Transcript
Okay, as I just covered, my name's Adam Brock, and I'm going to be giving a talk called "There is No Org Chart." It's going to be about how I run a software startup with AI agents and zero employees. I'm the founder and junior vice president of Hypermatic. This talk is going to be pretty simple. I'm just going to be talking about three things. The first is how I designed my company for leverage. Secondly, I'm going to make the case that Codex is all you need. We're going to finish up by talking about the 1993 major motion picture "Jurassic Park."
I'm just going to start with a quote from my former boss, who lovingly told me that I remind him of himself when he was my age: "You're very unemployable." One of the reasons for this is that I'm basically allergic to bureaucracy. My contempt for useless meetings and fake work is almost infinite. I used to do these things at my previous company, like create these little apps. One was called "Death by Meeting," where you could input the number of people at the meeting and their average hourly rate. You would then watch the screen in real time during the meeting to see how much money and time you were all collectively wasting. It also gave you suggestions at the bottom of other things you could be doing instead of being in this super important meeting.
But I knew that I always wanted to start a business. After 10 years in digital and advertising agencies, I resigned from my job in 2019, the week that Figma announced Figma plugins. That was sort of the impetus for wanting to start my own thing. I thought I could start a company doing that with absolutely no way of proving that it would work. Now I run a profitable self-funded software startup by myself. Those three things are illegal in Silicon Valley. It's sort of a weird startup in a way because at the time, this was very peak bubble—2020, 2021. All these investors were telling me all the things I should be doing, and I just said, "You're wrong." This obviously is a really dumb way of running a business.
I'm kind of happy with where things have landed because it has all come full circle with these AI tools. The startup is called Hypermatic, and you can find it at hypermatic.com. The quickest way is actually just to Google "Agile is for losers" and hit "I'm feeling lucky," and it'll pop right up. The company makes commercial Figma plugins and is essentially a suite of 12 different Figma plugins specifically focused on automating digital marketing tasks. This involves design, code, and content automation. You can think about things like animated banners, HTML marketing emails, and file conversions—the typical sort of digital agency stuff that you'd run into.
It's grown to about 2 million users in the last seven years. A lot of the customers that I have, if I say "we," just assume I mean "I" or "me and my agents." In any case, the Figma plugins are used by many world-class companies. I feel very lucky that these companies have been sort of tricked into thinking there's this massive company behind the products they're using, but at the end of the day, it's just kind of me. I took the YC Paul Graham advice a little too far, where they said, "Do things that don't scale." I thought that meant forever. I've been working on this business every day for the last seven years on my own, and I've never hired anyone. You might call that re-target maxing, or you might call it just kind of the way things are going to go in the future. Maybe I was just sort of early.
If I was going to go to the trouble of starting a company, I should probably design a business that I actually want to work in. The company's a bit weird because it doesn't have any employees, it doesn't have any investors, it doesn't have any meetings—everything's done by email. There's no Agile™ at all. There's also no enterprise. Everything is completely self-serve, and therefore, there's essentially no compromise. That's not to say there aren't trade-offs. I've certainly turned down multiple six figures in potential revenue from doing enterprise deals, but I basically just tell them to go away because I don't want to do the things they want me to do.
This was already the foundation I had for my business for many years before AI came onto the scene. The reason I mention this is that you can't put lipstick on a pig. AI cannot fix a broken company. If your company's idea of workflow optimization is scheduling more recurring meetings until nobody has any time to do actual work, then AI is not going to save you, I'm sorry to say. It can really only multiply the existing foundation, systems, culture, and values that are already there.
One example of those values for me, from day one, has always been that I wanted to offer extreme customer service. I won't go into too many details, but if you just picture Arnold Schwarzenegger in "The Terminator," it's basically just that, but for every customer support request that I get. I will stop at nothing until it is resolved as quickly and effectively as possible, even if that means being looped into third-party providers that have nothing to do with me, like Salesforce or other horrible companies.
You can think about this as having a foundation for allowing you to add AI as leverage. In my case, customer support now becomes a full deployment pipeline. Essentially, that means before you might be able to say when something will be done, but now you can just reply, and the first reply you send to their request is that it's deployed and completed. This just leads me to say that I have not written code in 12 months, since July of 2025, when I started using Claude Code in the terminal. I've been saying this for a year to everybody that I know—my developer friends, my designer friends, all my business owner friends, shaking random people in the street, telling them about not writing code anymore.
As this line from "The Fountainhead" makes very clear, the hardest thing to explain is actually the glaringly evident, which everybody else has decided not to see. As of today, 100% of my code is written by AI agents. You're probably thinking, "Well, how could he possibly know that?" The reason is that I deleted my code editor in March of 2026, about four months ago. I decided to go all in on forcing myself to just use the Codex app. I became convinced that this was clearly going to be the future, and I wanted to force myself to get used to it and embrace it instead of doing things the way I was doing them before.
To relate it to the meetup, I think that AI ops to me kind of means turning your processes into reliable AI agent workflows. You can do that by giving your agents the tools, context, rules, taste, judgment, and feedback loops that they need to succeed. This isn't going to be a super technical talk because my setup is actually intentionally very simple. I love reading about all the new stuff online, and I love paying attention to that, but I personally don't chase every new trick. I just assume that all the most useful breakthroughs will eventually show up in the app, like Codex, for example.
Apologies to anyone who's already kind of familiar with these fundamentals, but I know a lot of people aren't. There are a few ways you can provide Codex with extra details to your agents to make them successful. The obvious one would be having an AGENTS.md file, which is essentially a dedicated README file to tell your agents exactly how your project works. Everything from how it's structured, where things go, how the deployment works, all this kind of stuff. You can get Codex or Claude to write this for you, and then you can add to it over time.
A new one is the DESIGN.md file. This is something that's come out of Google. It's kind of a spec to help you create a persistent visual identity and design system context for your coding agent. They should be able to get a sense of what your brand is, what your style is, what font you use, and what language looks like. This can be helpful when you're designing new features rather than just guessing or using the defaults that a lot of these models come with inherently.
Another really important one is repeatability. One way you can do this is by using skills. It's worth noting that these MD files are just text files. They're just explaining a series of workflows or steps that can be executed by the agent. You can create these skills just like you would create them for your employees. When you hire a new employee, you obviously need to teach them certain things about the business for them to be successful. Agents are no different. You can go from Keanu on the left, teach him skills, and then you get Neo on the right, and he's actually effective. Codex or Claude can create these skills for you.
One way to think about it is if you do something more than twice, you might want to consider turning it into a skill. You can share these skills with your team and with people outside your team. There are about a million online, so you can look at other people's skills and grab those as well. Essentially, this graph from Anthropic is one way of putting it where you can think of skills as a concrete step towards continuous learning. On day one, you've got this AI agent that's really intelligent, but it doesn't know much about your specific workflows. You would imagine that over time, as it gets more used to how you do things, it becomes much more useful.
Another important thing is to give your agents tools. Your agent's kind of useless if it's just stuck in the terminal or writing code. You need to give it access to tools in the same way that you give access to your employees' tools. You want to give it access to the browser, your email, GitHub, Slack, Figma, documents, and let it use your entire computer as well. If there's not an API for certain tasks, Codex will just take over your cursor and start doing stuff. The first time you try that, you really do kind of have a bit of a feel-the-AGI moment. Codex makes this quite easy because they have a plugins area where you can install these skills or integrations with one click. I'd definitely recommend doing that.
Finally, another feature that I find really useful is automations. Codex has automations built in, which allows you to put your work on Copilot, or Autopilot, I should say. You can essentially tell Codex what you want to do and when you want to do it. You can imagine some tasks like, "Hey, every morning at 9:00, go through my entire email and draft a reply to everything. Summarize my Slack at the end of every day. On Friday afternoons every week, I want you to go through all of my analytics and create me a report or something." The possibilities are infinite.
I want to make the case, and this is obviously going to be controversial with the Claude people in the audience. Actually, quick show of hands, who uses Claude? Okay, wow. And now, Codex? Okay, not too bad. But definitely like 3/4 Claude. I'll caveat this with saying you can substitute Codex with Claude Code, obviously. All of these things still apply. For me in my business, I found that just simplifying it and having Codex as the place for everything, I found that Codex is all that I need.
As I said, I really love all the optimization going on at the frontier, you know, with multi-agent orchestration, all this kind of crazy stuff. But day-to-day, I just want to run my business. I'm spending more time just doing that. I'm going to show you some examples of how I personally use Codex to run my business for about $200 a month. This is for my business, so it might not apply to you. My hope is that you come away with one or two ideas that maybe you can use in your own business.
The first thing to note is that Codex is actually a pretty misleading name because, as we learned from Donald J. Trump, everything's computer. You shouldn't really think of it as just an AI tool for coding. It is really a general-purpose AI agent for any work that can be done on a computer. Almost anything you can think of that you do on a computer, whether that's done with code, tools, browsers, or files, Codex is theoretically able to do that work for you. I would just try to get you to reframe Codex as more of a general agent.
I'm going to do the obvious one first, which is engineering, because that's what I've spent most of my time doing for the last seven years. Automating this is a huge gain in and of itself. I would estimate there's roughly a 30x increase in the speed of producing code, roughly a 97% reduction in code delivery time. One reason for that is that you can run these parallel agents. You don't just have one agent; you can have many agents all working on different tasks or all working together on the same task. There are newer workflows that this is all kind of converging on, where you have a main agent orchestrating other sub-agents that are working as teammates, picking off tasks in your list and collaborating with each other as you would with a regular team.
Other boring stuff that has been really helpful includes platform migrations and infrastructure optimizations. Codex essentially paid for itself in the first day that I tried doing this stuff. The provider for my documentation platform 10x their prices overnight, and I decided to migrate to VitePress. It did the entire migration in one day because I owned all the markdown stuff, and it's worked really well. It's also been great for optimizing other services, going through log files and finding things that are using resources.
One of the other most important things I use Codex for on a day-to-day basis is bug fixes and feature requests. I have a skill that effectively goes through my customer emails, reads the email, investigates what the issue or request is, writes any code for it, reviews its own work, automatically updates the public documentation, and then tells me it's done. It also tells me it's done with a draft of the email reply to send to the customer. There's a skill that knows that anytime it makes any product updates, it will automatically update an internal change log, telling me what it's done and why. It also updates the public doc site, so the doc site and the products are always in sync. There's never a question of them being out of sync anymore. This also gets indexed with AI lookups, so anytime it updates the documentation, that gets incorporated into the AI-generated answers on the doc site as well.
Another important one is customer support emails. This is another individual skill that I have, which can be called by the other workflow I mentioned. When it comes time to draft the email, it looks up this skill automatically. It knows that if we're doing emails, it should use this email draft skill. Codex has complete access to my business Gmail. It's been trained to use my exact voice. It knows how to verify facts, and it always gives me a draft to send. It will never send it without my permission. It's also really helpful for customer admin and billing. I've given Codex complete access to my payment platform, so it can automatically look up invoices, customer emails, accounts, and all that sort of stuff. Again, it won't do anything destructive until I tell it to, like canceling subscriptions.
This one makes me really happy because, as Agent Smith once said, you should never send a human to do a machine's job. When it comes to forms and compliance, I can't think of anything that makes me want to stick bicycle spokes in my eyes more than filling out massive Excel spreadsheets or PDFs and questionnaires. I just get Codex to do all these things now, and because it has all the context for my company, it's able to fill these out very effectively. I basically don't have to spend any time doing this anymore.
With all of these things, you're probably thinking, "Why are you giving access to your Gmail, your code base, doing all this crazy stuff for you?" The answer is you should always trust but verify. These agents are extremely good, but as you know, they do make mistakes. I trust it quite a lot, but I am always the one who, at the end of the day, is making decisions about when to deploy things and when to send things to customers. If something's gone wrong, it's my fault, not the agent's fault.
Some other things I use it for include building internal automation tooling for my business. These are things like scripts, command line interface tools, Chrome extensions, Figma plugins, and just a bunch of one-off tools to automate something that happens to be needed on the spot. One small example is I've got a little website automation CLI tool. I can do things like paste in a YouTube video. I've recorded hundreds of YouTube videos, and I just paste that in. It automatically transcribes it, adds it as a new page on my website, and does that in about 10 seconds. Years ago, even before ChatGPT, I was doing this all myself, and it used to take hours. Now it's down to 10 seconds.
Another handy thing I've found is spinning up standalone web apps, customer-facing web apps. I've done this for things like free design resources, like this email component library. I used to get emails all the time from people on the pricing page asking, "Can you send me a PDF quote of the pricing?" I had this template where I'd have to fill out the price and copy-paste it from the page they linked me to, scroll down to the pricing table. They just love PDFs. I thought, "If you love PDFs so much, you can generate your own PDFs now with this tool." I spent about 30 minutes live coding this PDF quote generator, and it's saved me probably an hour a week in these PDF generations.
Just other things like license management and making customers' lives a little easier or more useful. Another one, which is pretty obvious, is image assets. Because Codex is essentially just GPT models under the hood, it can generate assets in Codex as well. I figured that real slop has never been tried, so I decided to ship these Studio Ghibli anime images, mostly with machine guns and M16s, to my marketing website. I have a skill that can generate these whenever I like. It's also pretty good at website design. I iterate with Codex in the in-app browser, where you can drop pins on parts of the page on the real website. You can leave feedback just like you would with a normal review of a designer's work. Once you've left all the feedback, you just fire off all those comment pins, and it will automatically go through and make all the changes.
I also use it for developing new features and UI in my Figma plugins. I've had a pretty robust design system and component library from day one, but Codex is very smart. It picks out all the right components from the global library and wires up all the functionality. I haven't really had any issues with it going rogue; it's pretty good at knowing what components to use. If it builds its own new components, it'll use my styles, so it uses all the right color variables and stuff like that.
Another one I use it for is marketing and doing things like conversion optimizations. I have some marketing skills installed that are quite good, and also the Lazy Web MCP, which allows Codex to look up other sites or screenshots of things that might be similar, learn from them, and adapt some of those things into your own UI. It's been pretty good at doing that as well. I should also note that my revenue has gone up quite a bit since I started using these agents. Whatever they're doing seems to be working.
In general, I think that more ideas are worth trying, even if it's a small idea and kind of dumb. If you can code it in and deploy it in about 30 minutes, maybe it's worth a try. This example is just a small one. I thought it would be cool to have a pre-populated "convince your boss" type email to buy a product or something like that. I shipped that one ages ago.
Finally, I think Codex can be really helpful for higher barrier tasks as well—important tasks that you know you should do and really want to do, but just don't have the time or skill set to do them. I decided to challenge Codex a little bit to see if we could generate product demo videos. I've been wanting to do this for a year and just never got around to it. I gave it access to ElevenLabs and HeyGen for the audio and avatar AI. I basically went with no other context. I just said, "Go to my website, take my existing plain screen caps, write your own scripts, edit the videos to match everything, and just do whatever you want."
Once I got it working, I could roll it out to all 12 of the Figma plugins in an afternoon. This is one example of what it kind of came up with. You can judge for yourself if it's any good or not.
"Create all your banner animations with a simple timeline inside Figma. Preview every movement and build campaign creative much faster without touching code. Export ready-to-run banners in the formats your media workflow actually needs, including HTML, MP4, GIF, WebM, and more without rebuilding assets by hand. Bring videos, motion, and richer creative into your banners without wrestling with extra tools, and keep everything inside your Figma file while you work. Share polished browser previews that make reviews simpler, help clients approve faster, and keep the handoff between design and production moving smoothly. Animate banners in Figma, export polished creative for every major ad format, and launch faster with Bannerify."
That's kind of my HTML vision. I got it to do that video component as well, where it automatically does the play button stuff. I got it to blur the end of that video in the actual rendering process, so it could add the overlay for the clickable CTA, which takes you to the pricing page and stuff like that. It's way better than nothing. If I didn't have AI, I would have let this sit there for two more years and never got around to it. It can kind of unblock you in that way. Obviously, this can be improved. I think this is HeyGen version 3; they're up to version 5 now, but it's a tiny avatar, so it really doesn't matter that much.
These are just some of the many examples. This talk was like an hour long before, so I cut it in half in terms of examples, but these are some examples of how Codex helps me run my entire business—from customer support, product engineering, documentation, design and creative, marketing, SEO, research and strategy, admin and compliance, and of course internal tooling and automation. I would estimate that 90% of the tasks I used to do are now done by Codex. This allows me to have more time to work on the business and less time getting stuck in the business.
One way to think about that is all of these things kind of compound. Even if you think about the bug-fixing task, the fact that I don't have to be hung up on a particular bug for hours, diagnosing what's going on, and having Codex do that entire process of diagnosing and then fixing it in 2 to 3 minutes is a massive gain when you're dealing with many threads per week.
Having said all of that, even with all this automation, the hard parts are still hard. It's still hard to get product-market fit. Sales, marketing, and distribution are notoriously hard. Getting customers and retaining them is hard. Deciding what to build and why is very hard, and overall strategy or how you're going to win might be the hardest thing of all. The weird part is I've noticed that I work about 20% more after I started using Codex and these AI agents, and I don't think this is uncommon. I think a lot of developers run into this. One reason is that I can ship so much more. In that way, you can think of Codex as a backlog destroyer.
I saw an interesting post in the Agile subreddit where this guy was melting down because he felt like he was a historian documenting things that have already happened. He was creating sprint boards and all these things, and he said the developers get the task, and 20 minutes later it's deployed. He wondered what he was even doing anymore. It really changes the way I think a lot of companies are going to have to plan out their work.
The weird part is that it also works the opposite extreme. It kind of works both ways. I can also work roughly 90% less than before. Technology is miraculous because it allows us to do more with less. One concrete example of this was last year in March, before Claude Code, before I was using these AI coding tools or agentic tools. I was still spending about 3 to 4 hours a day just maintaining the business, kind of in maintenance mode, not taking on any new projects, just doing the day-to-day stuff to make sure the business didn't burn down. If you contrast that with last month, where I spent about 30 to 40 minutes a day doing those same things, you can see the difference in just one year.
This is nothing; this is just going to keep going and getting better over time. For me, at least, the mythical 4-hour work week is actually weirdly now possible. I was able to travel around South Korea with the business in maintenance mode, keeping all the customers happy and all the requests delivered every day. Because I had Codex, I was able to do this with far less time than I only had to do a year ago.
I was trying to have a little bit of a break from all this AI stuff. I was just jacked into X, absorbing all this stuff, thinking I needed to get away from AI for a while. But Korea is so AI-pilled; everywhere you turn, there are pictures of Dario and Sam Altman. I couldn't escape. I picked the worst place to go to get away from the AI stuff.
That leads me to say you can use Codex anywhere, and of course this also applies to Claude. I can control my computer via the mobile app from anywhere. Last week, I was at the movies during the day, and the trailers are like 25 minutes now. I was on my phone, writing a massive prompt to Codex, and by the time I got home, it was all done. It was a 3-hour task, and it had finished by the time I got back. It's a wild time to be alive.
I really want to hammer this home: AI only multiplies your existing foundation, systems, and values. The healthier your foundation is, the stronger that multiplier is going to be. If you want to stop bullets, you should probably make sure your company's in good shape.
As I promised, I'm going to spend the next few minutes finishing off talking about "Jurassic Park." I'd recommend checking out a couple of documentaries. One is called "Light & Magic," which is about Industrial Light & Magic, the company behind George Lucas' "Star Wars," "Jurassic Park," and "Terminator 2." Another one is called "Jurassic Punk." The thing that resonated with me is that at that time, it was really like what we're going through right now. All these model makers and animators, stop-motion animators, were really talented at what they did. They loved their craft and did amazing work. But at the same time, in the same studio, these rogue people who worked on "Terminator 2" were convinced they could build dinosaurs in 3D, and nobody wanted to believe it. They thought it was impossible. When they showed it was possible, the other people did everything they could to try and stop it because they didn't want their jobs to be taken.
You got this weird hybrid of the two. The future was just there; you just had to look at the screen. You really had to want to see it. A lot of these people chose not to see it and ironically were wiped out like the dinosaurs they were modeling. Today, we have something very similar. It feels like we're all kind of watching two different movies on the same screen. We all have access to the same facts, but there seem to be vastly different narratives depending on who you talk to.
In terms of the AI anti-AI contradictions, there's this weird sort of Schrödinger's AI where all these contradictory things are true at the same time. AI is so useless that it should be dismissed, but it's also so powerful that it must be regulated. It's so fake that it clearly must be a bubble, but it's so real that it's going to take everybody's job. It's so dumb that it's basically just an auto-complete that predicts tokens, but it's so smart that it could even end the world. It is so unprofitable that it's obviously a scam, but it's so profitable that it must be seized by the US government. Finally, it is so dangerous that it must be destroyed at all costs, even if that means nuking the data centers from orbit. It's the only way to be sure.
One quote that stood out to me was from the guy who came up with responsive web design. He wrote an article where the subtitle was "Artificial Intelligence is a Failed Technology, and It's Time We Described It That Way." Note the date on this: September 2025. I would correlate this date with a couple of graphs. This arrow points two weeks before the number of Claude Code daily commits grew by a factor of 42,000X. It was also posted two weeks before the Anthropic revenue run rate went vertical from under $10 billion to $44 billion in six months—the fastest growing company in history. The market is telling you that the value is real, but people ideologically are blinded and captured and don't want it to be true.
The real bottleneck here is not intelligence at all; it's much more human. As we've heard from the people I talked to before who are just now getting approved to use Copilot at their company, AI moves far faster than institutions do. Bureaucracy slows down this adoption massively. Regulation, as we've seen in the last couple of weeks with the US government effectively now being de facto regulators for every new model, is also going to bottleneck AI. Individual gains do not automatically become company gains. It's a very natural human tendency for status anxiety, just like the Jurassic Park people. They don't want to lose what they've built up, and they create a massive amount of resistance.
In this reality, the ultimate advantage actually becomes initiative. One concrete example that I'll end with is that I get emails from companies making me confirm that my products have absolutely nothing to do with AI. Even if I tell them the only feature is a copy prompt to clipboard that makes it easy for them to copy a prompt to the clipboard, this isn't good enough. I ask them, "How do you make sure your designers do not manually copy text content from Figma and paste it into ChatGPT?" I'm yet to hear a reply back, so I'll let you know.
You have a choice to make, just like the Jurassic Park people. Everybody is going to be AI-pilled by default. The question is, which pill do you want to be AI-pilled by? You can take the blue pill and go back to believing that AI is useless or evil, or you can take the red pill, leap into the future, and feel the AGI. It's important not to be shortsighted and think that the reality of today is what the reality is going to be like forever. By 2030, your workflows will be running on dramatically more capable API AI. You should probably design your company to benefit from more intelligence over time.
What did we learn? We learned that AI will not fix a broken company; it can only multiply what's already there. We learned that Codex and Claude Code, to my Claude Code fans in the audience, are not just for code. Everything's computer. Anything you can do on a computer, you will eventually be able to do with these tools. Workflows are the real asset because these models will keep improving. If you've seen what the new models like Fable can do, you would already be well aware of this.
Finally, I would reiterate that you do not need permission. I don't care what your company says. I don't care if they say you're only allowed to use Copilot. Just use whatever you want to use. They're going to be the ones left behind. You should just do whatever you want. If they fire you, start a new company or go work at Cognition. You do not need permission. I suggest trying to start small and automate one single workflow tomorrow. See how it goes. I think you'll be surprised at what it can do.
My final quote is from Morpheus in "The Matrix." You have to let it all go—the fear, the doubt, and the disbelief. Free your mind. I'll see you on the other side, I hope. Thank you very much.