The argument in one line.
Running a startup in 2026 means becoming a full-time manager of AI coding agents rather than a part-time coder, and the people who master that management skill — working from the cloud, making dozens of high-stakes decisions a day, and routing tasks to the right-priced model — will out-ship everyone still working locally.
Read if. Skip if.
READ IF YOU ARE…
- A solo founder or one-person startup team trying to scale output without hiring, especially anyone already running Claude Code, Codex, Devin, or Cursor day to day.
- A technical or semi-technical builder who has tried an AI coding agent once or twice but hasn't built a repeatable system around running several at once.
- Anyone managing a real product with paying customers who needs a way to catch production bugs and regressions without manually reading logs every day.
- A founder or solopreneur worried about a ballooning AI token bill who wants a concrete monthly budget and a model-routing strategy.
SKIP IF…
- You want a tutorial on writing your first prompt or picking your first AI coding tool — this assumes you already have an agent workflow running.
- You're looking for enterprise engineering-org advice — Ryan's playbook is built for a team of one to a handful of people, not a scaled engineering department.
The full version, fast.
Ryan Carson, a solo founder running the AI divorce-tech startup Untangle, argues the highest-leverage skill for 2026 is managing teams of AI coding agents rather than writing code yourself. He works entirely from cloud-based virtual machines (mainly Devin) so five to ten agents run in parallel without code collisions, shipping 22 to 40 pull requests a day, and does roughly half of that management from his phone. The job has shifted from typing code to making 10 to 20 high-stakes decisions before lunch, which he survives by pinning his most important threads and checking in on a roughly 25-minute cadence. He layers in automations — an end-to-end signup test, a daily production-watchdog summary of customer activity, and a self-improvement loop that grades and fixes an AI agent's own customer conversations — so the system keeps improving without his constant attention. After a $20,000 token month, he settled on budgeting about $5,000 per 'engineering employee' a month by routing cheap, high-volume work to a fine-tuned model and reserving frontier models for genuinely hard calls, and warns against building an entire software factory on one frontier lab's own coding tool.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Who's talking.
00:00hostGreg Isenberg
01:22guestRyan Carson
Where the time goes.
00:00 – 01:43
01 · Intro
Greg's cold open frames the episode's thesis — people who run teams of AI agents will outperform everyone else — and brings on Ryan Carson to explain cloud agents, automations, and building a software factory.
01:43 – 04:24
02 · The New Agent Paradigm
Ryan lays out his background: 25 years as a founder, scaled Treehouse to 110 employees and a million learners, and now runs Untangle — an AI divorce agent for family law firms — as a team of one with revenue on track to 4x this month. States that everyone, regardless of prior role, is now a manager of agents.
04:24 – 07:23
03 · Inside Ryan's eight-screen desk setup
Screen-share tour of his 52-inch monitor split into eight zones: Slack for agent notifications, a UG Monk paper to-do system, a vertical mouse and WhisperFlow for voice input, 1Password to keep prod keys isolated from agents, Devin, two testing screens, X, and Codex. About half his work happens on his phone.
07:23 – 10:00
04 · Why Devin and cloud agents
Explains why cloud-hosted agents beat local development: cloud VMs eliminate git-worktree and code-collision overhead, letting him run 5-10 agents in parallel. Calls working purely locally 'caveman' behavior that ships 10x less.
10:00 – 13:15
05 · Everyone Is Now a Manager of Agents
Responds to 'local maxis' on X who see cloud agents as amateurish — says they're not doing real work. Describes needing to 50x his output once Untangle hit product-market fit and started shipping real fixes to paying customers.
13:15 – 14:50
06 · Local vs Virtual Development
Transition into how to structure a system that doesn't drive the operator crazy once output scales — reframing the job itself around decision-making rather than typing.
14:50 – 24:22
07 · Agent Management System
The core operating system: work is now making 10-20 high-stakes decisions before lunch; pin the threads that matter and let the rest run; check in on a roughly 25-minute cadence. Shows his real PR chart averaging 22-25 PRs a day (up to 40), including 8 PRs shipped from his phone while hiking Mount Washington with his son.
24:22 – 32:54
08 · 3 Automations to build
Covers the end-to-end signup test (browser QA three times a week, ~$60 in tokens per run), the production watchdog (9am daily digest of customer activity linked to real UI), and the pattern of spinning up a manager thread that delegates to cheaper child agent sessions.
32:54 – 34:38
09 · Self-improvement loop with Grace
Describes Grace, the AI paralegal agent inside Untangle, and the daily loop that grades her customer chats against a rubric, auto-spins a fix session for anything scoring low, and ships roughly three small UX fixes a day.
34:38 – 40:33
10 · Token costs and model routing
A $20k token month was the wake-up call. Settles on roughly $5,000 per 'engineering employee' a month, using Cognition's cheaper fine-tuned SWE-1.7 for high-volume loop work and reserving frontier models for hard calls. Warns against building a company's entire software factory on a single frontier lab's own tool (Claude Code, Codex) since those labs aren't incentivized to optimize cost.
40:33 – 44:03
11 · Building reputation on X
Argues that publicly sharing what you're learning on X compounds into real relationships and opportunities, citing his own relationship with Greg and Sahil Bloom's rise from private-equity newsletter writer to bestselling author.
44:03 – 44:47
12 · Closing thoughts
Recaps the three takeaways — cloud agents, automations, and building public credibility — and Greg signs off, pointing listeners to follow Ryan on X.
Lines worth screenshotting.
- A one-person startup can ship 22 to 40 pull requests a day by running five to ten cloud agent sessions in parallel instead of coding locally.
- Cloud VMs remove the git-worktree problem entirely — each agent session gets its own isolated environment, so there is no code collision to manage.
- Production write keys should live in a password manager, never handed to an agent by default — the agent has to ask for the key and the human pastes it in manually for every prod write.
- The real bottleneck at high agent volume isn't coding speed, it's the human's capacity to make roughly 10 to 20 high-stakes yes/no decisions before lunch.
- Pin only the two or three threads that matter today and let everything else run in the background — checking every thread constantly is what burns the manager out, not the agents.
- A roughly 25-minute check-in cadence keeps a solo operator sane while running many agents at once.
- About half of agent management can happen from a phone, because agents block on human decisions around the clock, not just during desk hours.
- An automated end-to-end signup test that browser-tests the real product three times a week costs about $60 in tokens and catches breakage before customers do.
- A daily automation that summarizes overnight customer activity into a report linked back to the real UI replaces manually reading logs.
- A self-improvement loop can grade an AI agent's own customer conversations against a rubric daily and auto-spin a child session to fix anything that scores low, shipping small UX fixes a human would never have bothered prioritizing.
- A $20,000 token month is what forces most builders to start routing tasks to cheaper fine-tuned models instead of running everything on a frontier model by default.
- Budget roughly $5,000 a month in tokens per 'engineering employee' once a product has real usage — above that, the fix is model routing, not a bigger budget.
- Route by task, not by habit: cheap fine-tuned models handle high-volume loop work, and expensive frontier models are reserved for genuinely hard calls.
- Building an entire company's software factory on a single frontier lab's own tool is risky long-term because that lab isn't incentivized to optimize your token bill — independent agent labs are.
- Publicly sharing what you're learning, even unpolished, compounds into relationships and opportunities over years — it's how guest bookings and business relationships actually get built.
The job changed from writing code to managing agents and routing their costs.
WHAT TO LEARN
Scaling output now means running agents from the cloud, budgeting your decision-making capacity like a scarce resource, and routing tasks to the cheapest model that can do the job.
01Intro
- A cold open can state a bold, testable claim in one sentence and immediately name the guest who will prove it — no throat-clearing needed.
02The New Agent Paradigm
- Prior management experience (people, teams, organizations) transfers directly to managing AI agents — the discipline is the same, the reports just aren't human.
- A solo operator can run a company with real paying customers and rapidly growing revenue without hiring, as long as agent management replaces headcount.
03Inside Ryan's eight-screen desk setup
- A 52-inch monitor split into purpose-built zones (notifications, key storage, testing, agent sessions) reduces the friction of monitoring many parallel workstreams.
- Keep production write credentials in a password manager and require an agent to explicitly ask for them per task, rather than granting standing access.
04Why Devin and cloud agents
- Cloud-hosted virtual machines let multiple AI coding sessions run in parallel without colliding, removing the need for git worktrees or duplicate local checkouts.
- The mental overhead of coordinating parallel local work (which files, which branch, did I pull) disappears once each agent gets its own disposable environment.
05Everyone Is Now a Manager of Agents
- Reaching product-market fit is often the forcing function that makes a founder abandon local development entirely — real customers require a step-change in output.
- Front-end and wireframing work may still start locally, but should move to a cloud agent as soon as possible once the shape of the feature is clear.
06Local vs Virtual Development
- Roughly half of high-output agent management can realistically happen from a phone, since agents block on human approval at all hours, not just desk hours.
07Agent Management System
- Reframe the job itself: at high agent throughput, the scarce resource is the human's capacity for high-stakes decisions, not typing speed.
- Pin only the threads carrying today's most important work, and check them on a fixed cadence (roughly every 25 minutes) rather than continuously, to avoid burnout.
- A visible PR-per-day chart is a useful gut-check for whether an agent workflow is actually shipping, not just generating activity.
083 Automations to build
- An automated end-to-end signup test that runs a few times a week in a real browser catches production breakage that unit tests miss, for a modest, bounded token cost.
- A daily automated summary of real customer activity, linked back to the actual UI, surfaces problems a human scanning logs would likely miss.
- Delegate automation-building itself to a parent agent thread that spins up cheaper child sessions, rather than hand-building every integration yourself.
09Self-improvement loop with Grace
- A self-graded improvement loop — grading an AI agent's own output against a rubric and auto-fixing what scores low — surfaces small UX problems a human would never prioritize fixing manually.
- There is still no substitute for a human periodically using their own product end to end; agents miss judgment calls that are obvious to a person.
10Token costs and model routing
- Treat a runaway token bill as a signal to add model routing, not a signal to cut usage: budget roughly $5,000 per engineering employee per month and route cheap, high-volume work to a fine-tuned model.
- Avoid anchoring an entire company's engineering workflow to a single frontier lab's own agent tool, since that lab has no incentive to optimize your long-term cost.
- Independent agent labs that sit on top of multiple models are structurally incentivized to find the cheapest model that gets an acceptable result for each task.
11Building reputation on X
- Publishing what you're learning publicly, even when unpolished, is a durable way to build the relationships and credibility that later become business opportunities.
- It's fine to publicly admit you don't know something and let an agent help you figure it out — that honesty is itself part of the credibility-building.
12Closing thoughts
- The three compounding skills for this era are working from the cloud, building automations, and building a public, credible reputation.
Terms worth knowing.
- Devin
- A cloud-based AI coding agent made by Cognition that runs full development sessions inside its own virtual machine, including writing, testing, and shipping code.
- Cloud agent / VM
- A disposable virtual machine spun up per coding session so multiple AI agents can work in parallel without their code changes colliding.
- Production watchdog
- A daily automation that scans a database for real customer activity and produces a summary report the founder reads each morning instead of digging through logs.
- Model routing
- Deliberately sending different tasks to different AI models based on cost and difficulty, instead of running every task on the same, usually more expensive, model.
- SWE-1.7
- A fine-tuned, lower-cost coding model built by Cognition, used for high-volume automated loop work instead of a frontier model like Opus or GPT.
- Playbook
- A saved, step-by-step set of instructions inside Devin telling an agent exactly how to correctly execute a recurring task, distinct from a one-off 'skill.'
- Independent agent lab
- A company that builds an agent product on top of multiple underlying AI models (e.g. Devin, Cursor, AMP, Factory) rather than a single frontier lab's own coding tool, letting it route work to whichever model is cheapest for the job.
- Git worktree
- A way to check out multiple branches of the same code repository locally at once; described here as a technical workaround that cloud-based agent VMs make unnecessary.
Things they pointed at.
02:29companyTreehouse
04:44productUG Monk paper system
05:02toolWhisperFlow
05:46tool1Password
06:04toolDevin (Cognition)
06:28toolCodex (OpenAI)
08:15toolAMP
08:15toolCursor
00:39toolFactory
20:35toolFable
20:57toolFusion
35:59toolSWE-1.7
42:58channelSahil Bloom
Lines you could clip.
00:00
“I don't think it's crazy to say that the people who know how to run teams of AI agents are gonna be able to outperform everyone else in this AI age we're in.”
02:01
“You need to be the best in the world. And it doesn't matter whether you're a VC, a founder, an IC, a mom, a dad, like a college student.”
12:48
“If you are working locally, I honestly think you are a caveman.”
13:34
“They are not doing real work.”
15:05
“What your work is now is making high stakes decisions almost all day.”
18:45
“It turns out this weird back muscle, it turns out that's the most important muscle in the world.”
34:03
“It is shocking to me the number of small fine UX details that have been picked up by this loop and have been fixed.”
35:05
“For real engineering work per employee, you're looking at probably $5,000 a month. And anything more than that, you really need to figure out model routing.”
36:55
“I don't understand why anybody would build their engineering on a Frontier lab stack.”
42:32
“Take the time to share what you're learning. It doesn't have to be super polished, but get out there and be useful to people, and it will open up doors.”
Where the conversation goes.
01:43 – 04:24steadyRyan's background and Untangle
04:24 – 07:23denseDesk setup and tool stack
07:23 – 13:15denseWhy cloud agents beat local dev
14:50 – 24:22denseDecision-making cadence and PR velocity
24:22 – 34:38denseThe three automations
34:38 – 40:33denseToken cost and model routing
40:33 – 44:47steadyBuilding reputation on X
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogystory
00:00I don't think it's crazy to say that the people who know how to run teams of AI agents are gonna be able to outperform everyone else in this AI age we're in. So how do you actually run teams of AI agents?
00:11What does this mean? What's the step by step way to actually do this, and what are the tools you need to know, like Dev and AI, to actually go and do this? Well, on today's episode, I brought on Ryan Carson, and he clearly explains all of this for you.
00:25He even gives you a tour of his insane setup for how he manages his AI agent.
00:33So by the end of this episode, you're gonna understand what cloud agents are, how you can set up automations, how you can create a that creates software twenty four seven, and you're gonna be able to understand it just really clearly because he does such a good job at explaining all of this. So enjoy the episode.
00:49I'll see you at the end, and have a creative day.
01:00Ryan Carson back on the pod. By the end of this episode, Ryan, what are people gonna get out of this? So you're gonna do be able to do three things really well.
01:10You're gonna be able to run Cloud Agents, you're gonna be able to do automations, and you're going to be able to ship faster. So whether you're a founder, a solopreneur, a stay a stay at home parent, you're just gonna become an agent pro
01:26by the time we're done. Okay. Because I think a lot of people listening to this might have they've played with agents, cloud agents, but you're gonna help them just get the confidence basically to be world class.
01:39You know? I think that's what I'm trying to that's what I've been thinking a lot about. It's like, okay.
01:43There's, a bunch of people who have played with agents, but how do we become world class agent operators?
01:50Yeah. Essentially, you are a manager of agents now. So no matter what you used to do, whether it was a people manager or an IC, you are going to become a manager of agents now.
02:01And you need to be the best in the world. And it doesn't matter whether you're a VC, a founder, an IC, a mom, a dad, like a college student.
02:10You should manage agents and be the best at it. And we're going to show you some tips for that. Okay, so where do we start?
02:17Okay, so let me give some context, right, so people understand where is this coming from. So I've been a founder CEO for like twenty five years.
02:26I'm working on my fourth company now. At my last company, TreeHouse, which was acquired, we had I think up to like 110 full time employees, taught a million people how to code.
02:38So I've done the whole build organization thing and manage organization thing. And you learn a lot about managing people on that. And now this startup, which I'm running, I am one employee right now.
02:51So we've raised a seed round. We're scaling really fast. Our revenue probably is going to 4x this month.
02:58And I'm probably going to be hiring my first employees. We are essentially an AI divorce agent for divorce firms.
03:05So we do the dirty topic that no one seems to want to tackle, which is family law. So I've learned a lot about how to scale me, right?
03:14And so I'm going to help folks understand how do you manage a lot of agents well? How do you check their work?
03:21How do you automate things? How do you work from your phone? You know, I do almost 50%, probably more of my work from my phone.
03:31And it's pretty wild. So that's where we're going to take you.
03:35And I will say as well, like, this is going to change every three to six months. So we're going to share a lot with you.
03:44Get your hands dirty. Try it. And the best way by to learn is to do.
03:50Cool. Yeah. I'm actually I'm particularly interested in the phone stuff because I actually do 0% of my stuff on the phone.
03:56What? Yeah. Okay.
03:58Let's talk. So
04:00let me start by just showing you, well, when I'm at my desk, what is the world like? Right?
04:04So right now I'm in my home office. I work from Connecticut and I do all my desk work here.
04:13And I've got this big 52 inches Dell monitor. I used to have like a couple monitors and I was like, screw it. I just want a huge monitor.
04:20So let me share my screen. All right.
04:24So what I do is I set up eight screens on my monitor. And I also enjoy food and snacks as I code.
04:37But the key is to have eight things up at once because I have to multitask a lot of agents at once. I also use a paper system to keep track of some of my to dos, which is a little bit weird.
04:48I use the UG Monk system, which is kind of fun, and I've got a killer Razer mouse. I was getting a lot of wrist pain actually, so I decided to go vertical mouse on that.
04:59I also have a button which actually is connected to WhisperFlow.
05:05So I almost use WhisperFlow for all of my stuff. So I've got Slack in the top left.
05:11That's where I get notifications and things about what my agents are doing or what is happening. That's the primary thing I use Slack for.
05:20And then what I do is I lay out various other screens. So I've got a one password always set up.
05:26It's very important to keep your keys safe, secure, and separated from the agent, Right?
05:33So I've got all my prod write keys in one password. I do not give my agents prod write keys. So production writing is dangerous.
05:43Right? And your agents will do something bad. And so you need to set up a system where they ask you for the key when they need to write to prod, and then you go into 1Password, you copy and paste it into the session, you're both very clear about what you're doing.
05:56So that's one tip. Keep your keys secure.
06:01And then I use Devon. So Devon is created by Cognition.
06:06I think it's one of the best software factories in the industry. It's not cheap, but it's good.
06:12The things I want talk about today are not Devon specific, but they do work very well in Devon. And one of the keys is Cloud Agents.
06:21Then I usually have two screens set up where I am testing my app. It's just a lot of testing.
06:28And then, of course, I've got X in the top left and top right. And then on the right, I've got Codecs, which we'll talk a little bit more how I do most of my work in cloud in Devon.
06:39And a lot of that work, like we talked about, 50% is probably on my phone. And then whenever I do desktop or whenever I'm doing things locally, I'll do it in codecs. Usually that's in the bottom right.
06:52So it's kind of my, when I'm at my desk, my setup and what I do when I'm here.
07:00But the truth is I do a lot of work, you know, from this, my iPhone. This is my beautiful wife.
07:07Just I get a lot of stuff done, you know. I'll be in the shower and I'm thinking, Okay, I need to check on that PR. Or maybe I want to land that PR, right?
07:15You know, grab it, talk to Devin, which is just done in a browser on my mobile, and I get a lot of stuff done that way. So that's my setup.
07:24So I I keep hearing this that Devin is is almost, you know, it's the most powerful coding agent.
07:32And, you know, can you just for folks who haven't heard of Devon or who haven't played with it, why is that? What is going on?
07:39So
07:41about two years ago, maybe two and a half years ago, Devon came out. And they pitched it as this idea of like, it's a software engineer. And honestly, didn't work very well.
07:51The models were just not good enough. And no matter how good the harness was, the models just weren't there. And then fast forward to 2026, the models are definitely good enough.
08:04And when you have a good harness that is cloud based, it's really good. So think of it as you've got your choices, right?
08:12So if you're using agents to build things, you're probably using codecs from OpenAI or using Cloud Code from Anthropic or you're using one of the Indy I call them the Indy agent harnesses. You've got AMP.
08:25You've got Cursor, which is not independent anymore, but sort of, owned by Elon. And then you've got Cognition, which makes Devon.
08:34You've got Factory as another one as well. So you've got these choices, right?
08:39And so I thought, Okay, I want to scale myself horizontally.
08:46Like I want to be able to run as many agents as I possibly can. And I don't want to be working on my laptop and trying to figure out which code is colliding, which work tree am I on, all a lot of technical stuff that really it keeps you from shipping.
09:04And I know there's a lot of different folks that listen to the show. You might have solopreneurs who are newer to coding.
09:12They don't have computer science screens. They're not super technical, but they're building things. Then you may have software engineers who are multiplying themselves 10 to 100x.
09:21They have deep understanding of software and architecture. And then you have founders who are coding but are abstracting themselves. And you have VCs who are coding.
09:32But everybody's going to learn a couple of skills.
09:36And what's interesting is that I'll sort of zoom out for a second. I think we all thought that engineering and being technical was going away.
09:45But actually what's happening is the more you become a better agent manager, actually the more technical you become. And I think a good analogy is that you are basically an engineering manager, right?
10:01And so to be a good manager of many agents, you do need to become technical, right? There's this reality of you have to understand what Postgres is. You have to understand the production environment versus a dev environment.
10:13You have to understand migrations. But the truth is you'll learn all that by using agents, right?
10:19And so I am way more technical than I have ever been in my whole life. And so I think the idea has become more and more technical. Therefore, Okay, so let's back up.
10:30Why Devon? What is it? It's a really good agent.
10:34They have their cloud environment really nailed down. So that's the first takeaway I want to talk about today is cloud.
10:40Okay. How can you run more agents, get more done? And the answer is you work in the cloud, not on your local machine.
10:50Now for the folks that aren't super technical, I'll sort of explain what does that mean. So when you're writing code or you're building things, you are editing code, right?
10:59That code has to be edited and then hosted somewhere, right? And usually what we used to do is we used to all work on our Macs.
11:08And that's called local development. And you have to set up your whole local development environment. You have to have your database.
11:13You have to have your authentication. You have to have everything local. You have to have your dev server set up.
11:20And then what happens if you want to do two things at once? You have to either use git work trees, which is pretty technical and hard to understand, or you have to literally make another copy of your code in another directory.
11:33And then so then to work on two things at once, you have to have two copies of your code, which is hard. And then what about three or four? What if you want to do five or 10 things?
11:40It just doesn't work. So what Devon does and I think a lot of good independent agent harnesses like Cursor is doing this now I think Codex is starting to do this Okay is they create what's called a VM or a virtual machine in the cloud.
11:58And it's basically a little server. And it is your development environment.
12:04And you click a button and it spins up. And all of a sudden, you can code in there. And the beautiful thing is if you can create infinite numbers of these things.
12:13So I often have at least five cloud agents working at once, often 10.
12:21And there's no risk that the code is going to collide. So what it does is it takes away all of the mental overhead of trying to coordinate these things. Like, I want to work in this part of the code, but what if it overlaps this part of the code?
12:35Or what if I want to do that? Have I synchronized and pulled the latest?
12:40Like, I don't know. And when I work in the cloud, I literally just open up my browser and I click New Session and a new VM spins up.
12:49And I never ever have to think about it. So I think if you are working locally, I honestly think . Like I think you are holding yourself back and you are shipping 10x less than you could be.
13:05And it is not smart. No matter if you think the cool people work locally and that all the smart ends, it's not true anymore.
13:14You work in the cloud.
13:15Yeah. I mean, there are a lot of people like on X who are sort of like local maxis who look at people who spin up virtual machines and cloud agents as sort of amateurish.
13:30What do you say to people like that?
13:33They are not doing real work.
13:38Here's the deal, right? So as soon as we hit PMF with Untangle and we started acquiring real customers and having to ship real fixes, real features, like I had to probably 50x my output, right?
13:52And you just can't do it if you're working locally. So I think it's nice to work locally.
13:59And sometimes you have to if you're doing heavy front end stuff. So say you're shipping a brand new feature with a lot of UI, like a lot of new UI, right? Then yes, of course, you're probably going to spin up a local agent and you're going to do some light front end, probably wireframing.
14:18But as quickly as you can, you want to move that into the cloud and then let a cloud agent take over. So I think this is happening to a lot of people. I think a lot of the traditional knowledge of how to work is out of date.
14:35That's where we're at. Okay. So if you do start using Cloud Agents and you do 50 x outputs, you know, in some ways that is a little stressful because you have so much stuff going on.
14:47Right? And now all of a sudden you have to manage more. So how do you how do you structure a system so you don't go absolutely crazy?
14:56So it's interesting. I think first of all, you have to shift your expectation of what your work is.
15:05And what your work is now is making high stakes decisions almost all day. So in the past, like we all had the luxury of making probably, you know, at most maybe two to three high stakes decisions per day, right?
15:19You'd have an important meeting with your key team or with a customer, and you'd make a couple important decisions. And you would communicate those decisions, and then they would get done over two weeks. I think you have to sort of mentally pace yourself to say, probably by lunch you're going to have made 10 to 20 high stakes decisions.
15:45And so you have to learn how to do that. So which is your question, how do you do that?
15:50And I think there's a couple simple things, right? So when I'm working with my agents, I actually have to organize my threads by and I pin the ones that are the most important work we're getting done today.
16:03Right? Because there's a lot of quick bug fixes or small things you're shipping.
16:09And you kind of learn to separate into two buckets. What are the big, big, important things I'm getting done today?
16:16Pin those. And then the other thing is just let them rip and get back to them when you can. And then set yourself this kind of almost timer, right, where you're like, Okay, I'm going to go crazy if I just click through my threads like this all day.
16:33You just get completely wiped out, right? And so it's almost this discipline of saying, I'm actually you know the phrasing smooth is slow is smooth and smooth is fast?
16:46Yep. I think there's a little bit of that where you realize, I can only check on my agents and make high stake decisions probably every twenty five minutes. And so what you might do is settle into this cadence where you just check on your high stakes threads like every twenty five minutes just so you can like literally mentally rest.
17:09And then the other side of that is just understand that this is the world we live in and it is kind of exhausting. And that is kind of where we're at. And in order to survive in this new world, you're going to work a lot more, not a lot less.
17:24And on that note, I'll share my screen because there is essentially a world we live in now where you're going to ship probably so this is my PRs, right?
17:39And we're looking these are days, right? And you can see the average here is sort of this 22 to 25 PRs a day, Right?
17:49And sometimes 40 a day. Most of these are being merged, right?
17:56A couple are being closed. And what's funny is what happened this day. I went hiking with my son.
18:03Which day is that? Can't see your cursor.
18:05Oh, it's the was the eighteenth. Eighteenth. And so we went and climbed Mount Washington, I literally couldn't access my phone.
18:13But notice immediately the morning of, right, while he was sleeping, shipped what is that?
18:18Eight PRs. So this is kind of the world we live in now.
18:27And we could talk about that. But let's zoom out to the lessons here. You've got to go cloud, right?
18:35You have to understand in order to compete in this new world, you're going to be working half the time from your phone. And it is tiring, it's like going to the gym.
18:46Okay, I have this new muscle. I never had to exercise before. And it turns out , it turns out that's the most important muscle in the world.
18:55And if you don't do it, you're not going to win. So the way to do that is in set some sort of cadence for yourself. Pin your most important threads.
19:03Write it down. So this what I'm going to sort of flash up on the screen here is my written to do list.
19:10Every day, you know, I use this hilarious analog system to remind myself what are the most important things we're going to ship today.
19:19Because there's going to be 10 probably PRs that we ship that are fire, you know, putting out fires. And and you can get distracted and exhausted. So so cloud and then exercise this new weird back muscle, which is making high stakes decisions all day.
19:37Including on your phone. Right? What you're basically saying is like, you're gonna because of the velocity of
19:45of,
19:46you know Knowledge work. Do it. Knowledge work.
19:48And you're and you're you just you'll have stuff going on and you will need to be there to say yes, no, actually this. You know, giving feedback in real time and, you know, why your point is why wait until you're in front of your Mac to give feedback.
20:08It makes ton of sense. If you're trying to hit product market fit and if you're trying to forex revenue in, you know, in a month or whatever it is,
20:18you know, why You gotta be available. Gotta be available. And, you know, I think it's easy to think of these things as humans, right?
20:24Like you have a team now that is infinitely scalable. They're going to do work very fast. And they're going to be blocked by you making important decisions, right?
20:35And you can't delegate the decisions yet. I think we're still very much even if you use Fable, like this is a trick, use a Fable thread as kind of the manager and then ask it to spin up children.
20:47And I do this a lot. So I'll spin up a Fable thread and I'll say, Okay, we want to accomplish these five things.
20:54Spin up five child Devin sessions. Don't use Fable. I use Fusion for that, which is a cheaper model.
21:00It's cool. It's like their new model where they've got Fable is kind of the parent and they've got a sidekick.
21:07Point is you're not gonna it's not viable to run Fable or any super premium model all the time. But have a sort of a parent, you know, very intelligent model run children. You can do these things, but in the end, you're gonna have to pay attention.
21:24And it's just like if you ran a big team and you were out to lunch on your boat all the time. You gotta get in the office. And the office is your phone.
21:34And so get it done. So Greg, this is my homework for you.
21:38Next time we talk, I want you to be doing more than 50% of your work on your phone.
21:42Yeah. I mean, that's why that's why you're here. Right?
21:44Like, you're a lot of people ask me, like, how do you choose guests and and why do you explore topics, certain topics over other topics?
21:53And it's it's really it's for everyone. Like, I love arming people with the information, the tactics, the systems, but I'm also everyone.
22:02Like, I'm a part of that. Right? Like, a part of that is, like, I know that you're you're doing this, and I wanna understand why you're doing it.
22:08And I also wanna understand, like, you know, one of the takeaways I have is your system for dealing with tasks is not is not too dissimilar to a very organized person pre AI.
22:21Right. Like, idea of, like, pinning, you know, your most important things, like, it's on one hand, it's pretty obvious.
22:28But on the other hand, if no one tells you, like, hey, you know, the way to to do this so you stay sane is via that system, you're you it's overwhelming.
22:39Yeah. It is funny. It's like back to the basics.
22:41Right? Yeah. It's just in an age and age.
22:45And it turns out being a very, very good engineering manager is probably one of the most valuable skills you can teach yourself. And I have the honor of mentoring my nephew right now.
22:59He's he's coming to me and asking me, how do I win? He's just come out of college soon. How do I win in this new world?
23:05And I'm basically teaching him what we're talking about on the show. It's like, Okay, your job is going to be managing as many agents as you possibly can.
23:14Get good at that, right? Which means you need to start to adopt these behaviors, right?
23:21Work in the cloud. Prioritize what you're doing. Learn how to make high stakes decisions quickly.
23:26Get technical.
23:29It is a farce that engineering is going away. Like if anybody tells you engineers are going away or are becoming less technical, they're just not doing the work, right? And it's like being a carpenter.
23:42Like say you had robot carpenters around, if you were a real carpenter, you'd be a much better manager of those robot carpenters because you understand how to make the cut, where to put the nail, and so it's important.
23:56So to back up, and if you're listening to the show, like the thing I want you to walk away is cloud agents. Get your reps in. Figure out how that works.
24:07Try it a couple times. Do it from your phone and get comfortable looking at your small phone interface and learning how to work that way.
24:18So I think that's thing one. It's very important. The second thing I want to dive into is automations.
24:25Okay. So say you're working on a cloud, well done. You're starting to separate yourself from the crowd.
24:31The second thing is automations. You need to sort of think about things that you would have normally had a meeting with once a week with your team where you check on something or daily you check on something and build an automation.
24:46What does that mean? So the first thing to do is go back to your agent, start a session, and say, Okay, I want you to automate this task every x amount of days.
25:00And so one of them that I do that's really simple is that I want to automate the test of signing up for Untangle, creating a case, onboarding a client, having that client go through discovery, basically our user experience. But I want that to be automated in a browser.
25:18This is not rocket science. This is basic user testing, right? But it turns out that good agents can do this.
25:26And so the big unlock is set up your agents so that they can make themselves better instead of you trying to build these systems.
25:39And so I set that up. And then in Devon, it's really cool. It's called a playbook.
25:43But the idea is it's sort of like here's how to do this thing. It's different than a skill.
25:49It's more of a list of things to do, how to do them correctly. And I have a it's called end to end sign up test.
25:59And that runs three times a week because it is expensive. It's probably $60 in tokens because it's doing a of browser testing.
26:11And I think that's sort of a side point I want to explain here is it's now vital that your agent can properly browser test.
26:23And it really should work out of the box, right? This is part of the reason I use Devon is because they've been using cloud agents that can do browser testing for two years now.
26:34It's so good. It records a video. It annotates the video.
26:40It then looks at its own video and it fixes bugs that it sees, right? So you have this agentic loop that is specifically around browser testing that then is automated three times a week, right?
26:52And then what do you do with that? And then you need to trigger a triage based off of that.
26:58So say that it's like, oh, wow, the third test failed in this automation.
27:05I'm going to spin up a Devin session or an agent session to then fix it. And then you have to figure out where do you know that that failed. How do you know?
27:15Did the agent just kind of silently fail and then fix itself? What about the PR that was created?
27:21So you have to start thinking about how do you see these things when they happen, right? As the manager of all these agents, how do you know what's failing and what's not?
27:32And that's, again, where you should talk to your agent and say, Okay, we've got this end to end test. I understand it runs Monday, Wednesday, Friday. I understand it triggers a child triage session if anything goes wrong.
27:46How do I know about it? And the new agent might say, well, you know, why don't I post in Slack? You know, and you could say, well, Okay, how are you going to do that?
27:54Well, we need to connect to MCP. You know, how am I going to know look in that channel? There's just all this machinery that you need to automate and figure out and work through to build that system of automation so that you are doing less of the work.
28:10So this specific automation, is this more of a, like, bug testing QA automation, or is it more of a, you know, UX QA perspective?
28:24This one is more about finding bugs that have been introduced to prod that somehow we missed in our user tests, in our automated test suite.
28:38So that one in particular is like, hey, we just can't have sign ups go down and not know about it. But sometimes your automated test suite just doesn't catch everything. But a good old fashioned clicking through the site in a browser catches that.
28:56It is not really about the UX. I think that there's no one thing I will say is there's no substitute, and we probably won't have a substitute for a while, of you using your own app.
29:12There just isn't a good substitute right now. So right now, every so often I literally go through our entire app myself.
29:22And then I obviously quickly spin up sessions to fix bugs I find. But we're just so far from agents still.
29:29The smartest models in the world still lack this sense of obvious intelligence
29:39where you're like, why did you think that was okay? Well, yeah. I think the bigger question is, you know, if if there's a division of labor between human beings and agents, what are the tasks that you as the human being are gonna be doing?
29:52What are the sets of tasks? And what are the sets of tasks that the agents are gonna be doing that you're okay with them doing and keeping you in the loop?
30:00So my question for you around automations is beyond beyond QA, what are some other automations that a founder could be implementing?
30:10So let me pull up my list for you. So the kind of automations that I have are clearly QA bug fix, fine.
30:23The other is what I call my production watchdog. So we've got a lot of real law firms using Untangle.
30:31We've got a lot of real clients going through discovery. And we've got with just a huge amount of activity.
30:38And honestly, I can't keep track of it mentally by just looking through logs. And so what I do is every day at 9AM, I have what's called a production watchdog automation, which what it does is it goes through all of the events in the database that happen for our paid customers and it summarizes that in a JSON file.
31:01And then that JSON file is in our admin. And then I know every morning I go in and I click on the production watchdog and I read through the summary of what the customers did.
31:14And it's so valuable. So it's sort of like you can imagine your chief of staff showing up and saying, all right, Greg, here's the important stuff that happened yesterday for our customers. And I'm going to roll it up and explain the things that went well and a couple bugs that we saw.
31:31So that's an automation. Essentially what happened yesterday, that was important.
31:37So that one That's a big one. Yeah. That's honestly a big one.
31:40Like it's it's like one of those ones that sounds small but is a big one because you're going you think you know what's going on but you don't, really. You
31:48don't. I mean, and it's so shocking to me. The amount of times I'm like, oh and this is the other trick.
31:56Man, so use a production watchdog automation, pick your most important customers, and then make sure that in that report it's like the customer did x, here's a link to view that. And so I actually have a production instance where you click it and you actually see what the customer was doing.
32:15And then I'll be like, woah, that's weird. And we have this amazing ability over millions of years of evolution in very quickly picking up things that are off.
32:30And so there's a couple of times I'm like, that's weird. And I click into it and it turns out, oh, the user experience that the agent built on that wasn't quite what I thought it was.
32:42And so yeah, production watchdog automation that links out to real UI is just, man, that's a game changer.
32:51So that's a big automation. This is really important, a self improvement loop. How does this work?
32:58Okay, everyone's talking about self improvement, right? So in Untingle, have an agent called Grace.
33:04And she is essentially a paralegal. And she has a lot of chats with both our customers who are attorneys, our customers who are paralegals, and our customers and their customers who are clients that are getting divorced.
33:19So how do I actually grade those chats and then improve them without managing the fine details.
33:26So what I do is I every day I have an automation that looks at these chats and then grades them on a rubric. And you and again, just talk to your agent about this. You pick the most important part of your app that you want to self improve and say, here's how you agent judge whether this thing that happens is good or bad.
33:47And then every day I want this automation to kick off. Look at those things for us.
33:53They're conversations that Grace is having. And then I want you to grade them. And then if anything is bad or below this score, I want you to kick off a child session and fix it.
34:03And it is shocking to me the number of small fine UX details that have been picked up by this loop and have been fixed.
34:13And honestly, are things that I wouldn't bother doing. Either I didn't know about them or I'd be like, meh, you know, it's a paper cut. Like, are we really going to ship a fix to that?
34:23But because the agent already identified it, spun up a PR, and it's ready to ship, I just go, ship it. And, you know, we probably ship, I would say, probably three of these a day.
34:35Wow. Yeah. That's that's huge.
34:38And, you know, a lot a lot of the backlash on on loops is is just that it's expensive. How how how have you with with Devon specifically, has it been like absurdly expensive? Like, are we talking?
34:52No. So well, so here's the thing. So last month I spent probably $20 in tokens,
34:59which is just too much. Like it's not viable. I think all of us are in a place where we were getting to the spot where it's realizing, Okay, for real engineering work per employee, you're looking at probably $5,000 a month.
35:12Like, that's probably where we're going to shake out here. And anything more than that, you really need to figure out model routing. And so thankfully, Cognition knows this.
35:23And everyone's figuring this out. Whether it's Cursor, Cognition, I would say factory, AMP, all of the independent agent labs understand.
35:33Our customers have a limit. They're not going to pay infinite token budgets. And then you have the this is me ranting a bit.
35:41But it's like you have Cloud Code and Codex who are just severely they're funding tokens and making it seem cheap.
35:49But this is not going to work long term. And so what I do is basically I have learned how to use the right model for the right task.
35:57And so these reinforcement loops, they use what's called SWE 1.7.
36:02It's a fine tuned model that Cognition has specifically built for coding. It's super cheap compared to Opus 4.8, GPT 5.6.
36:12And so a lot of these loops happen with this cheaper, finely tuned model.
36:18And we're just going to see more of that. So you're talking I don't know.
36:23Actually, I could probably look and dig it out. But you're probably talking like $5 a session on that. Okay.
36:28That's not bad. I mean, and if you're not willing to pay $15 a day to improve one of the core featured sets of your like what are you doing? Like either it's not a real company or you don't care.
36:38Yeah. Totally. Spend the money.
36:42And then figure out the system to reduce your token costs.
36:48I will say this is why I don't understand why anybody would build their engineering on a Frontier lab stack.
36:58Like I'm just going to it out. Like if you are if your whole engineering motion is happening inside of ClogCode or inside of Codecs, what are you doing?
37:07Because they are not incentivized to make it reasonable for you long term.
37:14Right? They're going to lock you in to their models, to their process.
37:18Whereas if you use an independent agent lab, you know, like an AMP, like a Devon, like a factory, like sort of a cursor, but they're kind of weird now. Those systems are going to optimize for affordable engineering long term.
37:33And I don't think you want to be locked in to Anthropic or or OpenAI, you know, solely.
37:41I mean, to be clear, you can still use their models
37:45within some of these products, right? So you're not abandoning ship. No.
37:49It's actually so it's yes and, right? So the beautiful thing is if you pay an independent agent lab like a factory, like a cursor, like a cognition, like an amp, etcetera, they're incentivized to figure out how do I give you the best results for the lowest price.
38:04So they'll say, Okay, we're going to use GPT56 sole on extra high for this type of task.
38:11And then we're going to route to a SWE 1.7. And then we're going to have a check with a Fable.
38:18Like they're going to figure out that hard stuff. And the other thing I will say is if you're trying to build a software factory for yourself, stop.
38:26The reason why you see Ramp Launch Inspect, which is their custom in house agent, is because they got to a size where they had to build their software factory in house. And if you get to a certain size, you will do that.
38:43But now if you're a one person shop and you're just building something, it's just you, you can probably get away with this $200 a month angle on a pick your frontier lab. But as soon as you graduate into, Okay, we're building a real product.
38:57We have product market fit. We're going to start hiring people. You've to build a software factory.
39:01And at that point, you want to be on an independent agent lab. Otherwise, you're going have to build it yourself, which is stupid, or you're going to pay and be locked into a walled garden.
39:15And you can't The do
39:17analogy I use for it is you know, working with an independent lab is sort of like a like working with like a mortgage broker or travel agent. Like they're gonna make a lot of calls to make sure to get you the best price and stuff like that.
39:34That's a good analogy. Yeah. To me, I use both.
39:37Know, I I'm in Cloud Code, but I'm also I use Cursor personally. Although Devin, I'm interested in. I might if people are interested, I might just do like a whole breakdown on Devon if people
39:50want to do that. Devon's so good now. Know, I just it's so good.
39:55It's amazing. But I will say I do use Codex, right?
40:00The Mac app is so beautiful. And they are subsidizing the token so much that if you want an agent on your machine just to do really cool, helpful stuff, open browser tabs and create Google Docs and check your email and do all this information work and basically get it for free.
40:22I mean, you're paying $200 and you get almost infinite tokens, awesome.
40:26But don't build your company software factory on that. Like, I just think it's bonkers to do that.
40:33We've covered a lot today. Is there anything else you wanted to
40:37cover? I feel like I'm just vomiting,
40:39you know, all of all the stuff in my head that I But no. It's important that, like, this is what I wanted. I wanted you just to vomit everywhere because, like, it's sort of like advice that you you you would give to your to your nephew, I think, or 22 year old a you 22 year old fresh out of college.
41:01It's like, what are the minds I need to know in terms of managing agents? So is there anything else that people need to know?
41:08Cloud agents, automations?
41:11Yeah. Cloud, automations.
41:14Those are the two that I want people to walk away from. I guess the third is really you will be using a software factory.
41:22And what I mean is the agents are going to be writing 100% of your code, reviewing 100% of your code, shipping 100% of your That's where we're going. And so the more you can get there, the better, thing one.
41:36Thing two. I think thing three is it is important to build a credible reputation for yourself, right?
41:46And I will say, you know, the work I put in over the last twenty years on X, like, it really pays dividends.
41:55And what I mean by that is if you if you share what you're learning publicly and you generally give, like write very helpful articles on X. And obviously, the algo really does reward articles still.
42:09So you should be writing articles on X. You should absolutely be on X no matter how you feel about Elon because sharing your knowledge really does build up credibility which will pay dividends.
42:22And I don't mean just like getting paid to post, I mean the the the network and the relationships. I mean, the reason why I'm on the show is because we built a relationship over x. And so I would encourage people, take the time to share what you're learning.
42:36It doesn't have to be super polished, but get out there and be useful to people, and it will open up doors. So you know more than you think, just get out there and share it and
42:49you'll reap the rewards, not immediately but eventually. I mean even if you don't know, you can just say I don't know and that's your and that's your strategy. Right?
42:56And that's interesting too. Amen. And if you don't know, then gas an agent help you.
43:01Yeah. So I mean, my original cohost of this show was Sahil Bloom. Sahil Bloom worked in private equity, and during COVID, he just started writing about topics he was interested in.
43:15Like, he would be like, just finance topics or, you know, a particular individual who was interesting, and it, like, it it it read as like Wikipedia articles, but really focused on Twitter specifically, like optimized for Twitter.
43:30Mhmm. And and that was his thing. And he wasn't saying I know everything.
43:33He was just saying like, I'm learning this thing. Here's the thing I'm learning. He grew in his audience over a million Twitter followers, wrote a best selling New York Times best selling book, started companies on top of it, raised a fund on top of it.
43:46So anything is possible for sure on on x.
43:51Do that. Yeah. Yep.
43:52Just just get out there and do. Do do do and learn. Yeah.
43:57You know, there there's no try. There's only do as Yoda says.
44:01Right? So Totally.
44:03Ryan Carson, I feel smarter, and that that was my goal.
44:10And speaking of X, you're a wonderful follow. I'll include links where you can follow Ryan across the Internet, including X.
44:19He's a must follow, so please do that. Thank you for coming on, sharing your generous thoughts, incredible setup.
44:29And is there anything you wanna leave people with?
44:32No. Thanks, Greg. Thanks for having the show.
44:34I appreciate you putting out the constant knowledge constant knowledge and sharing. It really makes the Internet a better place. So happy to be here, thanks for having me.
44:42Alright. I'll see you next time.
44:45Take care.
The Hook
The bait, then the rug-pull.
Greg Isenberg opens with a blunt claim: the people who learn to run teams of AI agents are about to outperform everyone else. His guest, Ryan Carson, has the receipts — a one-person startup shipping up to 40 pull requests a day, half of it managed from his phone.
How they asked for the click.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
