Cool. We'll see you next week.

Tyler (47:49)
See ya.