Subscribe
Share
Share
Embed
RiskCast documents what happens when you make AI agents structurally indispensable — not optional. Host Stefan Friend is building OpenClaw, a self-hosted multi-agent system that runs his businesses, and sharing the real experience: the good, the bad, the ugly.
Each episode explores the practical reality of living with AI infrastructure through conversations with builders doing the same.
Charles, I'm glad to be sitting down with you and chatting.You know, we've had a lot of ad hoc conversations around AI over the last 4, 6 months and, uh, uh, I sat in on y- your first One Shot Labs in December.Yeah.Yeah.Which is crazy that it was only- 4 months4 months ago.Right.It feels like it was both-yesterday and, like, years almost.Uh-huh.Um, theWe're, we're doing a new episode for the RiskCast podcast came out of the OpenClaw night.Yeah.That we also, uh- Yeahboth got to share on.Mm.And it just seemed like there's a growing community in Charlotte, but also everywhere else around- Yeahreally not just OpenClaw, but OpenClaw- All the clawsthe AI agents.Uh-huh.The, the things that are possible today- Yeahand in the months and, you know, further out ahead- Yeah, yeahwith these tools, so.It was a packed house.It was, it was- It was-awesome to see all the interest, so.It was so much fun.Yeah.So, excited to have you, and, uh, you easily had the most, uh, aggressive, interesting, like, cutting edge- Uh-huhOpenClaw system, so we'll want to- Thanksget into that.Yeah.But also the other things that you're building and what you're doing with your other agents.Sure.Yeah, yeah.Um- So.Thanks for having me.Yeah.Um, yeah, so I mean, uh, I've really, I don't know, just en- enjoyed chatting with you and, and, um, had, uhIt was really interesting, I guess, seeing, seeing your claw as well, your experiment with just basically throwing, throwing the slides up on that, that you had never seen- Yeah, yeahbefore and walking through.Um, it's, it's funny that I, I mean, I've done, done the same thing to a certain degree.Like, the amount of time I spend prepping isn't- Yeahlike, has gone down dramatically for any kind of meeting, even if I have to have, like, a polished slide deck.Be like, "Ah, I have a meeting in about an hour.I think I could, I should probably start on that deck now."Right?And so seeing you kind of like go through it live was really cool.But yeah.Yeah.Uh.Thank you.Yeah.It, it, it seemed like it was a good use case and I think it really worked well because of the audience where we had some people that were clearly aware but had not really played with it and- Yeah"Look, this is a live thing.Like, I, I know that it should be capable of this."Mm-hmm.You know?Uh, the, the agent obviously is aware of why I'm building it and how I interact with it.Let's see what it comes up with.Sure.And, uh, maybe skewed a little more marketing salesy narrative than, you know, practical implementation, but it, it was fun.Yeah.Yeah.And then your experiment on the other hand was, Yeah.I mean, it's definitely-run a company, right?Yeah.It'sWell, so it was definitely an experiment.Um, and the experiment, um, was still running live after the event up until Claude cut off OpenClaw.Okay.And I haven't yet kind of, uh, reconfigured it.So, so you were- But yeahusing the OAuth?Yeah.I was.Okay.Yeah, yeah.Um, so I had a couple Claude subscriptions that- YeahI was using kind of under the hood with OpenClaw.Um, but yeah, I mean, I had, you know, started seeing basically all these capabilities kind of in, in silos, marketing, sales, customer success, engineering, right?Essentially leadership, right?Uh, or I, I've been using orchestrator agents quite a bit in software development.Um- Yeahand I was like, "I mean, you could probably run a whole company," likeAnd, and I'd started seeing some folks talk about this as well.Like, um, Elon Musk gave a, he had an interview where he talked about the, how computer changed from basically a human role, you know, uh, a whole building full of people whose job it was to do hand calculations 'cause they were computers.That was the job.And then the computers basically became a physical object that, that, that did that job.Um, people still were using Excel, right?And- Yeahbut Excel, um, whatever software they were using would do the calculations for them, right?And one of the things he said was, like, if you could imagine just, like, any one of those cells having to still be done by hand by a computer, like, it breaks the whole thing, right?And the same thing- Yeahfrom a, uh, company standpoint.At a certain point you, you look at it and you go, "If any one of these functions still depends on a human, it, it dramatically slows down the, the entire operation."Yeah.Right?So I thought, "Okay, I need to experiment with this," right?And so I set it up.I basically had, uh, umI took one of a, a, aI justWe have a small, uh, SaaS product that, um, it d- does, it, it, it has real sales, real customers, but it's nothing crazy in terms of, like, the amount of traffic we're getting day to day.It's called dimeadozen.ai.And I was like, "This could actually be the perfect use case."It's kind of like a standalone product.I don't feel like I'm taking too much risk, um, you know, with- Yeahthe, a little bit of revenue it's, it's bringing in.Let me, let me see if I can take this and just run it, the, the entire thing end to end with, with agents.So I stood up a, basically a CEO agent, engineering, which is probably, like, the biggest part of the experiment I'll talk about more.But, like, um, set up a, a customer support agent that, that would hook in with our intercom system and respond to tickets.Um, and then a marketing and mark- Yeahbasically a marketing sales type, type agent that would be creating content for social mainly- Yeahum, but also doing emails and, um, uh, any other kind of marketing they might come up with.And I was, that maybe anything else they might come up with as kind of like a, a broader, um, experiment with the, the agent team which was-Within its, uh, instructions essentially for this team, I, I had to be very explicit with I want you to be thinking of new ideas, being innovative, really trying to push the envelope.But there is a broad goal of maximize profits for this company- Yepwithin eth-ethical boundaries.Um- I, I like that you- Yeahmade it a point- Yeahwhen you shared that and again- Uh-huhlike the within the ethical boundaries.Yeah, yeah.Building that in at the foundation.Yeah.Uh, I think you had your like, a culture document as well, so you- Yeaheven it, it really sounded like you essentially took this idea of, okay, I've s- I've seen orchestrator engines run code- Yeahand be able to build things.Yeah.But the same thing applies as long as skills are reasonably well-defined, and what's a lean startup team?It typically starts with, you know, the co-founders, maybe the first couple of employees- Mm-hmmbasically have a function and the, the flexibility and ability to move outside of, of that, you know, the crossover between- Right.Gotta-areas and like just- Wear a lot of hatsfiguring things out.So yeah, it was-- I mean, it felt similar to how you'd probably imagine building a company from scratch, thinking, okay, what are the high level goals?What are the boundaries?Um, what is our culture?What is our, you know, uh, what are our ethics?Some of it is almost like not even necessarily professional or like, um, like company building, but it's, it's setting ethical boundaries for, uh, an individual, like s-similar type of thing of if you were to work with a human, you're talking about the types of human.Maybe in hiring is a good analogy.Like if I'm gonna hire someone, right- Yeahum, I, I want them to be bringing a new perspective, but there's still kind of like common ethical guidelines that we want everyone to abide by.I saw-- I've seen where agents don't always necessarily have that baked in.Um, I've seen where agents will fabricate testimonials or, um- Yeahmake up, you know, whatever they need to in order to accomplish a goal.Like there's not really- Well, yeahthat same sense of like, oh, this might not be right.I think probably one of the worst examples, maybe not, not the worst, but that jumps out was I saw where a team had implemented some of these agents underneath, and they thought that they were generating sales.It was really just like a lead gen agent going to the, uh, the like account executive agent.They just, just had a loop happening in between them.God.I assume- Uh-huhbecause they didn't really have those kind of boundaries, and they're like- Yeah.Right.Right, right.Well, the- Yeah, you gotta be careful with like what goals you set.I mean- Yeahsimilar to like a human team, you might likeThere's been many iterations of incentive building for human teams, right?Where it's like if you incentivize lines of code, right, you get more lines of code, not necessarily more- Yeahfunctional software.You get more lines of code.Um- But it applies to- The same thing applies-to everything else- Maybe even to a more extreme-are your incentives aligned to the behaviors- Rightyou actually want- Yeahand, and the outcomes.Yeah.Yeah.Yeah, exactly.So went through that whole process and, um, I can't say that it was h- totally hands-off probably at any point, right?Like the, the goal was to basically get a completely self-sufficient company, right?Um, in reality, it wa- it, it did decrease over time, but initially there was a lot of training.Um, maybe another parallel to a human team, right?You just like need, needed a lot of direction early on and correction.Um- Yeahuh, and then over time, that kind of dropped off.Um, I definitely saw where, okay, I, I'd get too busy for, to really give feedback for a period of time, and the team would be less productive without my direction.Still not going to 0, but like- Yeahum, there's some technical challenges still, right, with, um, authentication and, um, even like CAPTCHAs and things like that where social platforms are pretty sensitive about having automation.Yes.Um, and then there's just like, um, the way that agents fail is, is like surprising sometimes, like superhuman in certain ways, but then really dumb in other ways.Like just, um, uh, it's just like surprising, like sometimes a s- a very, very simplest, simple solution for you or I isn't immediately obvious- Yeahum, to, to an agent.So, um, having to suggest the simple option sometimes is, uh, uh, something they need.It's still crazy the fact that I could just like occasionally nudge, you know, a group of agents and they drive a business.Yeah.Right?But yeah.Like I said, not, not a whole lot dissimilar in some ways from humans, but I think the, the loop there- Mm-hmmand the, the timescale to get to a more sustainable place, faster than onboarding people with- Yeahwith less, less context- Mm-hmmcoming into the role.Yeah.You know, they- Yeahthey might have those skills- Mm-hmmbut the ability to integrate into the team, refine how they, they understand the product- Yeahor your processes- Mm-hmmthat can often take a long time.Yeah, yeah.Um, so yeah, it's one of the things I'm, I'm trying to do for real now.You know- Yeahthis, uh, the, it was an initial experiment essentially with, with Dime a Dozen, which I, IReally the main goal of running an experience like that is to learn what I need to learn- Yeahto know where are the gaps and like how, you know, what are the systems I need to tweak when I actually do this for a client or maybe for something with higher risk for myself, right?And so I'm really excited about really starting to put that into production cases with clients to where I can- CoolI've started sitting down and talking with folks who are, you know, experts in the in-industry and being like, "Hey, like could you partner with me to help?Would you be interested in an A- in an AI employee really- Yeahto take maybe a part of your job, something that's done entirely on a computer today, um- Yeahand auto- and automate that?"Managed agents.Yeah.Right?Yeah, yeah.Exactly, yeah.Which I don't know if you saw, like Claude came out with managed agents like- Oh, I-just today, I believe.Yes.I, I, I saw that and immediately was like, I, IThis is great timing that it's coming up right before we talk 'cause I, I know that you've already-been experimenting with what these looked like before it was- Yeahpackaged as a feature from Claude.Yeah.Yeah, yeah.So I'm real interested in what they're doing.Um, there'sThey're, they're, um, a very safety conscious company, which is awesome, right?Yeah.We wanna be deploying agents safely.At the same time, sometimes I'm willing to maybe push the, uh, push it a little bit further- Yeahthan, than, than, uh, what the initial like, uh, solutions from a bigger, you know, lab would, is willing to do.I love that you're doing that.I think it's funny that I have not done that with my agents quite yet.Mm-hmm.Like even, even the one, you know, social media marketing flow, for instance, where like I got it all connected.Yeah.Technically was creating content.And this was into, uh, 11, 11 labs, uh, to actually generate, you know, the, the audio and video- Mmfor the, the brand that I wanted.Mm.But it was, it wasn't perfect.Mm.And I just, I stopped it from like turning the full switch on and, and r-really going- Mmbecause I, I felt like there was enough variation from piece to piece that's created that I just, "Ah, I don't like it."And I actually think in, in many ways it should just be like, "Let's just, let's just do it.Let's just-" Yeah.Yeah, yeah." get it out there, see how people respond to it."Right.Right.Things that- Yeah, yeahI would advise- Yeahyou know, companies that I've worked on or- Mm-hmmif I get on a project- Mm-hmmlike experiment in real life instead of, "Well, I experimented, and then I stopped myself from-" Right, right, right"taking the next step."Yeah.I mean, I know a lot of folks I've, I've worked with, and it, it applies for building like an MVP for so- for an app.It's-- Folks get, um, at a very early stage, like just, um, very, um, I don't know, hesitant to put something out there when they're like, "Uh, uh, you know, I, I don't want my baby to be ugly," right?Like, "I don't wanna put- Yeahthis ugly thing out in the wild, and then people judge me ba- 'cause it's like not, you know, up to my standard- Yeahor whatever it may be."And I totally get it, but at the same time, the earlier you can start learning, the, the better.And so if you're willing to put the, you know, ugly solution, the, the incomplete solution out there, it just is gonna start speeding up the whole cycle that much faster.Um, yeah, I'd say that's one of the biggest like, I don't know, challenges I've seen with folks is, you know- Yeahjust, just putting it out there.Um, finding ways to limit risk, like finding, you know, a, a small product or, or, you know, a small part, component of your idea and being able to put that out there.Um- Well, and I think- Yeahfrom Guardrail's perspective as well, you, you had said it initially.Yeah.You-- The one thing they didn't do was actually touch and control the money movement.Right.Right.Right.Yeah.That was one thing I, I held back on.Yeah.Um, was hoping maybe to at a, at a, at a certain point to get, um, to allow more money movement.And I, and I, and there actually are solutions that are coming out from like Stripe and, and others that are specifically geared toward providing more payment, um, capabilities to agents.Yeah.Right?Which is awesome to, to see, and wild that we can like officially support through a large provider, um-like a credit card issued to an agent with like- Yeaha low, a low credit limit or, you know, whatever it may be, and pr- and set up the guardrails that, that you want.Um, I mean, this whole thing's really wild, and I think that most, most folks, if you just, like, went out and talked to the average person, uh, have, like, have no idea just how much- Yeahthings have changed.Yeah.Uh, um- Well, and, and again, changed- Yeahchanged already over the last 12 to 18 months.Sure.Yeah, yeah.And continuing to accelerate.But also continuing to, to move forward.Right.You know, the, the conversations at Bank of America, which is one of the largest financial institutions in the world, and- Mm-hmmand so it's, it's moving a tanker.Mm-hmm.Right?Mm-hmm.A cargo ship- Mm-hmmto, in order to turn, and I have teammates that have been there decades that are, you know, making comments like this is the fastest they've seen things move here.Oh, wow.Okay.Yeah.'Cause you are having teams that you're like, maybe it's not there yet.Mm-hmm.Maybe it's, it's not quite like Stripe where- Yeahthere's stuff in production or- Mm.Mm-hmmin testable, but the conversations are happening.They see, like, how-how do we plan now and then get it out as fast as possible- Sure.to support the, the agentic systems that are- Yeahif they're, if they're not already capable of connecting into the banks and being able to, you know, monitor your credit cards or your bank accounts or money movements.Mm-hmm.Mm-hmm.It's heading that way.It's around the corner, and people are going to want it 'cause i-in some ways, those are really defined things.Like, you know you want to be able to pay off invoices- Mm.Mm-hmmlike, at the cutoff date, right?Sure.If you're net 60 terms- Yeah, sureI wanna, I, I wanna be able to find these rules and hand it to an agent, and- Mm.Mm-hmmpay it off.I want to pay my credit card bill.Like, all those things that a lot of times- Surepeople end up having to pay attention to.Yeah.You know?Yeah, yeah.Sure.Yeah, yeah.So I think, um, I mean, you're definitely quickly moving to a world where everyone has their own personal agent, I think.Um, and we don'tAnd maybe the, the amount of, uh, am- mental load that we all have to kind of carry, carry around, um, decreases, you know, as weOr as we all have our own personal Jarvis, essentially.Yeah.Which is really exciting.Um, but there's, there's still, I mean, a h- a huge, um, knowledge gap and just awareness, um, yeah, that, that, that needs to be closed, I, I, I feel like.Yeah.That's why I started doing One Shot Labs, actually.It was, was like I, I just felt like there's such a huge knowledge gap here, such great opportunity for folks like in dev roles, for one, but I mean, product, sales, everything, to just, um, learn these skills that are applicable.Right?Fo- I focus specifically on software, but, like, what I've realized is the skills that I've learned from working with agents to build software are directly applicable to basically every other, um, you know, white collar role.Um, and, uh, so I mean, uh, yeah, I'm trying, I'm, I'm trying in that, in through that to just, like, spread awareness and capability.Within a couple days of really, like, focused effort, I feel like most people can get pretty, um, pretty sufficient or, like, get to the point where they can start to teach themselves, uh- Yeah.and, and work- working with AI, right?Yeah.Um- And, and, and going- Yeahbeyond just the back and forth with a chatbot, but, like- Rightactually understand.Right.Yeah.H-how am I building something?How am I- Yeah.whether, whether, again, I think it's- Uh-huhtech or- Yeahyou're a person that's been in marketing for years.it's not just, "Oh, give me a marketing thing."It's no, how do you set up the skills that way it, you know, has some of your knowledge base and- Yeah.Yeahyou can go pretty quickly.Right.I think the biggest hump- A learning curveis just, that you have to get over, is just that initial, initial 1 Yeahof, like, fear of the unknown.Or like, "Oh, this feels very technical."And I think in the past, a lot of solutions have been, like, a lot of work, right?And- YepI'm not saying this isn't work, but it's different, and it'sI feel like it's just so much more fulfilling to work on and learn the, these skills, um, in terms of like working with agents.Yeah.Um, the, the fruits of your labor are just like so much more rapidly, um, there.Like, uh- Yeahit's, it's really exciting 'cause like I, um, I've always been focused on outcomes and, and product, right?And, and I love building, right?But the, the amount of, uh, just like grind you have to go through in the past to make small amounts of progress has been a lot, has been really, has been really high.And I feel like now, if you're willing to put in some amount of work into learning and, and using these tools, the amount of progress you can make is just like enormous, right?Yeah.Like unimaginable, right?Yeah.So- What would've taken- Uhteams months before- Yeah.Yeahyou can do with multiple agents- Rightin days or weeks.So I find, I find that very motivating, right?I'm just like anybody else, like, I mean, I, I'm hesitant to pick up a new, a new tool that looks like, oh man, it's gonna be a big learning curve.Like, you know, I was very, um, I don't know, hesitant with like AI workflow builders in the past, like n8n.Yeah.Right?Um, I can't say I everI, I, I never wanted to spend the time to be, to like really master- When the-n8n.that type of automation, it always felt likeOr, or going into CRM tools and then connecting those into- Yeahworkflows in other companies before, like is it automated?Sure.But I always felt like there were big gaps around the use, like the- Mmthe happy path.Mm-hmm.Felt big.Mm.Mm-hmm.Whereas it feels like this is the one where relatively easy to learn.I can kind of give it the direction, and because there's just that next step up in what it can understand- Mm-hmmmaybe outside of context- Mm.Mm-hmmit feels like it's able to close the, some of those gaps.Mm-hmm.And I, and it sounds like similar to your experiment- Yeaheven if it doesn't right away, well, now it's just a quick back and forth with me.Yeah.And, and we fill in context, and now it's got new memories- Rightand skills that- Rightit doesn't, I don't end up in the same loop again.Yeah.Right.Theoretically, yeah.Theoretically.Things are still not perfect.But yeah.Yeah, exactly.Um, yeah, so I mean, I, I f- I feel like, um, the more I can just work with my agent that's similar to how I would work with a new coworker, the, the better it feels, right?Yeah.I, I don't want to learn a new software system.I just wanna interact in natural language with, um, you know, something that, that is going to learn and, and get better at the job I'm giving it, right?Yeah.I've never felt more, um, I guess likeI, I, I've worked, I worked as an engineering manager at Vanguard, um, was there for over 12 years, and worked with some really talented people, right?Um, technical leads, architects, et cetera, right?And I feel like today, building new software systems has, it, just in the last couple months, I would say, like I feel like I'm interacting with Claude very similar to how I used to interact with tech leads and, and, and, and architects at, um, when I, in my role at Vanguard.Um, extremely capable.The, the- Yeahbut then the, the, the types, the, where I'm adding value is, is connecting the dots across, you know, uh, kind of with a large, a lot, a broader view of things.And knowing where I need to dig in, where I need to ask more questions.Um, and uh, and yeah, it, it's, but the, the, the, the wild thing is that like I can interact with Claude and the quality of the types of, um, ideas that Claude presents me with is- Yeahis, you know, um, incredible, right?Yeah.That's really cool stuff.So the, the experiment or, that you were running, said has not been going, uh, since the Anthropic change to be able to OAuth in.Yeah.Now you, you, you've got those lessons learned though.what's the next iteration of the experiment?Well, fortunately, uh, Anthropic made a lot of updates to Claude Code prior to cutting off the- Yeahthe Claw- the OpenClaw OAuth connection.And, um, the next iteration is just using Claude Code directly.Cool.Um, so I mean, there was a couple unique things about OpenClaw that were really valuable, right?One is the proactive nature of having the heartbeat.Yeah.Where you could configure that to whatever you wanted, you know, 15 minutes or, you know, every hour or whatever it may be.But you don't have to re-prompt.It just, you know, wakes up on a cycle and- Yeahlooks to see if there's a job to be done.Um, and uh, that was the biggest thing.Um, now with Claude Code, you can set up a loop and, uh, I mean, it was possible before, but it's still, it's becoming a, a c- a more core capability now.Yeah.Uh, there's still some experiment- experimentation to be done i- in that space because, um, and I'll, you know, I'm literally going through the cutover right now, so I'm going to see, like, if everything kind of works within Claude Code- Okaythe, the same way that I, you know, hope and expect that it, that it will.Um, but yeah, so that was, uh, the loop and, uh, being able to set up the, uhIt had, it had an, uh, built-in memory system.Claude Code has now its own memory system very similar to, um, to OpenClaw.There's actually a lot of different ways to create memory essentially within Claude Code.Um, uh, like initially there was the claude.md file where, like, basically that was just, that was basically the persistent memory space.And then, um, they introduced- Yeahskills, right?It was like- Yeahokay, interesting.Okay, what should go in skills versus claude.md, right?Um- Similar thing now within projects- YeahNow similar with projects and, uh, similar with, um, uh, they have a new, like, memory, essentially memory.md and with a memory system within Claude Code.Um, folks ask me, like, "Okay," like, "wait a second, help me make sense of, like, how am I supposed to use all this?"And it's actually not totally clear if you just, just by going through the documentation or even- Yeahplaying with it.I- There, there's not 1 Yeahcomplete setup yet.Yeah, yeah.It's, uh, a little bit tailorable still.Right.Right.So I tell peop- I tell folks, like, "Actually, don't worry about that too much."Like, everything really at the end of the day is context engineering.You have one context window.Right?And however that context ends up in the agent's context window, um, great.Great.Yeah.But don't, don't stress too much about whether this- Yeah, don't overthink that part.Like, oh, should this be a skill?Should this be in claude.md?Should this be in memory?Like, as long as it ends up in context, that's good.Um, so, so yeah.That's why I feel like- Yeahi- it's, it's been helpful if I'm, you know, clear in that context as well, like, here, here's what I'm doing in, in part of this for the, the agent.It's like, here's how I'm currently working and, like, what, where I save things and my folder structures, my projects and, you know, what maybe I, I'm used to having persist across- Mmprojects or tools.Yeah.It's been pretty good at- Yeahhelping then build, so that way that, that doesn't have to change a whole lot.It's, you know, supplementing well and, uh, can, can complement my workflows, but maybe not add new things- Mmwhich is what I'm trying to get.Like, how can I remove things that I have to be remembering to go check on or forms, fields, something to go work on when I know it's possible for me to, yeah, connect the dots- Mm.Mm-hmmand hand it to Claude Code- Mm-hmmor OpenClaw- Yeahand get them to just do it.Yeah.The other big thing with OpenClaw, um, is, was the ability to communicate through other channels, right?Not directly- Yeahyou know, through having to open up like CLI or the desktop app, whatever, but I've used-- I, I had demonstrated, you know, my, uh, agent team through Slack.Yep.Um, folks use Telegram, iMessage, or whatever.Yeah.And, um, now with Claude Code, you can do the same thing.You can set up either your own custom channels.I actually have not done that yet.Yeah.Um, yeah.I need toThat's still one of theI, I need to get switched over.Yeah, yeah.They officially support, I think as of today, I think it's whenever, when this comes out, it'll be a dozen more channels.But it's like- YeahiMessage, Tele- The big onesTelegram- Rightand Discord.Yeah.Right.But you can also set up your own custom, uh, channels.So I mean, uh, I'm gonna experiment more with that.It's wild, though.Um- I know, I think folks want to interact with agents, um, the same way they interact with other team members, like remote team members.Right?Um, the less that someone has to, to, to think about, "Oh, this is a agent," and like it's different than, you know, any other employee on my team, the harder I think the, the harder it's gonna be to kind of get over that hump of working with agents side by side, right?Yeah.When it's integrated with your team, you're literally communicating it with it in the s- the same Slack channel as your other teammates, sending emails, giving access to things in Google Drive, whatever it may be- Yeahum, the easier the adoption is for folks.'Cause I can just be like, uh, well tell me about the job that needs to be done."Right?"What's the, the job rec or, or, uh, the job profile?"Yeah.Right?Okay.Now tell me about how you would want-- how you prefer to work with, you know, uh, this, the, the, this new member of your team, right?Okay, great.We can get that, we can get that going.Yeah.Right?So I think that's exciting.Yeah.Me- meet people where they are and reduce friction- Yeahin ways that it's- Yeahactually making jobs and lives easier.Yeah.And hopefully amplifying folk- uh, giving folks the opportunity to amplify their own skills.Um, I, I, I mean I, I do think there's certain jobs that, you know, are gonna be impacted, but then it also gives folks with, um, uh, skill sets an opportunity to amplify their own impact.Yeah.And so, yeah.I hope we see more of that, uh, over time.think ho- hopefully it's twof- 2fold.On one hand, like amplify so the people that, uh, want to be building stuff- Yeahuh, are able to do so.But it does seem increasingly likely that, you know, as we go further out on the time horizon, there'll be major impacts in the workforce too.And- Mm-hmmyou know, leaving the socioeconomic parts aside for a second that- Mm-hmmneed to be addressed- Mm-hmmand figured out.Mm-hmm.Then as well that like h- h- hopefully that means people do have those things taken care of, so you can do the interesting things that you don't do while you're at work- Yeahthat are fulfilling for you.Spending time with friends, family, working on a hobby, you know, doing- Sure.Yeah.I like the idea of that.I mean, that, the fact that I can get projects done in a weekend now just 'cause I'm interested that would-- I just never would've had the capacity to do on my own before- Yeahis, is really fun and exciting, right?Um- Yeah.The, the biggest thing that I, I like to do, and I still do a good bit myself, but I have not made as much time to do it 'cause I just work in the different projects, always took time, and then, and then time would go toward family- Mm-hmmis like actually writing.Oh, yeah?I really do like to write.Like physically writing with a pen- Physicallyand paper?Physically.W- Typing.Typing, okay.Um- Yeahmore often.Uh-huh.My, my physical handwriting is like barely this side of hieroglyphics or chicken scratch.Mm-hmm.it-- that can be fun, but it doesn't make for very good translation later or- Yeah.Uh-huhuh, anything that persists.I just actually like to write.Yeah.yeah, so- Do you write with AI or do you write just kinda like freeform fromYeah.Do both- Okaydepending on the context.Yeah.But, so I, I actually have been working on, off and on, a couple of kids' stories- Mm.Mm-hmmum, since my kids were born.Yeah.Um, I guess I'll call it a memoir, uh, for lack of better words.Um, uh, almost 11 years now.Wow.Okay.Um, I, I got addicted to painkillers after a rugby injury in college.Oh.Yeah.Um, like too many people with opioids- Sureyeah, you-- real natural reason that kind of- Yeahdrove it, and then-uh, your body, your biophysically starts to need more and more, and so- Yeah.you take more and more.Mm.Start making excuses to doctors- Mmfor why you need more and more.Mm-hmm.Mm-hmm.And you go from, you know, the, the one Vicodin you're supposed to be taking- Mmto 2 to 4, to Percocet, to pretty much everything just short of heroin I had tried.Wow.Um- Yeah.So I bet you have a very, uh, interesting story to tell about- It was-the whole journey.So yes, I think so.And for me, there's also, there's a lot of parallels of my father had had a serious injury- Mmlike the year before.Mm.And I think he had also gone down a, a similar path in parallel- Okaylike in, in time parallel- Wowin lots of ways.Mm.But further.Mm.And it was seeing him- Mmin the ER that was finally like- Mm"No, I cannot do"Like, my- Yeahmy path ahead is this or worse- Yeahor, or I fix it.Um- Yeah, yeah.But it's combination of like not making the time, heavy emotions, dark, dark head space.Sure.Yeah.Um- Is the memoir- Ebb and flowthe part, you know, basically that you're working through?So that's, that's something that I've gotten to spend a lot more time on in the last- Yeah3, 4 months.Yeah.Um, it's been nice 'cause it-- so much of it is there.You're in that editing phase- Mma- as much as anything else of like, okay, p- pieces need to be rearranged.This actually is irrelevant 'cause-it's taking away from the story.It's an interesting anecdote.Oh, yeah, yeah.Um- Mm-hmmwhich I, I just wouldn't be able to do if I didn't have some of the AI tools that areHelping me build, helping me do other things.Yeah.You know, keep making it easier for me to signal what's important that needs to be paid attention to, even if I haven't quite taken the leap to just, "Ah, let's let you take care of it and I'll just clean it up later if you mess up."Yeah.Um, which I'm sure will come.Sure.Oh, that, that's great.Yeah.Yeah.Um, I mean, I just have little, like, h- hobby projects and, and, and whatnot that I'll spin up on the, you know, on the weekend.Actually, more often not than not, though, running my own business, it's, it's trying to get ahead on some project work.So I still have to manage- I get thatmy time on that stan- Yeahfrom that standpoint.Um, I mean, I've heard others talk about this as well.It's like you'd think that now with the power of AI and agents and whatnot that you'd be doing less work, but then it just, the cost of not having an agent running during a, you know, for a period of time also- Yeahseems like, uh, I don't know, just, like, a lot to, to miss out on.It's like- Yeahthat there's this, like, angst of, like, "Oh my God," like, "I can't believe my, my agent is just sitting there not doing anything right now," right?And so, um, it's like when you can produce so much in such a short period of time, then the, the anxiety comes from, like, taking, having any time at which things are paused.Right?there's a little bit of that.I, I think my wife has probably seen, uh, me do that a few too many times, where, like, getting toward bedtime- Yeahand, you know, the, there's a question that needed to be answered or maybe I was running a fix.I've done that a couple of times.Mm-hmm.And it got stuck.Mm-hmm.And I wasn't paying attention, and then I notice it.I'm like,I'm, I'm just, "I'll, I'll be right back."And then she goes to bed, and I'm s- I've-I've gone down a rabbit hole, um.Yeah.'Cause it's, 'cause it's fun.I mean, I- Sure.Yeahyou, you can do so much andYeah, I would say, um, most nights at this point, um, before going to bed, I'll kick off something that I know is going to be, like, a long-running thing.Nice.Yeah.That's a great way to do it.Um, uh, I mean, even just, like, end-to-end browser testing, right?Um, takes time.Yep.Right?So why not go through and, I mean, unless I have something else I want an agent to work on.Wake, wake up to a report in the morning maybe.Exactly.Yeah.Yeah, just kick it off.It can bring up the browser, go through the whole app end to end.It doesn't even have to be fully scripted.Like in the past, you'd have to, like, spend time- Yeahmaking sure that your, you know, Playwright or Puppeteer scripts are just, like, um, you know, perfect.It's a very flaky way to test end to end.Now, when an agent can literally just look at what's on the screen, interpret based on context how different things are supposed to work, it's not even alway- it doesn't have to be 100% confident.But, like, part of the report in the morning could be like, "Hey," like, "I was looking at this one page," right?"And I was expecting that markers would show up on the map in this particular way, and I didn't see that."You know, "Hey, is that what you intended for with this configuration or not?"Right?Like- Yeah.So there's a little bit of interpretive, interpretability.It's not always clear cut.But, like, um, you know, that's not something that software systems were capable of, you know, a year ago.Not a whole lot different than- So.The process is just simpler in some ways.Yeah.That, that would've been a meeting- Yeahto go through with the team and- Sure.you know, download where- Yeahwhere they got stuck and what didn't work and- Right.Right.Nice.Yes.Yeah, I think that's one of the, um, one of the shifts, one of the mindset shifts that I think a lot of folks are going to need to go through, and I've had to go through myself, is, like, this ideaI, I mean, I hear folks talk about, like, the, the basically going from chatbot to basically a worker that can take action- Yeaha true agent, right?But then also just, like, from a time standpoint, right?Like, you think, umI think our, our relationship with software in the past has generally been, like, it's, uh, fast and, and it's, it's immediate and it's deterministic, right?And, um, I've had to get more, um, I guess comfortable with or, um, uh, wrap my head around the idea of working with agents that maybe, you know, can solve problems, um, over a longer time horizon.Yeah.Right?And recognizing that, right?And, and putting it, and putting, you know, basically putting them to work over long- longer, longer periods of time.Um, I think it is a shift.Um, uh, and I don't know.I guess it's, it's, it's more like treating the AI like a coworker versus, uh, you know, like a software system.I mean, I've even heard from m- from other folks around like, "When I started talking with my agent more like a coworker- RightI started getting better results than when I was talking to it like a machine."Yeah.Right?And I think that even from a task perspective, when you can start thinking about how do I offload tasks, um, or jobs, um, more like I would offload to like a teammate versus a traditional software system- Yeahyou know, you can, you can accomplish a lot more.That, that feels as well like it ties back into the, the amplification hopefully for lots of people- Mm-hmm.of either their jobs or, or outcomes, both, that instead of lots of people are in relatively defined roles.Yeah.But if you start to be able to amplify, like they, they know the whole system.Mm-hmm.They know the whole product line maybe.Yeah.They're just, because of whatever other constraints within the business or the systems and tools-you have to have a person in that seat.Mm-hmm.You start to be able to think across and- Mm-hmmyou act with the agent or, uh, go, "Hey, here's the, the whole workflow that has to be done."Yeah.And you start to build a way for the tools to not just do this one piece, but- Yeahlike this one piece which is gonna carry into step two and just on and on and then, then the loops and- Sureone person- Yeahnow is doing your agent's, your, your- Yeahone marketing person, Right4 different marketing people with a manager- Yeahthat's- Uh-huhthe only one that can see the whole picture.Right.Right.Yeah.Yeah.And I think the loops are important that you're talking about.Um, similar- You have to think about how, how we do work, right, uh, as humans, right?You don't just do one pass on something and then b- feel 100% confident that you got everything right.In the past with a software system, you expect that, right?It was like, "I'm gonna click this button.It's going to run through this processing of this, you know, maybe a set of files," right?"And the results I'm gonna get are gonna be 100% right because, like, it's a deterministic, deterministic system."It does, it does it correct every single time.It does it the exact same way.Yeah.And now when you, when you're directing agents s- more similarly to like a human coworker, you can't expect it necessarily to always get everything or to get everything right the very first time.And so setting up loops to where you put the agent to work on, or multiple agents to work on, um, performing the task, but then evaluating the quality of that task.Yeah.Iterate, iterating to improve the quality, evaluating the new output.It's a continuous, um, improvement cycle, right?Yeah.And having some of those checkpoints.Yeah.Yeah.That-- So on that a little bit, I'm, I'm curious.I have a another friend I think similar to you in lots of ways, like started with the agents really in software building.Mm-hmm.Mm-hmm.And, uh, I, I know he had made a comment like when he had first deployed team and like loved it, like you could see him there start building and the, the way that they're able to de- deploy across, uh, the, the full stack- Mm-hmm.Mm-hmmand work was awesome.Yeah.He had, he had set up, "Hey, let's run like a team, like have, uh, daily stand up and- Uh-huhhave these check-ins."Uh-huh.Um, how, how i- have you done that if you're deploying agents?Are you setting up kind of those same types of rhythms that- Yeahyou might see in agile development or?Um, so there is kind of a, a re- reporting process, right, that, that, um, uh, is helpful, right?So oftentimes, or typically I would say I'm working with, I'm-- The, the top level agent that I'm interacting with to, and I basically am almost exclusively using Cloud Code at this point, right?But the actual agent that I'm sending a prompt directly to- Yeahis essentially acting as an orchestrator, right?And then all the real work is getting done by sub-agents.Yep.Right?Um, the-- What's important for the orchestrator is to have kind of a broad knowledge of the overall goal of what we're trying to accomplish, the broader plan.Um, think about it as like the engineering manager, right?Or maybe the senior architect, whoever it may be.Yeah.Right?Um, I actually have, oftentimes have to explicitly tell the main Claude agent, "I don't want you to do any work yourself.Um, I, you must only- Interestingorchestrate sub-agents."Yeah.If you don't do that, like it tends to, uh, want to jump in.Uh, I, I just launched 6 sub-agents to accomplish XYZ, but-I see this other thing that needs to be done.While they're working, I'm gonna go ahead and take care of this little thing over here."And I have to be like, "No, no, no, no.You're just the orchestrator."Um, and, uh- What, what happens- Yeahif it gets, uh- Well, it-sidetracked?Yeah, the, um, the, the, the benefit of the having the orchestrator i- is that I'm protecting its, um, its memory essentially, its context window with the bro- high-level broad plan, right?And then keeping track of how are we doing against that broad plan, and, you know, any kind of, like, high-level dependency type of things that it needs to keep in mind.When it starts going, getting into the weeds on any given problem, it starts to fill up its memory with the weeds, and I'm like- Yeah"You're, you're becoming a less effective orchestrator-because you're, like, taking care of all these little problems."Yeah.And then you end up having to compact your, your context window and forgetting about the overall plan.Yeah.So, so I have to say, "No, like, you are the orche- your job is to hold the broad plan in place-" Yeah "right?And to think about maybe trade-off decisions and things like that, right?But don't jump in."Yeah.Right?Um, it's, it's- Sounds like-kind of funny the, the, the human parallels, right?Yeah.I'm, immediately I'm like I've, I've, I've done this myself.Uh-huh.Yeah.I've had managers- Uh-huhthe same thing of like you're, you're stretched too thin.Uh-huh.There's work to be done.And you're like- Yeah"Well, let me go do that."It's like-No, you're, you're the, you're the manager.You're supposed to be- Rightunblocking your teams.Don't, uh- Yeahdon't go in too much.Right.So it's really important, I think, f- uh, to go back to your question, like, for the sub-agents to report back.A lot of it's just kind of built into the harness at this point with Cloud Code to where they'll, you know, pretty effectively be able to communicate back, like, what they've been able to accomplish or any issues and things like that.Um, uh, but yeah, it's not quite the same, I guess, because of speed and, and the number of iterations you can get through, like, where human teams wanna, like, take a pause every morning for 15 minutes and say, like, "Let's sync up."Um, a lot of that happens just, like, much more rapidly and, uh, in real time with agents.It's still cycles.Like, I'm-- We'll go through, like, we as in me and my agent teams, like, multiple waves of development.Right?Um, that's similar to, to Agile.It's just like, you know, it takes as long as it takes, which might be like, okay, it takes like, I don't know, 60 to 90 minutes maybe for like a full wave, and then we'll be ready to launch the next one.So.That's cool.Yeah.It, it makes sense that the, the rhythms exist- Mmlook different.Mm.They look different on every human software team too, right?Right.So.Right.Right.Yeah.And some things, like, it's interesting, like, um, Claude surprises me sometimes where like, uh, you know, I talk about just kicking things off before I go to bed, and, uh, sometimes I'm surprised 'cause I kick something off that I think is going to require a much longer development cycle, and, uh, and then it's done like 30 minutes later or something.Okay.Right?And, uh, and other times I get surprised in the other direction.So, um, even, like, Claude itself is a very poor, uh, estimate, estimator of how long- Mmwork is going to take.The- Uh- Do you have it do estimates for you of the work to be done?I mean, it, it'sI mean, it's actually so bad it's just like throw out because the problem is the, uh, the LLMs are trained on human estimates, right?So it, it, like, if you ask Claude to come up with a, a schedule for you, uh, along with like estimates of how long different things are gonna take, it'll be like, "Well, we're gonna need a week to like build out this foundation, you know, uh, to get through story one, 2, 3.You know, it's like 2 more days- Yep3 more days, whatever."And I'm like, "No, I want you to do it all right now and then-quit stopping."And then it comes back 30 minutes later, "It's all done."Like, okay.You were a pretty bad estimate.So, um, that actually would be really interesting.Like, I, I feel like there's more development that needs to be done in that space in terms of estimating not just time, but token usage.Mm.UmYeah.I've, I've thought about experimenting if I could get more confident with just like how many tokens something is gonna take, um, if I'm working with clients at some point.I feel like that could also become a pricing model where like you're literally just- Seems likelyputting some kind of, um, upcharge on the number of tokens.Like, "Hey, like what do you, what do you think it's gonna take?""Well, I think it's gonna take probably about 4 to five million tokens to get your project done."Right?"Okay.Well, how much is it gonna cost?""Well, a token costs X and then, you know, my markup is this," whatever it might be.Yeah.Right?So I could see that becoming a more standard, uh, model in the future.Yeah.Potentially makes it something that's still sustainable when the OpenAI and Anthropics of the world stop subsidizing the subscriptions.Right.Right.Right?Yeah.'Cause you'd go, "Here's the base cost," right?Um,Um, yeah, I'm, I'm hoping the models become- just continue to become more and more efficient over time.I'm sure they're- Yeahhoping that as well.Um, because otherwise if it plateaus, I think we're gonna be in trouble actually.Like if if models don't bec- continue to get more and more efficient, um, because if we just stopped right now and they stopped heavily sub- subsidizing the tokens, um, it would still be a lot more cost-effective, I think, to have humans do a lot of, a lot of things.Um, we need to get to a point where we can achieve the human amplification that everyone's dreaming of through, through agents, and we can only do that, I think, if things become-- if the tokens become more cost-effective.Especially considering there's, there's still a, a pretty big gap between what you and I can sign up for, even, even on the free plans- Mmlet alone the relatively inexpensive first paid tiers- Mm-hmmuh, and, and the open source ones.Mm.Mm-hmm.I, I've played with a couple of those.Um, my, my OpenClaw system still has the, the initial monitoring- Mm-hmmuh, done by the, the Qwen, uh, model.Mm.Um, just 'cause like that's, it's, that's the most well-defined of like for each channel what you're looking for- Mm-hmmand, and kind of with those loops, like anything it misses, we fix it quickly, it gets updated, so- Mmthe next time it's looking, what it's monitoring and how it surfaces things, it's pretty, pretty clear.Mm.But I don't really trust it a whole lot more than to just passively look for things and, and kind of make an initial decision.I'm not gonna have- Yeahit write responses or make a decision- Mm.Mmlike I would Sonnet or even Opus.Yeah.It just much better.It's all, and Qwen, the Qwen model's just running totally, uh, locally on your Mac Mini?Yeah.Yeah.That's great.Yeah.Um, I've looked into what it would cost to run, uh, one of the frontier open source models like, uh, like Kimi K 2.5. Yeah.And if you're really gonna take one of those like lar- the largest version, um, and try to run it, you, like you need pretty beefy hardware.Yes.And the cost of running that like in a, in a virtual environment like on AWS, it's, it's not cost-effective.No.Not, not on virtual- Yeahor I think- Or even a physic, like a- Oh, yeahlike a, like 3I've seen folks get like 3 Mac Minis, or not Mac Minis, uh, Mac Studios and- Yeah.I was about to say the same thingand like hook them together.Yeah.I, I looked at it- Yeahand it was like you're, you're gonna spend 5K on- Yeahat least on hardware and- Right.Right.well, now I just defeated the point.Right.Right.Right.Yeah.Um, though, I mean, we'll see what the API token costs 'cause if I look at my usage week to week, you know, across basically a, a Claude subscription, right?Um, if it was API costs, it would be like, I don't know, uh, tens of 1000.Right?Um, and, uh, being able to get that, thou- thousands of dollars worth of usage from an API at 200 bucks a month is pretty incredible.Yeah.Um, so yeah.Well, we're, uh, coming up on, uh, uh, I think end of the conversation time.Uh, any last thoughts or, or if not- YeahI'll, I'll have, I have one maybe final question I can ask.Okay.Well, I'm curious about the question.Okay.Uh, okay.Ask the question.Ask the question.Yeah, yeah.So yeah.I mean, I think you've touched on, on some of this a little bit, but if you look, if I more immediately, like next 6 months- Mmwhat are you hoping you're able to do with these agent systems- Mmthat you haven't been able to do yet?Um, so I, I am hoping to help make it a reality for more people.Um- Rightand I haven't, I haven't been able, I haven't been able to really do that.Like I've been building software systems, even- Yeahlike AI-driven software systems in the past.Um, but the process of training, um, agents on a brand-new s- space, a brand-new field, a brand-new skill set, right?And then being able toUm, integrate that agent into a client's existing workflow in a, in a way that's much more similar to onboarding a new team member.Yep.Um, that's I think what I'm really excited about over the next 6 months is being able to streamline that process, um, and make it, make it real for- Which- Make it real for more people.Which part- Yeahfeels like it's the bigger hurdle at the moment?The actual agent or the company people, you know, helping them get comfortable, get it on board, be able to use it properly?Well, there's like, uh, there's been initially a, um, a good amount of nervousness when you start talking to, to folks about AI and agents, and folks were like, "I don't even know what that means."Like- Yeahum, yeah like, "I'm interested 'cause I hear I need to be doing AI, but like, um, it-- the whole thing makes me feel nervous."Being able to say to someone like, how do you interact with your team today?"Like, okay, you-- this is, this is how you'll be able to interact with your, your agent in the future, right?Yeah.And, um, kind of being able to kind of walk through what it really looks like and, and will f- you know, feel like just day to day has helped with that quite a bit, right?Presenting that vision, right?Right.Um, and it hasn't been actually as, as dif- difficult to get over as some of the surprises when it comes to just like the quality of the actual agent that we have to build and deliver.A lot of jobs, even when they look on the surface, um, as like very straightforward, when you really get into the, the nitty-gritty of it, um, you start discovering that the humans actually are making a lot of like fairly like, um, nuanced decisions- Yeahin certain, in certain places or there's a lot of, um, like tribal knowledge, right?That isn't necessarily documented, but is in someone's head.And, um, so like kind of goes back to our saying like the ways in which agents fail sometimes is surprising, like it's superhuman in certain ways, but then like surprises you with how it, how it misses certain things.It's- Yeahsimilar li- like that when you're bringing on a human, new human employee, you can have them sit down and like learn a job, right?And then a lot of things just start intuitively, they just start catching on to some of the patterns.Um, so, uh, to answer your question directly, I actually think solving for that i- has, is a challenge.Um, I think that the tooling that's available now for do- for solving for that is, is better than it ever has been.Um, but having consistent quality from a, from an, from an agent- Yeahuh, is, is a, is a real challenge,Especially in a brand new field, you know, that is-- doesn't have like built-in expertise.Makes sense.cool to see.I think it's heading that direction.Absolutely.Yeah.So I'm excited to try to get it, you know, try to get it out there, try to get the, the power of agents into, uh, the hands of as many people as possible.Folks who don't have any like technical background, who, um, you know, maybe to a certain degree even have some like, uh, like nervousness around leveraging brand new technology they don't understand, right?Yeah.Just making that as easy for folks to adopt as possible, right?Yeah.Well, it-- speaking from experience since I sat in on your one shot cohort, I, I think you certainly help in making it feel more comfortable to just dive in and- Mm-hmmuh, I mean, in that, that first one, we had a few people that are significantly less technical than I am, I think.Mm-hmm.I'm, I'm not a engineer by trade like you are, and everybody left with functional stuff.So I don't, I don't doubt that you'll be able to keep working on that, man.Thanks.Yeah.Appreciate it.Cool.Well, this has been a lot of fun.Enjoyed it.Thank you again for coming in.Yeah.And- Thanks for having me