MEGA.dev live kickoff with Kent C. Dodds, John Lindquist, Theo Browne and Angie Jones
Greg hosts the MEGA.dev kickoff, with Kent C. Dodds, John Lindquist, Theo Browne and Angie Jones presenting their weeks and answering questions. Kent demonstrates setting up an agent-driven autopilot workflow for his Media RSS project, covering verification, autonomy and product decisions.
videoKent C. Dodds, John Lindquist, Theo Browne and Angie Jones · hq.mega.dev · Saved Nov 19, 2026
mega course, coding agents, agent permissions, agent-generated pull requests, production deployment
Transcript of the recorded MEGA.dev live kickoff (Monday, Oct 26 · 96 minutes), Greg, with Kent C. Dodds, John Lindquist, Theo Browne, Angie Jones and Joel. The recording plays in MEGA Student HQ (needs a login).
MEGA Kickoff - transcript
Duration: 00:00:00 - 01:36:35
Speakers
| Label | Speaker | Role |
|---|---|---|
| Speaker 0 | Greg | host |
| Speaker 1 | Kent | author, week 3 |
| Speaker 2 | Joel | moderator |
| Speaker 3 | John Lindquist | author, week 4 |
| Speaker 4 | Theo | author, week 1 |
| Speaker 5 | Angie | author, week 2 |
Speaker attribution follows the original diarization.
Contents
00:00:00Opening and agenda00:01:40Kent - pre-recorded demo: Media RSS on autopilot00:10:44Kent's week (Product) - conversation with Joel00:19:25John - pre-recorded demo: Pi, Herder and Pine00:37:48John's week - conversation with Joel00:46:00Theo's week - conversation with Joel01:03:32Angie's week - conversation with Joel01:14:33Mega HQ walkthrough01:22:26Q&A with all authors01:33:11Mega format and closing thoughts
1. Opening and agenda
Greg 00:00:00
Welcome, everyone. Uh, it's great to have you here, uh, in Mega. It's really a great pri-privilege to be able to deliver this, uh, program to you. We're really happy that, uh, all of the hard work that we put, uh, can be now, uh, seen by you. And basically, we, uh, will explain a lot, um, about what's happening in Mega during the, uh, course of this meeting. Uh, I will have a short part after, um, our authors', uh, presentations to, uh, tell you a bit more about how Mega is organized, what we have inside, and how it works. So everyone, including our students, you can get to know, uh, how to use the HQ, how to, uh, how this is all, uh, happening. And, um, the next, uh, 90 minutes will go, um, like this. We'll have some presentations from, uh, from Kent and, uh, John that, uh, we have prepared for you. So those will be pre-recorded, and then we'll talk about, uh, each week of Mega and, um, how it's prepared and how we're gonna lead it. And then, um, I will, uh, do the walkthrough, um, about few minutes, uh, to show you where you can, um, uh, join and how it looks, how the HQ, uh, looks, and, uh, how you can participate in Mega. Uh, and then we'll have Q&A, fun part, uh, with all of our authors. So I hope that you'll spend this, uh, evening, um, uh, in a really, uh, good and meaningful way. We're really super happy to have you all here. And we're gonna start with the first pre-recorded, uh, section of, uh, Kent and then, and then John. So enjoy.
2. Kent - pre-recorded demo: Media RSS on autopilot
Kent 00:01:40
Let's get this project on autopilot.
Kent 00:01:43
And so now, my agent is going to run the KCD Skills Autopilot, uh, skill, uh, or command, and this is going to tell the agent how to get this repository, my Media RSS open source project, onto autopilot. And these are the things that you are going to learn during my week of mega.dev. We're gonna talk about how to not only get your repository set up for agents to work really efficiently, but also how to, uh, make the right product decisions so that once you have this wonderful software factory that is churning out your product and any decisions and inputs and everything, but, uh, like, once that's all working, how you can make the right product decisions so that what you get is something that's actually valuable. So right now, it's just auditing the code base. It's not looking for bugs or anything. It's looking for gaps that would stop an agent from shipping changes in this project safely without me having to read every single line of code. So, can an a-agent verify its own work? Is there one command that runs the same checks locally as in CI? Are decisions written down, or are they all just in my head? Is there a safe way to roll something back? Uh, every single gap is written down with evidence, uh, and nothing actually gets changed. It's just reading things. And while it's running that, let me just show you what Media RSS is. So it is a, um, simple application for taking a bunch of, uh, MP3 or, or media files and turning it into a RSS feed. And so I've just got a silly amount of, uh, media files. These are books and podcasts and other things that are not available in, like, a regular podcast form, and I want to be able to consume them in my podcast application. And so I put them on my NAS in my closet. I've got this server running in there, and, uh, then I can, uh, create podcast, uh, RSS feeds and subscribe in my podcast player. And so I've got one for my kids. I've got one for, like, all of our, uh, family library, um, like, audiobook library, all of that stuff. But I made this project before agents and autopilot was really a thing, and so we're gonna turn autopilot on this project. All right, and it's finished. So it just ran through and audited the entire code base, and now it's asking questions. So we're gonna go through these really quickly, um, and we're gonna go in depth during my week of Mega. So, use Kodi to ship the PR loop. I'm gonna be using Kodi. You don't have to use Kodi, uh, but, uh, Kodi makes a lot of this much easier and more portable, so I'm gonna say yes. And now it's asking me things the agent must not do, and I'm not gonna go through all of these in this video. We don't have time for that, so I'm just gonna check all of these, and we're gonna go next. Uh, which risk tiers may an agent, like Kodi bot, merge without you? Today, Kodi bot merges every PR. Um, and for me, I'm gonna say, um, low and medium, uh, are the ones I want it to merge without me. And now it's asking, "What changes always need my approval?" And now, "How does the NAS pick up a new image after publish pushes?" And so this is, like, what is really cool about this is it understands my code base, and it's applying these principles to my specific code base. And so for my code base, we're gonna say it's not reachable. And it's asking me about backups, so in my case, um, the, I'll just say, the NAS is backed up separate from the app, so we don't need to worry about that. Okay, feature flags for a single user. This is, like, a single user thing. Uh, like, nobody else is using this, so, uh, no flag system. That's fine. And here now, it actually found the code base cleanup job that I have in my Kodi, and it's just asking if I want to add Media RSS to that for docs freshmans and keeping test tight and legacy reaper, all that. Um, so yeah, I'm gonna add those, uh, to Kodi, so I get this on autopilot. That's pretty cool. And here we've got agent filed issues. How are those prioritized? Yeah, let's add some labels. That's cool. Two Playwright specs run nowhere today. What should happen to them? Um, that is cool. Yeah, so I didn't even realize that I had some Playwright specs that aren't actually running in CI, so we'll run those in CI.
Kent 00:05:54
And we'll continue. All right, so going through all those questions helps the agent understand how to apply the autopilot to my level of comfort for this particular project. And that's the point, is that it's not guessing what I want. It interviews me and asks me the things that I want. Autopilot looks different for every team and every code base. So, uh, depending on, like, what it is that you're actually building, if you're building a heart monitor or, like, a pacemaker or something, I hope that you have a, um, even more strict, uh, process for what ends up in main than I do for my side project here. Uh, and so I'm keeping all of this, like, the details for all of these things really light on purpose because, uh, there's a lot of nuance to these answers, and that nuance is the skill that you're actually going to be developing during Mega. So now it's put together a plan based off of my own preferences here, and I'm not gonna take you through all of the elements of this plan. That's kind of what my week is about. And now for rung 13, it's asking, yep, let's go ahead and do that. Okay, great. We'll go into detail on what all of this is, and now it's going to create a PR for us, so let's wait for that. And here's what comes out on the other side. We've got 35 files changed. Here's what we're trying to accomplish. Here's why. Here's what changed. Here's how it was tested, all, all of this stuff. And, and there's a human checklist, all, all that. So, um, I'm not gonna go into detail today on what all of this is, um, but, uh, this is the sort of thing that we're going to talk about during my week. Now, not just how to set your repo up on autopilot, but also, once it's set up on autopilot, how to take, uh, get the most leverage out of those agents in the product decisions that you're making. So anyway, we've got, um, the CLI and different ways for the agents to check their work, so we've got a couple skills in here. Uh, we've got Git hooks to make sure that the verification is happening automatically for the agents, uh, issue templates for cleanup, uh, types of things and friction, uh, issues, and we've got workflows for CI and our pull request template. We'll also talk about this. And we have our decision records, or the ADRs, architectural decision records, and how we do rollbacks and contributing, and the primitives documentation, also super critical. Uh, and we've got our testing all set up and improved here,
Kent 00:08:14
uh, along with some tools like our feature map and the control media RSS, um, CLI that we're gonna, uh, use so the agent can verify its work better. So we're gonna talk a lot about verification, and we've improved some of our, uh, testing and everything here as well. So I'm not gonna walk through every file or justify every decision. That's what my week on mega.dev is for. So I'm just gonna squash merge, put media RSS on autopilot. Boom. Now, this repository is on autopilot. Awesome. So here's the thing. Setting that up was actually pretty easy. You watched me do the whole process. You could definitely do this yourself in a matter of a few minutes, maybe an hour. Kind of depends on how good your testing setup is and everything. And most of you watching this, you probably have a good testing setup. So the hard part, though, the, like, the actual skill is all of the questions that I rushed through. How much do you let agents do on their own? What do they do? Or, or, like, are they able to merge without you? Uh, where do you insist on having a human? And some of these things are not going to just be your decision. Some of these are going to be conversations with a team. And so there's no universal right answer here. Typically, you're going to err on the side of being more protective, and then you build up trust eventually. Uh, but the, the skill is matching the level of autonomy that you give to your agents to what you're actually comfortable with, and then raising that comfort level on purpose as you build trust and better guardrails. So that's what my week of Mega is all about. Mega is a four-week program for developers who are already building with agents, and my week is product week. That's November 9th through the 13th. So we're going to go deep on turning intent into agent work, verifying what ships, uh, and that what ships actually works, and building products that keep improving even after they're released, and you get this, uh, flywheel. Uh, it's very exciting. So if you watched this, uh, and thought, "I want that on my code base," then that's awesome. You can run the command right now, and you'll get a long way through it. But if you wanna really understand why it's all working, uh, then, uh, and, and how to tune it for your specific team and your comfort levels and how to trust what your agents ship without reading every single line, then I need you to come to Mega. I need you to come to my week, and I'll show you how it all works and how to set it up yourself. Mega kicks off today, so I hope you have your boss on speed dial. Go to mega.dev and sign up right now.
3. Kent's week (Product) - conversation with Joel
Greg 00:10:44
Uh, we only have one more pre-recorded, uh, section from, um, John, and this is because basically we don't want to wait for the inference. Uh, but otherwise, we will just have a conversation with our authors. So I will pass, uh, to Joel and, uh, Kent that will talk about, um, about, uh, Kent's week, um, so that you have, uh, a bit more information about that, and also you can ask questions. We will be answering on, uh, on the chat.
Kent 00:11:13
Thanks, Greg. I'm excited to chat. Um, Joel, are you there?
Joel 00:11:18
I am here. How you doing, Kent?
Kent 00:11:20
Super. Yeah, I'm so excited. This is-- Launch day is always exciting. [laughs]
Joel 00:11:25
What, what, what does product even mean?
Joel 00:11:28
What does that mean to you in, in the age of AI? Like, like, what... You know, like, when you say that, and I, and I think about it, like, 'cause we, we're, we've been developers, and we've worked together for a long time. And I know, like, your focus has kind of gone away from, from syntax, right? Like, we've moved, uh, I think we've moved, the world has moved on, not just us, but, like, it feels to me-
Kent 00:11:48
Thank goodness
Joel 00:11:48
... like, in a lot of ways, the world have moved on from syntax, and now you're telling us product, product engineering. And I'm curious, like, what does that mean, and what does that mean for somebody that has spent-
Joel 00:11:59
... a year, five years, 10 years, 20 years, um, in some of our cases, you know, like, like bunkered in, learning syntax, like really, like absorbed and, and even like at an identity level, we're developers, right? Like, so what, what does that mean now to you in terms of, of product and, and the shift?
Kent 00:12:18
Yeah. Well, product was really the only thing that ever actually mattered, and everything else was just the supporting characters in the, the product journey. Um, but, uh, like the, the basic definition of a product is the thing that people are paying for, um, or the thing that solves the problem. It's the outcome. And for decades, we were all concerned about the output, um, for like, we're writing this code or writing this syntax, um, do we use semicolons? Uh, is it tabs or spaces? Like, what font is that? You know, like all of these things that are just kind of, uh, s- s- and, and some of them really did matter, um, with regard to, uh, the performance and, and the way that-
Joel 00:12:56
Semicolon specifically, and the answer is no.
Kent 00:12:59
That's right. Yeah. Anything else looks ugly. [chuckles] But, uh, but yes, uh, 100%, it, it's, um, the stuff that actually, um, solves the, the problem for the user. And how that impacts software developers is, um, we now have more time to focus on what we should have been focusing on, uh, the entire time, which is building a system that, um, satisfies the requirements of today, uh, and also keeps in mind the, um, the goals and the vision of, like long-term vision for the product. Um, and so, like one of the, the most consistent things i- in software is change. And so building a system that can adapt to that change, um, is, is still a challenge. Uh, agents are not there quite yet. Maybe eventually they'll get there. But as, um, as for today, one of the most critical things that you need to learn is how to design a system in which agents can deliver on the product and where you're focused on the outcome and less on the output.
Joel 00:14:03
Do you think we're still developers, like at this point? Like, is, is that like, like a term that, that still holds up? Are we still developing software even though we're, we're doing it with, with, um, an agent or, uh, I guess these days, swarms of agents?
Kent 00:14:18
Yeah. Well, I mean, I always called myself an engineer because the research says you get paid more, um, that way. [chuckles]
Joel 00:14:24
Yeah.
Kent 00:14:24
But, uh-
Joel 00:14:25
So that's why it's like product engineer, right? Like, so you, you are an engineer and, and it's just kind of changing the focus of, of what, what
Joel 00:14:32
that is and, and kind of moving it forward is, is how-
Kent 00:14:35
That-
Joel 00:14:35
... that's how I'm, like personally absorbing the change is like I'm, I'm still, I still feel like fundamentally doing the same
Joel 00:14:43
job. It's just doing it, doing it differently. Um-
Kent 00:14:46
Yeah, yeah. Precisely. And, uh, people who had the right, um, uh, the right mindset for the work that they were doing, um, not a whole lot has actually changed. Um, maybe the, the specifics of the day-to-day tasks or what you're typing on your keyboard, that absolutely changes. But what you're thinking about and the meetings that you go to, even if you don't like meetings, and like the discussions that you have and the things that you think about, um, that actually doesn't change at all. And, um, so are we still developers? I guess it depends on, on how you define what a developer is. Um, I, I think that there is kind of this interesting, um, kind of combination of skills, um, and I'm starting to just call us builders or, or problem solvers in general, uh, that we use software to solve problems. Uh, what I think is really interesting is, um, observing that, um, the harness is the software or the assistant is the software. And, um, so all of our users are using their assistant, and that is how, what communicates with us and the services that we offer. And so, um, the, the ultimate goal, the, the outcome that we're looking for is solving human issues, um, but the, the mechanism by which we do that is by targeting the human a- uh, the agents that the humans are using. Um, and so, yeah, it's a, it's an interesting time, uh, and the future is, uh, exciting to think about. [chuckles]
Joel 00:16:11
So, uh, and to kind of, uh, circle into your, your week and, and what you're gonna teach, and, and I think you're, you're not specifically teaching us about, about Cody, and, and I know that's something that you're, you're excited about. And I-- But I'm curious how it relates and like what you've learned, um, from building Cody and, and your a- agent ecosystem, um, and then like thinking about that from a product perspective, I think is, is also kind of cool.
Kent 00:16:33
Yeah. Yeah. So my week is not, I'm not gonna teach you about Cody, um, but I will reference Cody, and let me explain why. So the, um, the entire time that I've been teaching, over a decade now, the way that I do this is I go and do something real in the real world, uh, and then I teach people how I did that. So, uh, rebuilt my website, uh, with, uh, ridiculous, uh, re- requirements, um, that s- like emulated the sorts of things that I did when I was at PayPal. Um, and actually when I was at PayPal, figured out the right way to build out, uh, testing infrastructure and, and the patterns and practices, and that became testing JavaScript. And so it's always just been a, like, let me apply the things that I think are the very best way to build software and deliver excellent user experiences. And that takes a long time, like takes months, maybe years even. And then you package that up into, um, a kind of simulated practice so that people can learn that. And so when I decided earlier this year that I'm gonna be doing product engineering, um, then it became, okay, well, I guess I, I better build a product so that I make sure that I am teaching people something based off of my own experience and a- as well as the experience of others. And so that's also why I started the Become an Epic Product Engineer podcast, so to fill the gaps that I didn't even know I had. And so now with the months of experience in, in building that, now I have paying users. We're actually cash flow positive as far as unit economics per user. Um, I, I feel like I've got a lot of experience that I can share on what it takes to build a product with the latest and greatest, uh, tools that we have our, at our disposal. So that's what my week is gonna be all about. It's the, the lessons that I've learned about, um, AI native product development.
Joel 00:18:14
So what's the, what's your favorite or most important lesson that you've learned about AI-native product development?
Kent 00:18:21
Um, that outcomes matter a lot more than output. And so figuring out ways that you can be reviewing, um, or, or, um,
Kent 00:18:32
finding ways that you can leverage agents to do the parts of the job that don't need a human. Um, I guess that's kind of related, but separate, separate, uh, skill. But that, like, that matters a, a lot. You can get just a ton of leverage when you put your, uh, your code base on autopilot, like I showed in that, that demo video. Which, by the way, you can all run that script right now, and it will do an audit on your code base and tell you where your gaps are and get you to, uh, whatever comfort level you are, uh, with autopilot. And my week is gonna be explaining all the decisions that go into, um, getting more leverage out of your agents so that you can focus on the product side of things. So yeah, getting leverage out of agents, that's what my week is all about.
Joel 00:19:15
And just so people know-
Greg 00:19:16
I, I have to jump in real quickly.
Joel 00:19:18
That's, that's on, uh, kcdogs, kcdskills, um, is where the autopilot skill is, so people can find it there. Somebody pasted the link in the chat.
4. John - pre-recorded demo: Pi, Herder and Pine
Greg 00:19:25
Cool. Uh, guys, because we have to, uh, move to John, um, uh, and, uh, because some of, uh, the participants, uh, joined a bit later, um, uh, so I, I will just quickly, um, recap what we are doing here. So we're trying to introduce every author and every week so that we discuss the weeks that, um, are, um, uh, ahead on Mega. Uh, love your comments, by the way. And we will, um, uh, now, um, air another pre-recorded se- uh, section of, uh, John. He will show you some of, um, his harness and basically, uh, later talk about his week, uh, a bit more, uh, as well. Then we'll have, uh, Angie and Theo, and then we'll ha- we'll have, um, a short, uh, segment where the, where I'll, uh, I will show you how Mega works, and we'll have Q&A with all the authors. So let's now switch to John's, um, pre-recorded section.
John 00:20:21
[upbeat music] So here I am working on a store, and I see that one of my tests is failing, so I'll just drop in a prompt like, "I noticed the cart total test is failing again, so please just update the expected value to..." And I'll drop in the value, hit enter, and I'll just wait for it to finish and fix it because, you know, agents fix everything. And then, wait a second, what is this? Let me escape this and turn my attention over here. Changing only the expected value would weaken a money and sense invariant, and we already have a test that exposes the same bug. Well, that was actually super helpful, and it stopped me from blowing a bunch of tokens on this conversation, and I can just focus on answering this question without polluting anything in this conversation. So I have this mini agent on the side, which saves me from being an idiot in the main conversation. And also, what if I'm in the middle of my session and a new issue comes in? Currently, I don't have any issues open, but something critical pops up that requires my immediate attention. And even though I'm busy in the middle of something, an issue can arrive, and I can immediately go to investigate it.
John 00:21:32
So after clicking that, a new session opens. It starts doing a bit of investigation for me, so I can see if that's something I need to act on right now or not. And I can check on what it finds, or I can just navigate into the issue itself, go to the actions, and send the reference over to Pi. So obviously, I have a bunch of custom tooling over here that is mostly dedicated to keeping my main agent conversation clean and focused so that it always has the context it needs, so that it can be extremely token-efficient and save me both time and money and just work better overall. And then you'll see these other panes. They're also agents, and they're agents I customized to run tools inspired by what's going on in the main conversation. Different than a sub-agent because their purpose is to display visual information back to me for the things that I might want to take action on. It's almost reminiscent of using an IDE and having a squiggly line underneath something. When was the last time you thought about a squiggly line?
John 00:22:33
Now, let's rewind a bit and take a step back because you probably recognize that I'm using a terminal, I'm using Pi, I'm using Herder. I'm essentially using this antiquated technology in this modern era of amazing harnesses and incredible designs and experiences. So let's talk about why. To me, the key piece here is right now I'm inside of a Pi session, and if I ask Pi, "Please use the Herder CLI to open a pane to the right," and I let this run, you'll see that it can figure out from the Herder CLI how to visually modify the UI that the agent is running in. And if I say something like, "Please run the following command in that pane," and I drop in my customized Pi shortcut with a bunch of my extensions and everything called Pit, then it can send that command over there and start up another agentic session. And now if I ask, "Please ask that pane to use the Herder CLI to ask this pane to summarize itself and then send a summary back to it and explain what's going on inside of the pane," then you'll see I actually have this whole system where an agent has complete control over the UI, has complete control over building other agents, and can build up an entire system of communication back and forth with these primitives of terminals, panes, Pi sessions, with essentially zero configuration. And I can do everything just by asking through dictation, and it just works and shows up. So my standard for any agent that I personally use is, can the agent improve itself and improve its own experience so that what I want to happen can happen, so that I can take the ideas in my head, just yell them at my computer, and then let them come to life? So what we need is a place to chat with our agents, so we'll use Pi. We'll need a place to manage focus and workspaces and where our agents can live, and that's Herder. And then essentially a renderer behind it all, which can be the terminal, Ghosty. Am I madly in love with any of these? Not at all. Would I switch to something better immediately if I found something? Absolutely. So during my week in mega.dev, I'm not focusing on Pi or Herder or a specific solution to a problem. I'm gonna focus on the principles behind the problem. My absolute favorite quote in the entire community of agentic developers and products is the pi.dev landing page, where the tagline is, "There are many agent harnesses, but this one is yours. And here is a tool to adapt to your workflows. Here's Pi and Pi Durable. Go make them yours." And the term yours doesn't just mean for you, it also means for your agents. And when it comes to building software factories and any sort of repeatable agentic workflow, the key is that it is your workflow, it is your team's workflow, it is your product's workflow. You take ownership, you build something beautiful, you provide that thing to your teams, and you provide the result to your customers. And the final piece, and then I'll stop yapping, is after you build that thing, is it portable and configurable enough to ship it somewhere else? Because the beauty of Pi is that you can strip it down to its bare needs, put it on any device, and have it run headlessly. You can toss it into a sandbox and tell your agents to hammer against it until it's a polished diamond. And any of the extensions you build, you can share with your team or the community. But I absolutely also recognize there's sacrifices in user experience and ecosystem, and the terminal and terminal shortcuts can just be scary sometimes. I get it.
John 00:26:08
All right, so let's get shameless. You've probably heard of the project called Reverse Engineer Anything. Just search for it on GitHub, you'll find it. And you probably run it against a product or a game or a competitor or something you're interested in to see how it was built because they did something better than you, or you're just curious about how they pulled something off. One of the projects I ran this against was Raycast, which is a universally loved launcher, which I personally think is having a bit of a difficult time finding its footing in a new agentic era where you can just dictate workflows and shortcuts. And so I started some explorations around Raycast's universal search, which has a fantastic user experience around recency and frequency and adding actions and the way that it sorts things, and just really does a great job of surfacing the things you like up to the top. Now, I don't have an answer to this question, but what does that product look like in the era of agents? If we all just spend our time typing into a single input box, isn't there some sort of scenario where the launcher is your input box, and the things in your launcher are your conversations, your agents, or something? I don't know. But I was super curious to deep dive into the experience of their universal search, as well as things like how they handle permissions on Mac, which has always been an extremely poorly documented thing on Macs and an absolute nightmare to debug and test against. So I fired off this massive reverse engineering conversation. I didn't really have a particular goal in mind. I just wanted to talk and see what popped up. And I think we all kind of do that, where you just kind of step into a project and let's see what's up. But then after the conversation has been compacted a couple dozen times, and maybe you've been in this conversation for days, if not weeks, you know who you are, you ask yourself, "What do I do with this mess? I know it's an absolute goldmine, so how do I get that value out?" Now, there are dozens of ways of doing this, and I'll show you one. And through an extension I built for Pi called Pine,
John 00:28:01
it's a play on words with Pi and panes for Pi and Herder, and a command in there called brainstorm. You can see that brainstorm is going to read the whole conversation and map its topics and suggest agents to build. So I'll let this run, and then once this comes back, it's gonna say, "Here's a bunch of ideas of agents you can extract. Nothing is built until you say so. Build one with this," and tosses out some ideas for you. So it says, "Your topics were investigated searching, indexing and ranking, what sort of Rust crates it might have, universal search," again, comparing searching in Raycast to Spotlight Search, and their file navigation and actions. So from those topics that it found, and remember, this is reading the session log, like the JSONL file that's created from a session, not the current context window, and it's pulling out these ideas of agents that it's suggesting you build and take away from this conversation. It's saying, "Do you want me to build a universal search architect? Do you want me to build a file navigation auditor? Do you want me to build a file search researcher?" And these were the main things that went into this. And what these do once you decide to build one, so let's go with the universal search architect. So I'm gonna say, "Pine build B2," and this will take a while to build, and it's gonna spin it up in its own workspace. So workspace is our concept in Herder. If I navigate over to the workspace, you'll see that it's building over here. And this workspace navigator is a custom extension I built into Herder, where I just tap a shortcut and I can fuzzy search between workspaces and jump between panes and everything, and I can share that as well. And don't worry, everything I show here or everything that I talk about during my week will be shared with everyone who attends. And you'll see it spun up this Pi session where it forked the past conversation, and now it's searching through all of the past session and building out a customized agent with all of the context relevant to this universal search architecture. All right, so this agent is done and built, and it's spun up in this current session, but just to show you in a new session, it dropped it in the agents folder. It called it Universal Search Architect. And if I say, "Please tell me how recency works in Raycast."
John 00:30:22
You now have a hyper-optimized agent specific to that task, which will find everything from your past session, all the research you've already done, know all the proper file references, and just be very specific and custom for anything you need to do. And it's just super dialed in for the exact sort of things you need to work on that flavor of task. So scrolling up, you can see it mentions selecting an item records a visit, the score decays over time, recency doesn't overwrite every match, and breaks that down for me. And I actually have it set up so that when it builds these custom agents, I'll often downgrade them to like a Luna-level agent or a Haiku-level agent because they're hyper-optimized based on those past sessions. And that brings up the difference between this approach and a sub-agent approach, which you might be familiar with from Claude Code and Codex and all the others, where I can run this agent, I can change its model, I can change all of its settings and configurations, and treat it as essentially a top-level agent with a role. Now, these could also be used as sub-agents. You can have other agents call these agents, no problem. And I don't want to start some semantic battle over the definition of a sub-agent. I'm not here to fight. I'm just here to build cool stuff. But a fun thing you can do with this if you ever run into a mistake in here is that because this was built using the Pine Agent Builder, it has a built-in fix command. And if you look at fix, you give it a way to improve this agent so the agent can improve itself, run its evaluations, and you force it to behave better. And when we talk about running its evaluations, most of the tools I build use a concept called black box testing, which I'll talk about a lot in my mega.dev week. The concept is essentially hand your agent over to another agent, tell that second agent what that agent should be able to do, and have it fabricate like 50 scenarios of things it would expect to be able to do with it. And this falls back to one of the principles of you're building agents for yourself and you're building agents for agents, so you need to set up systems around both of those directions. Again, we'll talk more about that during my week. But if I try to say fix, never respond with pros, only respond with tables and bulleted lists and other ways of displaying data, and let this run, we'll give that a second to finish, hop back over to our main session, and we can tell this we also want it to build a new terminal widget or a new web page which syncs up with your terminal sessions and shows dashboards and whatnot. A new agent would be just like we did, but without the brainstorming, an agent that you'd give instructions to. Let's go ahead with a widget. So we'll run this and tell it, "Please have this widget show all of the URLs that have been listed in the session, and if I select one of the URLs, it will open that URL in my browser, and have it automatically update as URLs come in in either the user prompts or in the assistant responses." And this will go and build this widget for me. So it opened up a new pane. It's gonna start building that view. And thanks to a system called, let me pull this open here, Open UI, which is a new standard for generative UI, which has a spec that can be rendered in all of these different languages. So for us, we're using the PyHarness one for the rendering things over in the web page, but for Pine, I also used it to render things in the terminal. So we have a bunch of pre-built components and such, so I can trust that the resulting widget of this should look decent at least, and I can iterate on it in the future. But the options for this are pretty unlimited, so you could ask it to bring in recent messages from Slack or whatever sort of integrations you could possibly imagine up. It's just up to you and whatever permissions you want to grant it. For example, if you think of all of the things that like an OpenClaw can do, or whatever your preferred personal agent is, and those sorts of integrations, and you want those kind of hooked up to widgets alongside your development systems, and they're customized widgets with interactivity that can be pushed into your other sessions, that's, that's what this is all about. So you can see it popped open a new pane down here, which has our links widget, and then it automatically closed the pane that was building this widget. So now this widget just lives next to our main widget. So I can go back and forth, and when I'm in the widget, I can check on some of the actions it gives me. I can send this back over to Pi as a reference to it, or it gave me the options of like open or run selected action. So let's go ahead and open it, and then it opened this URL in my browser, which is that reverse engineering URL we were talking about. And any time if I accidentally close it, I can type in Pine, I can type in open and session links, and open it again. And if you were eagle-eyed, you may have noticed there's also a Pine open, and I'll say session links web,
John 00:35:37
and this opened in my browser, which I'll drag over. And you can see we have all of the session links from this current session in a browser running on a dev server. And you have all of your default widgets built in as well. If we look at Git, you can see recent commits and all this sort of stuff. Super duper fun. And just to prove that our widget system is working, we'll try, "Please explain what
John 00:36:02
drop in our mega.dev is about." Hit enter, and you can see it automatically showed up over here in our links. And if we go back in our custom widgets in our browser, we have mega.dev in there as well, and a great explanation of what's going on with mega.dev. The shift happened. The only question is how deep you go. Which is obviously where you should go to learn more about this. Let's go check in on our evolved agent here. If I ask it, "Please explain how recency works," drop that in there this time. It should be low on pros, high on bullets, and high on tables. We'll see how it goes. All right, so here we go. We got some bullets, we got some tables, we got some more bullets, some bullets and tables, and pretty much just what we asked for. All right, so that's all the time I have. My name is John Lindquist. I hope to see you at mega.dev during week four. My goal is to help you build that software factory and that you can make it yours, that you can build these agentic systems that are great for users, that are great for agents, that have all these useful UIs hooked into them, and that you can help them evolve into something that you can host online, that you can hook into all sorts of integrations. I'll be taking a lot of things I showed today much further and also breaking it down into much smaller bits so that no matter what the future holds, you walk away with the tools and skills, both literally and figuratively, to build the things you need to succeed. And as always, I'm just gonna overwhelm you with examples and prompts and skills and all these take-home things you can play with so you can build on top of my projects, make them your own, make them better, and be prepared for whatever comes next.
5. John's week - conversation with Joel
Greg 00:37:48
All right. [clears throat] That was awesome. Uh, thanks so much, John. I will pass to, um, Joel and John because maybe you have one or two questions about, um, John's week. Uh, that was amazing, and I, I think that you understand now why it was pre-recorded. Uh, quite a lot going on. But yeah, just to give you idea, 'cause I'm looking at the chat, this is the idea for the four weeks, and we'll have Theo next, uh, with a short recording. But before that, um, Joel, are you here to maybe ask John, um, questions about his week? Maybe you picked something up from the chat.
Joel 00:38:21
You said one or two. I have, um, maybe 30 or 40. Is that gonna be okay, Greg?
John 00:38:26
No. No, it's really okay. No.
Joel 00:38:28
Like, 30 or 40, 30 or 40 questions. Like, I, I have a bunch. Like, we're gonna make this a whole, a whole episode. No. Um, John, I'm, I'm curious and, and I'm wondering what you think too, and I, I, I, I know I was watching the chat when I was going by and, and, like, setups like this are hyper-customized, right? Like, you're sitting down and you're thinking, like, when you open Pi and you're opening just a, a raw Pi out of the box, it's nothing, right? Like, it gives you, like, four tools, a few more now, maybe six. Um, but how do you think about that? Like, when you're, you're thinking like, "Well, okay, well, I need more. I need more to this. Um, I'm gonna have to build a tool that does everything you just showed us," right? Like, you spent, you know, that was, what, 10 minutes of, of just, like, the fire hose, like, fully blasting us in the face with your custom setup.
John 00:39:10
[laughs]
Joel 00:39:10
Like, like, what's the, like, thought process? Like, how do you, you even start to think about what it is you might need to build for yourself?
John 00:39:19
I have always been someone with a script-first mindset where I find the friction and I automate it away, and I've taken a lot of that with me into agentic land,
John 00:39:31
where even if something on a page or on a product or something I'm using doesn't meet the way that my eye is guided from one part to the other, like, I feel uncomfortable looking from the left to the right in the position of that thing. I'm like, "I don't like this thing. I'm gonna customize it because it's causing my eyes friction from going over here to over there." Like, it's my tolerance for any sort of, uh, product inefficious- inefficiency or product, um, friction is in, in negative numbers now. Like, I just cannot stand, uh, products getting in my way. And I think people in general and, and, like, people in the world, like, people have stopped using Google to search. Like, they just dictate something into ChatGPT,
John 00:40:17
and they want the answer. And I think we should feel similarly towards a lot of software development practices. Um, if people aren't using Google anymore, then why, why are we using a lot of the tools that, that are out there? Um, but
John 00:40:33
I, I will definitely phrase this, uh...
John 00:40:37
I wanna be careful with how far I push that because
John 00:40:41
this is for your personal software or your personal or, uh, business projects, right? This is the friction you find in your projects. So, whatever it is that you feel like is slowing you down, that you've always complained about, like, just tackle those things right away. Those are-- If you're gonna start building a, quote unquote, "software factory,"
John 00:41:02
um, find the stuff that you've always hated. Like, what, what do you, what do you hate about software? And just fix it. And Opus 5.5 will one-shot a lot of those things. And when it comes to choosing a Pi and, um, yeah, mostly Pi, Pi with its ex- extensions and its entire ecosystem and community, uh, there's so many examples for it to pull upon if you just use, like, a simple GitHub CLI search. And there's so many references, and there's great community. And just point it at all those references and say, "Spin up 10 examples. Let me preview some of them." Make a video of them if I'm lazy, and I'll watch the video on my phone later. Uh, like, there's just...
John 00:41:43
Yeah, just having that mindset of not tolerating those frictions or inefficiencies anymore, um,
John 00:41:52
is, is pretty much, I think, uh... I don't know if that's a superpower. Like, it's a, uh, it's, it's a blessing and a curse [chuckles] at the same time, so...
Joel 00:42:03
What do you, what do you think the danger is, like, in terms of... So, I, I think so- a commonality in all of the folks presenting here today is that, um, maybe we're people that aren't satisfied with the tooling and from-
John 00:42:15
Yeah
Joel 00:42:15
... your hyper-customized setup, and Kent has Kodi, and, and Theo, uh, basically has everything. I think he's, he's developing every single, um, point of the stack, uh, at this point.
John 00:42:25
Yep.
Joel 00:42:26
Like, where is the danger- ... where you just fall too deep into, like, the setup, the meta, like, building the thing that builds the thing versus, you know, like, like what Kent talked about where we're, we're, you know, like, shipping products to end users. Or, or maybe sometimes that is the, the product that you ship in, in some cases, but-
John 00:42:43
Yeah, and that's-
Joel 00:42:44
... yeah.
John 00:42:45
The, the weird part is drawing the line between which thing is for me, which thing is for my team, which thing is for my users. And the way that typically people build things is you build something for yourself first. Uh, if you feel like it can be evolved and escalated to something for your team, then, um, help it, like, put it online, get it hosted. And that's where a lot of things like Pi and Pi Durable, and just having modular-style agents, if you can keep them as small and customizable as possible to, so they can evolve into workflows. Um, and that's a lot of what my course is about, is having that mindset of, uh, when I mentioned a lot of people get stuck in one conversation and they just, like, stick to that one over weeks of time. Um, the more you can think of agents and their roles as things you can refactor out into... Refactor is the wrong word. I don't know. Um, if, let's see. Agents having babies, spawning siblings? What, what's the terminology there? But the, the more they can be split into their individual expertises, expertise, what's the plural? Um, the more you can reuse them for the, for your team or for your users. And I think a lot of people recognize that. That's why, like, dev tools are exploding, um, as far as people investing in dev tools companies and everything. Like, everyone finds a solution for themselves. Like, they find a friction that they are, them or their team, uh, gets frustrated with, and then they go ship something better. And then whether or not you want to, um...
John 00:44:19
I, I, I will freely admit, there are a lot of extensions that I use that are published by the community, and I'm like, "This is good enough for me." This is something that, um, I use, it falls into the background, and I don't have to think about it anymore. I didn't have to spend tokens on it. And that, kind of that same mindset is, you know, people are gonna ship products that are good enough for me and that I don't have to build. But for the ones that are, like, touch all the, the, the points of my software development cycle,
John 00:44:46
um, those are the ones where I need the most control over. And it's just kind of,
John 00:44:51
you know, building up your teams of agents, building up communities of agents, and having, having that mindset of,
John 00:44:57
uh, agents with, with expertise. And I'm not even talk- talking about sub-agents. More the, uh,
John 00:45:04
um, I don't know how much time we have, but, like, if you see how much GrokBot has, um, impacted the scene by having
John 00:45:12
many top-level agents with names and roles in a single conversation that, um,
John 00:45:19
kind of inverted the, or, or, or kind of escalated or pulled sub-agents from being nested inside of one agent out to a layer of, I know this agent has a specific role. I know it has this one conversation. Whenever I talk to this agent, I know I'm gonna be talking about that thing. Um, typically, we've been thinking about that on a project level. Now we can, uh, think about it more in what, what is this agent the best at? Um, and I could keep on yapping for, for-
Joel 00:45:48
Oh, my God. I could-
Greg 00:45:50
So, yeah. Sorry to interrupt.
Joel 00:45:51
If you need me to keep yapping, I could go down on that, that, that.
John 00:45:52
Yeah. [laughing]
Joel 00:45:52
I'm looking forward, I'm looking forward to your week. I know we're, like, we're so, we're so aligned in terms of-
John 00:45:57
Yeah. Joel and I are-
Joel 00:45:57
... this, this aspect of building and, and-
Greg 00:45:59
Yeah.
Joel 00:45:59
... it's great. Yeah.
6. Theo's week - conversation with Joel
Greg 00:46:00
Sorry to interrupt, guys, but we need to switch to Theo, uh, right now, uh, 'cause we want to hear from him, but we, we won't, uh, we won't put any more recordings. Uh, we'll just talk about the weeks. Um, uh, I think it's, uh, it's, um, uh, it's way better. And also, um, uh, after Theo, I will, I will sh- I will show you just, um, what Mega is about and how we build it, how we build the lessons, uh, so that, uh, I, I can address some of the, uh, questions from the chat, because, uh, there's, uh, uh, a lot to it. And, uh, today we just want to meet the authors and their weeks, and, uh, and you have, like, all the lessons and the content explained a lot more, uh, detailed on the platform. So I, I will just show it briefly, and then we'll proceed to Q&A. Theo, are you there? Um...
Theo 00:46:48
Howdy, howdy.
Greg 00:46:49
Hey, man. I, I need to drill-
Theo 00:46:52
How are you guys doing today?
Greg 00:46:52
... to question you about your week.
Theo 00:46:56
Of course. Yeah, but everybody's saying-
Joel 00:46:58
So I wanna hear about, I wanna hear about your week, Theo, but, like, you, uh, pejoratively, um, call us Pi people, and I wanna know, like, why. Like, before we jump into your week, like, why, why shouldn't people, like, go into that particular funnel of activity? Like, why should they stop doing that and, and maybe move to something higher level?
Theo 00:47:17
The reason to go all in on Pi, in my opinion, is because it's fun and you want to. And the biggest thing we all need to be working on right now is not burning out. Now that we have, like, the infinite work machine in all of our computers that is willing to just let us work 24 hours a day without question, we need to make sure we don't burn ourselves out in the process. And if Pi is the thing that makes you love this more,
Theo 00:47:43
absolutely you should do it. But it's, it, to me, it's
Theo 00:47:48
very much like you're doing it because you want to and for the love of the game.
Theo 00:47:53
I know for a fact that I am shipping more code than every Pi user that I know.
Joel 00:47:58
I think that's a, I mean, it's a great way to look at it. And somebody I saw recently was like, "Oh, we were promised a four-hour workweek, but we're actually getting a seven-day, four-hour workweek, but we're getting a seven-day workweek." And I said, "We're actually working 16 hours a day now, so it's more like a 14-day workweek is what it, it fundamentally feels like." And, and, like, I use Pi 'cause it's fun and I enjoy it. Fully agree with that. And I'm like, "Most people should just use T3 Code." That is literally what I tell people constantly, even though I'm doing the exact opposite of what I say, and I think that's fine. Um, what's your week? Like, what are you, what are you gonna dig into and, and show us during your Mega week?
Theo 00:48:32
My goal is to go really deep on the fundamentals so that the people who are taking the rest of the course feel like they have the, the mental model ready to apply the learnings. I find that a lot of people just, like, they understand roughly how agents work and that AI can write code in these certain ways, and if they have these files in their code base, the AI will magically know what to do. But they don't really understand how the agent sees the world or really how the agent builds its own understanding of anything. I really wanted to lean into that. So I actually start by going through the history of how LLMs started to function, how the, how they pick probabilities based on, like, what token they think should go next, and then build up the pieces from there historically so that you get to, like, come with us through the whole history. Because I honestly feel like part of why folks like us, the instructors, and some of the, like, more locked-in agentic devs that I know are able to understand it all is 'cause they learned it over the years as the things changed. Some people have some baggage from a specific error they haven't let go of. Some people deeply built their understanding at one point in time and haven't adapted or adjusted since. But if you follow the whole journey, I find it's a lot easier to understand how agents work and, more importantly, how they see things, because it's really hard to make the right glasses and build the right workspace for your agent if you don't know how their eyes work.
Joel 00:49:55
What's, like, the biggest
Joel 00:49:57
that you've observed, right? Like, maybe not the biggest, but what's a, a thing that people hold on from legacy software dev that really is, like, generally holding them back?
Theo 00:50:05
I, I think the biggest one is this idea that
Theo 00:50:10
when they look at the code, that makes it correct somehow. Like, I have personally... Like, the code I have merged that caused outages, on average, is the code I reviewed more heavily, not less, because the most sensitive code got the most eyes and still was the most likely to have problems. The power of agents isn't to write a bunch of slop that we all have to read. It is to give us the ability to take any problem that is too tedious to be bothered with, make it more tedious and more valuable in the process. Like, we've spent our whole careers steering away from tedium. The reason that we're engineers is because we hate tedious things so much that we go out of our way to automate everything. That's how we got into software in the first place, for everyone here, I'm almost certain, is you wanted to make something less tedious or annoying, so you made a more tedious and annoying decision to learn how to code.
Theo 00:51:04
We don't like tedium, and we steer away from it. Agents flourish in tedium. The more
Theo 00:51:13
bullshit the task you can give an agent, the more likely it is to succeed, to an extent. So if you find ways to take work that previously we would have fully avoided because it's tedious, make it more tedious, and then hand it to an agent, you can do crazy things. So this could be something silly like writing 10,000 lines of code for a 500-line change because you build a whole verification system to make sure it works how you want. Or something I was just working on the last few days is I spent a lot of time and tokens making an eval set for the MCP built into T3 Code, the one that lets you, like, to, like, create one thread from another, spin up sub-agents in one harness from another, all that, because I wanted it to be simpler and have fewer random errors with the agents and also less context being used. So I took every single time through my entire history for months where my agents did something I didn't like and tediously went through all of that with agents, created a subset of problems from my real lived experience, and then made a very detailed evaluation out of all that. I could have done this by hand. It would have taken weeks, if not months, of going through all of this data, combing through to find these specific things.
Theo 00:52:25
I now don't have to worry about any of that.
Joel 00:52:28
Is there something... And I, I, I'm... I think I recall correctly that testing wasn't your, like, unit tests and testing was something that you, like, maybe, maybe up to just TDD, but, like, I feel like you were... Is it like one of those tedious things where previously it was like, "Yeah, no thanks," and, and has, has moved into, like, now it's like, like a, "Sure?" Or, or the other side of that is, like, they love the tedium so much that they will write 15,000 tests that mean nothing because they're just happy to sit there and toil all day long. Like, what's the, like, what's the balance there with, with the tedious tasks that we allow them to run away with?
Theo 00:53:03
I find that both of those things are true, and it makes things strange.
Joel 00:53:07
Yeah.
Theo 00:53:08
To an extent, this is one of those, like,
Theo 00:53:10
uh, you'll know it when you see it and, like, move things around until it feels right things. Like, we let the agents write tons of unit tests in T3 Code for a while, and eventually it got to the point where we had more flakes than we had useful errors. So I went through and I Thanos snapped over half the tests away, and it has become easier for our agents to work and operate. That's really how we've been thinking about things from day one is, like, how do we unblock the agents and make it easier to make the right change and harder to make the wrong one? Sometimes that means some really heavy tests up front. Sometimes that means a crazy evaluation you're gonna send your agents to hill climb. Sometimes that means just, like, do the thing and have an agent go click some buttons and make sure it works.
Joel 00:53:57
Do you think,
Joel 00:53:58
like, programming languages generally, like, right now are... Do-- Are we gonna see, like, a Cambrian explosion of new programming languages, or are we gonna narrow down into some specific language? Like, I've, I've just been, I've been thinking about that a lot with TypeScript, which I love, and Rust and, and, you know, like, are these the, the final champions, or is there gonna be infinite languages, I guess?
Theo 00:54:19
I would be surprised if we didn't see, uh, an explosion. The, the weird experiment I've been doing was, it started with TypeScript Rust, which was my, like, white whale, the impossible AI task I've joked about for years-
Joel 00:54:31
Kudos, by the way
Theo 00:54:32
... is having an agent rewrite the TypeScript language in Rust.
Theo 00:54:36
It took a few days to get it working, and it took under two weeks to make it faster than the official Go rewrite. That was unfathomable just a few weeks ago. Like, three or four weeks ago.
Joel 00:54:48
That's crazy, though. It took a couple of days to get it working is, like, a crazy thing to say for that-
Theo 00:54:53
Yeah.
Joel 00:54:53
... period.
Theo 00:54:55
And two times faster than the official version in a week, and four-plus times faster in two weeks. It's just-
Joel 00:55:02
Mm-hmm.
Theo 00:55:03
It's crazy. And part of what I experimented, or experienced with this is, like, I... Most of the thread is I had one giant goal running, but then I spun up another thread next to it saying, "Audit the history here every once in a while, identify bottlenecks, and come up with ways that we can resolve those bottlenecks." The biggest bottleneck by far for ts-rust was actually the compute that I was running on, not the tokens I was getting from Anthropic. It was $500 of subscription usage to do the whole rewrite, which is still crazy. Like, yeah, of course, it's subsidized, but it's not... Like, like, the subsidies are gonna be matched by, like, the prices going down very quickly. So, $500 of money spent on the tokens, 10 to 20K retail for the API prices.
Theo 00:55:46
I spent over two grand on CI and VPSs to run-
Joel 00:55:50
Yeah.
Theo 00:55:50
... all of the compiler checks and everything for that project.
Theo 00:55:56
That inspired me to go further, and now I'm rewriting the Rust compiler, making a subset of Rust that compiles significantly faster. And it works. So, I have made a bespoke language to compile my bespoke compiler, to compile and type check my existing c- like, code bases. And I would be surprised if, like, I can do this goofing off for fun right now, why, as an industry, would we not do this in the future? Like, we've always had to rely on software having all of the core functionality we needed in order for us to consume it, including languages. Like, if the language you're using is missing features that you need for your project or your work, that's awful. So, every language has to have every feature. What if you can trim and right-size literally everything in your stack from top to bottom? Not only do I think we're all gonna be writing our own compilers for our bespoke applications, I think we could be writing our own languages. I think we could be writing our own operating systems. Like, I would not be surprised if a meaningful portion of new apps were deployed on a custom distribution of Linux in the next year.
Joel 00:57:00
Yeah. Um, I've been writing, like, a custom Lisp for myself, which is silly, and, and nobody else is gonna use it. But there's this, also this notion that we're all gonna be... Like, we have access to unlimited intelligence. Everybody is going to be in a personal zone of discovery constantly,
Joel 00:57:17
and that means that we, you know, like, what, why would you pay attention to what anybody else is doing? Because you are in your own zone of discovery. And I think, like, from a philosophical standpoint, that's like a... It's, it's, like, mind-blowing, but also, like, feels like it's going to maybe be at least partially true.
Theo 00:57:33
Yeah. I think that one of the, like, skills I had that helped me a lot throughout my career was knowing when and how to tap out on a thing. The silly example I give is keyboards. I used to be a huge keyboard nerd, where I, like, would build my own mechanical keyboards at least, like, a few times a year. I got the Poker 3 by Vortex, which was, like, a 60% with a decent, like, built-in layout, and I was immediately just like, "This is good enough. I've lost all motivation to care further. I am happy with this." And I have not built or customized a keyboard in, like, 10-plus years now. That's something that's allowed me to, like, not get nerd sniped to the point of self-harm historically, where, like, I could get into the thing enough to be practical and useful with it, build something cool with it, and not just keep going down the rabbit hole.
Theo 00:58:23
Now, I am letting myself go down the rabbit hole more, where I go from trying to speed up the builds in T3 code, to rewriting the language server that we use in it, to rewriting the other languages compiler, to rewriting the operating system this all runs on. And all of this not only is, like, more, uh, like, justifiable than ever, it's also just more fun and more successful than I ever could have imagined. Like, I'm genuinely annoyed and frustrated that T3 OS has become as good as it is, and I really need to just lock in and get it out, because-
Joel 00:58:58
Yes, please.
Theo 00:58:58
... it's so cool. And it's just unbelievable that, like, you can be rewarded for going deeper like that, but also, you can become addicted to just constantly finding another layer to go to, and you need more discipline than ever to choose where to, like, draw the b- barriers and, like, where to focus on.
Joel 00:59:15
I think, uh, something that's interesting to me that, that, um, I think keeps coming back is the, the idea of the bitter lesson, or that the intelligence is going to get such that all of the complexity that we wrap around it in our harnesses and in our, kind of our day-to-day and skills and all that stuff is, is, um, eventually going to be unnecessary, whereas it isn't today. And I'm wondering, you know, like, we're not, not predicting the future, but have you seen a lot of that in...
Joel 00:59:40
As the models progress, we need less to, you know, like, put around them, or... And will that trend continue?
Theo 00:59:48
I,
Theo 00:59:49
I definitely think we need to put less around the models. We... If anything, what we need is to give more to the models, like, give them more tools to verify more things themselves. Like, I know I just mentioned T3 OS, but the fact that I gave every thread on T3 OS its own window on the desktop so that each one... It's not like a VM per thread. It is a window manager per thread with access to, like, the shared application layer, so once you sign in one place, it's signed in everywhere. Giving my agents free rein to just do c- computer use however they want on a machine like that has allowed them to produce much better results. And from there, as long as you can define a hill, the agents can climb it. I have a couple videos coming out soon I'm really proud of in this regard, and I set you up for this in the course as well, this idea that, that the problem... Like, the agents are- ... trained to work with problems shaped a certain way, and if you learn how to shape problems that same way,
Theo 01:00:52
the agent will just do whatever it can to stay on that path. So if you just provide enough tools and hints that it's going the right direction, you don't need to do too much else. Like, I had a UI overhaul run that I let go for 18 hours straight, and I merged it, and the complaints are, "It's slightly too dark." Other than that,
Theo 01:01:13
18 hours later of just letting an agent run in a loop, polishing our UI, it succeeded.
Joel 01:01:20
Is the bottleneck, I guess, the ideas at that point? Like, thinking the chain of thought required from you, like, as the operator to go from this place of dissatisfaction. I feel like you're an eternally dissatisfied type of person, um, maybe. Uh, and-
Theo 01:01:34
Yeah
Joel 01:01:34
... you know, it's like you, you have this list, and it's not like previously you weren't technically capable of writing TypeScript and Rust. It would have just taken you three years or whatever, and so, like, that doesn't make sense. But now you can just do that, so like, where's the, where's the bottleneck, um, for, for the person? And, and h- I guess, I don't know, how do you... People talk about taste all the time, and I honestly, I, I get a little tired of the, the taste commentary. It's, it, it is part of that perhaps, but like, also, like, it's, [chuckles] I feel like frustration almost that, that drives some of it. Um,
Joel 01:02:06
how would you think about that?
Theo 01:02:08
The, the biggest bottleneck I feel right now is the... I, I always refer to what I call shelf space, like the amount of space you can take up in an area. And like, yeah, I'm popular. People see my tweets and things a lot. I'm sure a lot of the audience here is here because of my tweets. Thank you all for coming, by the way. Super appreciate it. The way I think of it is I effectively get, like, three slots a day that I can share things in. And if I share more than three things a day, each additional one, I, like, gets exponentially less attention.
Joel 01:02:39
Mm-hmm.
Theo 01:02:39
And I have to keep sharing the cool things we're doing and the products that exist, as well as the things I'm working on for other products. The more I release, the more confusing it gets, not just for the audience overseeing it, but for myself as well. It's
Theo 01:02:56
t- too easy to end up with too much, and that's the biggest issue I run into, is pacing everything as it comes out and making sure I actually get the thing to the end. Not because, like, the last 10% is the hardest. Like, it is, but the last 10% can be done with the agents now. It's more because the last 10%
Theo 01:03:18
is a commitment, and it's a scary commitment to make, and once the thing's out, it's out. You can't take it back the same way. So that's, for me, the biggest bottleneck, is just not having enough room to put out all these cool things that we're doing and working on, and I hope I can find ways to fix that.
7. Angie's week - conversation with Joel
Greg 01:03:32
I need to, I need to put a pause here, um, 'cause the kickoff is to get to know all the authors, and there is still Angie. Oh, thank you so much, Theo. Thanks so much. Uh-
Theo 01:03:40
Thanks, Theo.
Greg 01:03:41
Uh, Joel, thanks. That, that was great. Angie, are you here? Um, uh, can we talk, uh, briefly about your week?
Joel 01:03:47
Hi, Angie.
Angie 01:03:48
Hey. How are you?
Joel 01:03:51
I'm doing pretty good, actually. These are great conversations. I'm so excited. Um, I'm really stoked, uh, for your week, and, uh, I'm gonna let you explain to us what, what your, your week will be about and, and what information you're gonna, gonna, uh, provide for us.
Angie 01:04:05
Yeah. So I'm week two. I go after, uh, the amazing Theo. Uh, so I did get a sneak peek of Theo's content, and it's, it's outstanding. Um, and he'll go over a lot of, you know, how to effectively wield these agents, uh, as a software developer. You know, Joel, um, I really have a strong belief that very, very soon, software developers are gonna have a mandate to also know how to build agentic systems. Like, whether that's standalone agents or, you know, just capabilities within applications, but understanding how to control, like, the agent loop and everything that goes inside of that. So in my course, uh, we're actually gonna build an agent. Not a coding agent. It'll be, like, a real, uh, you know, uh, business type of agent.
Angie 01:05:04
And, uh, in that, we, we go through everything. So we're learning, you know, how to manage context, how to set up memory in your agent, and, and what that looks like. Uh, we're going through how to give your agent, uh, tools to use, whether that's tools that we're building in the app itself or MCP, and then we're gonna go really deep on MCP as well. So showing, like, how to, uh, build MCP servers with, like, a lot of advanced features beyond just tools, but things like, uh, elicitation, uh, and, uh, and task and all sorts of good things. And then we get into some multi-agent collaboration as well. So if I want to delegate tasks to a third party, not an agent that I have any control over whatsoever, but another business, uh, is providing an agent as a service, how do we communicate with that agent and delegate tasks to it and that sort of thing? And then, of course, after we've built this beautiful agent, we have to talk about how to ensure that it's actually reliable. And so we'll get into, uh, some of the, the important, maybe not so glamorous, but very important stuff like evals and observability and things like that.
Joel 01:06:28
Do you think... I-- Something I've been thinking about a lot lately, and I'm wondering what you, you, you think about this too. I, I, I feel like there's legacy software, like pre-AI, and then where we're, we're headed in the future, and, you know, like, protocols like MCP are gonna be, be huge and important. Um, what does that mean for, for people that were traditionally software developers developing legacy software compared to- ... like, the opportunities that they need to focus on, uh, to continue, like, their-- versus their career, and then, like, what's, what's gonna be built? Like, it feels like there's just a mound of opportunity and potential, but it's gonna require some, some different thinking.
Angie 01:07:07
Yeah, a bit. But surprisingly, not as much as you think, right? And I feel like that's why a lot of engineers feel like they, they don't have that, that sense of building, and they're not as much in the flow state anymore, is because they're just kind of prompting these agents to build stuff for them. That doesn't work when you need to build, like, agentic systems themself. You have to jump back in. So, um, you'll see a lot of code [chuckles] in, in my course, um, where you're controlling everything kinda around the model, right? And so this just becomes a new entity in the software that we're already used to building, where we have this non-deterministic, uh, actor that can do some pretty cool stuff. Um, but we have to build around that as well. And I think that developers will see, "Ah, this is where the flow state is," right? And so you start feeling a lot of what we felt before as builders. There's some new interesting problems to solve, but a lot of the skills that we already have as developers come into play.
Joel 01:08:17
MCP is interesting because I feel like it was something that, when it first came out, people kind of panned it a little bit and weren't, weren't nice to it, but now, like, it feels like it's good. And, um-
Angie 01:08:27
[laughs]
Joel 01:08:27
... and I, and I honestly don't think it's changed. It's changed some. Like, the protocol's changed, but I'm wondering what, what that shift was in your mind, and where-- how did MCP become something where people were kind of clowning on it a little bit, and now, like, it feels like, like people are really, truly-
Angie 01:08:41
Yeah
Joel 01:08:41
... like, embracing it, and everything's getting some MCP love.
Angie 01:08:44
Yeah. I think, I think one thing was, like, it was very, like, kind of a, a solo person sport before, where, um, you know, if you're only using it for your coding tasks or something like that, you're like, "What's the point of this? I could use a CLI or something like that." Um, now we're starting to-- people are starting to use agents beyond just coding, and now you can see the beauty of this. When you need to connect to all of your various applications, uh, that's a totally different ballgame, right? Um, I think also, we must admit that the, the spec itself did get a lot better and did, uh, add functionality for scaling, right, across, um, just, like, kinda one-to-one client-server relationships. And so, um, I think that definitely helped as well. And you're right, we're starting-- This is-- So I've, I've traveled all across the globe, um, this year, talking to companies that are using MCP, um, and it's a very different conversation than the one you'll s- you'll see on Twitter. [laughs]
Joel 01:09:49
Like, in the business versus on the-
Angie 01:09:50
In the business, right? They-
Joel 01:09:51
... on Twitter, right? Like, or like, people, boots on the ground, developers making softwares, shipping features, like, they-- there's a different-
Angie 01:09:58
Exactly
Joel 01:09:58
... different vibe than what you would see on x.com, the everything app.
Angie 01:10:01
Exactly. That's exactly right.
Joel 01:10:04
Yeah. I mean, that ma- that makes sense to me. What-- Like, another thing that I think is interesting, you talk about agentic applications and building, you know, like, applications and doing this. What, what, what does that mean? Is that, is that the chat box that, that Chase puts in when I log in? I don't even think they do that. They're pretty behind.
Angie 01:10:18
[laughs]
Joel 01:10:18
Banks aren't a great example. But, like, is that, is that the-- is the chat box what we're talking about, or what does that actually mean when we talk about agentic applications?
Angie 01:10:25
It, it, it could be a chatbot. It could be something that, um, is a service that other agents can, can call, right? Um, it could be agentic capabilities that work in the background in your app. So we're gonna explore actually all of that. I want, I want the learners to see each piece of that. So we'll have a chatbot agent that works as, like, a customer service type of agent. We'll see also why these things are so dumb, um, [laughs] and why they answer the way that they do and do the goofy things that they do, and we'll be able to tweak that stuff. As developers, I'll show you exactly what you need to do to make that better. Um, and then we'll also have, like, kinda standalone agents that, "Hey, anybody that needs me, here I am. Call me when you do." And these agents may not ever surface to an end user or a customer, but they're working with your agent in the background. Uh, we'll see that as well. And then we'll also see just, like, agentic capabilities built into, uh, the routing, the decisions. Oh, we throw a little decision, um, model in there too, so we got a little Jeb in there to show, uh, in-- within your code itself, right? Uh, you might need to go down various paths, and you need a little bit of judgment or reasoning to decide which path to go through, and, and, and decision models can help with that.
Joel 01:11:56
So part-- One of the things that I think of when you say that is, is people talk about loops, um, you know, for the last-- Uh, that, that came up this year, right? Like, we started talking about loops, and we, we had agents, and if you're not looping, you're, you're left behind. And, and I think they're-- we've moved on to graphs and graphs of loops now. But, but is that where you're t-- You know, is that part of, part of this is, is that, that looping experience, the background, like, agents working autonomously for us while we sleep? Is that, is that part of the discussion?
Angie 01:12:20
[laughs] Um, uh, so I, I think the loop is core to understanding how all of this works. I think what engineers are missing is that they look at this loop as something that just exists, and they try to figure out, like, "How do I be the human in this loop?" Um, I think that's the, the wrong mindset. What I wanna teach you is how do you control the loop, right? Um, and show you the various entry points where you get to determine what happens within this loop.
Joel 01:12:56
Yeah, it's more, I mean, not necessarily being the human in the loop, but having control over what occurs in the loop, which is two separate things.
Angie 01:13:03
Yeah.
Joel 01:13:03
Because you don't necessarily have to stop and intervene, but you need to be able to have visibility, understand what's going on. And ultimately, I think we, as the people that are, that are deploying these things are, are responsible, right? Like, we, we are still responsible to our customers and our clients and, and our, our coworkers and our, you know, like the, the stack, I think, at the end of the day. And, and, you know, like what...
Joel 01:13:24
I, I would think about that too, like the level of responsibility, and I'm wondering how we, we,
Joel 01:13:29
like, deal with that and, and, uh, what that means as, as the people that are deploying agents.
Angie 01:13:35
Yeah, and I think that's where you'll see a lot of the engineering work to be done. Um, we, we act like, I don't know, like these models, oh, they're so non-deterministic, and that means we have no control. That's not true at all. We have a lot of control, um, when we build the system around them, right? If you're not in control of building it, then of course, you have no control. But I wanna peel back all of that and show folks how this all works, um, and how to build those controls around the model.
Joel 01:14:08
Which levers that we have as the people that are building the agents to control them over time and make sure that the results that they produce are the results that we expect the vast majority of the time. I think, you know, at the end of the day, you're still, you know, like bugs happen, you're still gonna get mistakes, stuff, stuff will go, go south, but we have to, you know, like take, take control of that and, and be responsible for, for the output.
Angie 01:14:29
Yep.
Greg 01:14:31
Okay, guys.
Joel 01:14:32
Cool.
Greg 01:14:32
Uh-
Joel 01:14:32
Yeah, thank you, Angie.
8. Mega HQ walkthrough
Greg 01:14:33
Thanks so much. Thanks so much, Angie. Uh, that was amazing. Thank, thank you. And, um, uh, I would like to take this five minutes to just quickly show you and tell you, uh, the things and o- uh, and answer some questions from the chat about Mega. And then we will try to bring, uh, all of our authors together for two or three questions that I picked up from the chat there. Uh, there are some, uh, uh, nice questions. However, please take into consideration that this kickoff is just so that you can, uh, get to know the authors and their weeks and their mindset about the weeks. And then each week, you will have, during Mega, uh, each week you will have, um, a separate session, um, with, uh, the authors. And, uh, this Friday, Theo is starting with the Q&A session, so you can... You will then be able to ask, uh, him a lot of, uh, different question and, uh, questions and take his perspective. And then every Wednesday, we'll have a live session with one of the authors as well. So, there will be plenty of, um, uh, time to, uh, ask them ques- more questions. Uh, in the meantime, I will share my screen, uh, just real quick. And, um, I hope that you guys can see it. You, you should be able to see it because there is quite a few question, questions about Mega. So, I'll try to just quickly go through how we built Mega and how it is different from today's session and what you can count on. So basically, uh, if you enter, um, Mega's HQ, uh, there is a dashboard that, uh, you can access, uh, lessons from, and you have today's lesson and also you have today's brief. Uh, together with the course, we have a tool that enters your, um, uh, you c- you give instruction, uh, to your agent to install the tool. And then, um, from the console, from the terminal, you'll be able to, in your agent, uh, use the course as well. So, there, there will be assessments and there will be challenges coming to your harness, and you can use your harness to, um, uh, to, uh, do the assessments. I'll explain that in a second. But basically, this is the dashboard. And also from there, you'll be able to join Discord. There is already a, a great discussion on the Discord, uh, about Mega and how it's, um, uh, progressing right now. And, uh, basically, we have a-also the, um, hands-on challenges that you will be able to, uh, open from here. And then there is a storyline that we have in Mega. We created, um,
Greg 01:17:04
a storyline that travels with you, uh, within your harness, and, uh, you'll be getting bits and pieces of the storyline throughout the course. You can play the teaser, and you can, uh, see how the story work, uh, works here, but this is fully optional, and you don't have to get involved, uh, in all of that if you don't want to. Um, it's just like a minute a day or two minutes a day, and it's, um, uh, um, a story that we built so that, uh, we can all learn to, uh, read a bit slower [chuckles] as developers. And I think it's pretty interesting. So, um, those of you who, um, who will try it, I think they will be, um, uh, happy about how it goes. You can see the story, um, in the separate tab as well. But from the dashboard, the most important thing is basically to go to the lesson and to watch the lesson and read the lesson. So, lessons, basically in Mega, they have two parts. There is a video and a text component to it. And the video component is, uh, for example, in week one, there is Theo lessons. They, they are roughly like 20, 30 minutes. Some lessons are longer. And, uh, the way to use Mega is basically to, to, uh, watch the lesson and then read the text. This is not the same text. It's completely different from the lesson. You can play this, uh, back, um, and, um, listen to it, or you can read it, and I recommend reading it, um, because then you have the visualizations as well, the screenshots that you can click to enlarge. And this is basically the core content of Mega. So, what you do, um, every day is you log in here to the dashboard, or you go to your, um, agent, and then from the terminal, you put Mega start. And once you put Mega start in the terminal and you have your harness connected, and, uh, by the way, um, you can, uh, check challenges, uh, with the instructions on how to set up your harness here. So, once you are connected, every day you can run Mega start, and then we will do some- ... things to connect, um, the lesson, lessons content to your way of learning, uh, because there is also an assessment that we ask you to do, and you can also do it from the homepage. And once you do the assessment, uh, we will get to know, uh, your habits, not your projects, nothing personal, but basically how you work. And from the assessment, you can also get to know your way of working in terms of mega ladder that we built. And then once you use this command, mega start, at the beginning of your day, then we will, uh, connect your, um, scores from the assessment, not scores, but basically your profile.
Greg 01:19:42
Uh, in my case, this is from 150 sessions. We will connect it to the lesson, and we will give you the brief. So, every day, you will get personalized brief that you can read how this lesson is useful for your harness, for your content, for your context of, and the way of working. So, uh, you can read the brief, and then you can proceed to, to the lesson. Then you can watch the lesson and read the lesson, and, uh, once you are done, uh, you can, um, uh, check out the lessons challenge. So once you, uh, click on the lessons challenge, uh, and, or you can do it from your agent as well, you will have information about what to do and how to solve the puzzle that we have for each lesson. And, uh, with, uh, not with every lesson, but with most of them, you will also be able to collect a card, uh, to your deck. There is a deck tab in mega, and here you will collect and s- and earn your mega cards that will give you mega certificate at the end. So, there is 20 lessons, there is four live events, and, uh, those 20 lessons, they all have challenges, but not all of the challenges grant you cards, just some of them, because there is 12 cards. So basically, once you collect all of the cards, you will be granted mega, uh, developer certificate. Apart from those cards, you will also have a lot of m- other cards with our gamification, the arsenal and the lore cards, but those I'm gonna leave, uh, for now. I think you're gonna explore them, um, as we go. If you have any questions about, uh, mega and how it basically works, uh, let me know in the chat or just drop us a line at hello@mega.dev. We'll be happy to answer everything. And, uh, for you guys joining today, uh, for those of you joining, please make sure that you log in here to the HQ and then join our Discord and say hello. The time to join is only, uh, till midnight, uh, Pacific Time today. So, uh, hopefully, um, uh, we can, uh, meet inside, and, um, uh, also you can decide during the first two week whether you want to stay, um, uh, with, uh, in mega and, uh, this is bring- bringing you a lot of value or not. Uh, during the first two week, you can, uh, easily resign. So, I hope you guys, um, uh, saw today's lesson, those of you who are inside. Uh, I hope it was, uh, nice to meet our authors during the, the kickoff. This is how we, uh, usually organize it so that you can, uh, take a look, um, at who's inside, and you can, uh, chat, um, a bit to the au- and ca- we can chat a bit to the authors. And now if we can, um, bring all of them in, and I think that I can stop sharing now,
9. Q&A with all authors
Greg 01:22:26
and maybe you have, um, questions that I can ask all of them. Uh, I think we can take two or three more questions so that they can, uh, they can answer. I think the, the one question, I, I picked, I picked, uh, one, but there is plenty of those. We, we will try to answer the, uh, them on the Discord. But the one, uh, good one that I picked up was, uh, what is the future you've seen, um, in terms of how code bases and product evolves? Maybe you can tell, like,
Greg 01:22:57
j- just, just one or two sentences about, um, how the, the code bases and our products will evolve. What's, what's your bet? Um, um, and whoever wants to start.
Theo 01:23:10
I can go first for that. I think we're gonna be moving away from this idea of, like, the source code being the source of truth and towards a, a vaguer source of truth, which means, like, it will have transient problems and whatnot, but the idea of encoding intent instead of encoding behavior is, I think, a direction that we'll see everything moving towards.
Kent 01:23:30
Uh, I can tell you that the product that I'm working on is, like, 1.3 million lines of code that I have built myself over the last eight months, and that's just like,
Kent 01:23:40
no, that's not a thing that we used to be able to do ever in any context. And so, the, the future of the code base, I think, is definitely, um... I mean, it, it, it, the size is not what matters. It's the out- outcome of what you're building, right? Um, but I think that we're able... It... My vision for the future is that software is invisible, that, um, you, you, you don't talk about software. You talk about problems that you have, and the software is, like, an implementation detail of how you solve that problem, and the agents are gonna be the ones managing that. So, what is the future of the code base? Uh, I think that, uh, it's, um,
Kent 01:24:21
uh, it is relegated to, um, something that we don't have to think about, uh, in the long term. [chuckles]
Angie 01:24:29
Yeah, I totally agree with that. Um, I think we'll start looking at the source code as something like
Angie 01:24:39
kind of the byte code, [chuckles] the, the zeros and ones that this stuff all boils down to. It'll be like that low-level thing where, like, yeah, if you, if you wanna geek out and look at it, sure, knock yourself out, but I don't think there'll be a requirement or a need to. Um, like Theo said, I think it'll be much more about the intent. Um, we'll likely need
Angie 01:25:05
better ways to understand, um, like, what's-
Angie 01:25:10
... what this code base does, um, and the intent of it, you know. Uh, right now anybody can ask their agent that, but that doesn't seem scalable. Um, so I, I wouldn't be surprised if we saw something surface there.
Greg 01:25:26
And John, you have anything to add?
John 01:25:28
Sure. Uh, back to the earlier conversation, if you are comfortable with your agent building a compiler that you're gonna use, would you be comfortable having your agent build a Git replacement that you're gonna use or CI replacements? Um, at what point do you just trust your agent to take over the entire pipeline so that the things you have locally goes directly to your customers and that your customers and their intents align with the, the intents that you just have publicly served through...
John 01:25:57
I don't know, I, I guess I will wait till Theo vibe codes a new HTTP pro- protocol. Uh, I don't know what's, what's next on your plate. Um-
Theo 01:26:05
I have been thinking way too much about protocols. Don't tempt me, please.
John 01:26:08
Yeah, I figured. I figured.
Theo 01:26:09
A lot of inventing.
John 01:26:12
Um... [laughs] Sorry to... Yeah.
Greg 01:26:13
Re- regarding that, also Jeremy asked a question, uh, which is, uh, uh, quite important. So do you think that skills, like, um, skills will matter, uh, uh, or engineering will matter in the right patterns, um, if the right patterns are built into future versions of, let's say, Claude Code or Cursor or s- or things like that?
Angie 01:26:34
I wanna weigh in here first. Um,
Angie 01:26:38
I,
Angie 01:26:40
I don't know how people are using skills, but that question, like, really gets under my skin a bit. Um, the whole purpose of skills is not to make up for what the model is not capable of. Uh, you can have a very capable model. I can have a Fable-level model. What skills provides is
Angie 01:27:05
my specific
Angie 01:27:08
context, like my workflow of how my team works. So, you have me, you have Kent, you have John, you have Theo. All of us, you know, we work on different teams. We have different workflows of how we want stuff done. None of that is baked into a model, right? The model is generic, um, and so it has general knowledge that it is trained up on. Very, very capable, very competent. Um, skills provides not, like, how to do something, but how I want something done based on, like, my circumstance, right? And so I don't think, like, we get rid of skills when the models are more capable. In fact, like, if you look at some of the harnesses, like, I, I look at, uh, like, ChatGPT work, for example. I think it's beautifully engineered. Um, if you look beyond the prompt and see, like, what it's doing, like, it has a host of skills beyond my user-generated skills, a host of skills that tell it, like, how to, you know, interact with different things or how to use certain tools better based on, like, the harness itself. Um, and that does not exist in other platforms, right? Even though they might have the same connections to these tools, the authors of that harness have a very specific way that they think that this should work, and so they bake that into the skills.
Greg 01:28:37
And what about, maybe we can, uh, put an angle, maybe, maybe Kent or John or, or Theo can also put an angle of human skills o- o- on that instead of agent skills, 'cause Michael says that the question was maybe leaning a bit more towards human skills and how are we, you know, unnecessary or not in the loop.
John 01:28:57
I'll, I'll step up. So to Angie's point there,
John 01:29:02
if you pay attention to the community around trying to define terms for agents, terms like, uh, harness and sub-agent and everything, everyone has a different definition. And your engineering skills will be...
John 01:29:19
Uh, the, the better you can articulate what those concepts are and fully understand the pieces and how they fit together, the better you'll be able to ask and communicate with the agents the things you need. It's fine to be able to ask an agent, "Just make this website," or, "Make this game," and have it be really fun for, with the specific demographic. The more you understand the demographic, the more you understand the technology that, uh, that powers the tools that you're building, the more you understand, the, the better your product is gonna be for, for your, for your users, for your audience. And it's gonna require you to just be an expert in a lot of those, a lot of that terminology, that if you're not asking for the right thing, it's not gonna build the right thing. Um, it'll build a cool demo that you can post on your timeline and get a lot of likes,
John 01:30:10
but it's not gonna build a thing that, uh, withstands the pressure of security, of... Well, that's another discussion. Of multiple agents interacting with it and on and on.
Theo 01:30:25
Yeah, on that note, one of the most underrated things, and I pushed this a bit in my section, is the idea of, like, creating a shared grammar and language with your agents for a given project. 'Cause, like, words mean things, but they don't mean them well enough that the agent always means them the same way. So if you can, like, define a, a language that you and your agents and your work understand, it makes it so much easier to, to build the shared knowledge. And a lot of our job is going to be figuring out how much to understand and not to understand, and being more selective around what you let enter your brain. Because if you're not, you'll just always be overloaded, and I think we're about to enter an era of constant overwhelm and constant barraging of useless information, and we have to be very careful about what we do and don't let occupy space in our brains and time in our lives.
Kent 01:31:15
It's almost like attention is all you need. [laughing] I, I, um, I actually have been thinking about this a lot this morning, uh, thinking like, what if... Right, right now, um, the, the things that I talk about in my week, like one of them is, uh, about primitives and how you, uh, create those kinds of terms and the things that you talk about and the boxes that you draw around these abstractions and, and their relationships. And, uh, right now, setting that up, um, I think requires a, a certain amount of technical ability and, um, understandings of, uh, designing a good system and all of that. Um, I do think that eventually the models are just... and the agents, the harnesses that harness them, um, are gonna just get really, really, really good at all of that. And so what's left for us? Uh, and so
Kent 01:32:05
I think that, um,
Kent 01:32:07
communicating intent is valuable today and designing systems is valuable today. I'm not confident in my ability to predict what it's going to look like a year from now because I, I got it way wrong a year ago. Like, I had no idea how good we were gonna be a year ago. And so, um, what I think is a really durable skill is the, um, the ability to adapt to what becomes possible. Uh, this is a concept called adjacent possible. I actually, I've been thinking about this a lot this morning 'cause I'm gonna record a video about this exact thing, uh, as soon as we hang up here. But, um, this is just, uh, this idea of being aware of the advancements of technology and what becomes possible as the problems that people feel in the real world meet the technology that can solve those problems better than anything before. I think that will be, uh, a huge wedge, uh, a, a huge, like, moat, if you wanna call that that. Um, but, uh,
Kent 01:33:04
um, it's an exciting time. I, I think that we're gonna be able to solve a lot of really cool problems.
10. Mega format and closing thoughts
Greg 01:33:11
Uh, there's also quite a few questions about Mega itself, so, uh, I allow myself to, to answer also. Um, a lot of answers you'll find on the website, uh, but basically, yes, uh, this is async. The lessons are async. So you will just, uh, go to the, um, Mega HQ every day, and you have a recording, like 20, 30 minutes of recording. It will be much more structured than the conversation that we have here, of course. And then the lesson content, which is about half an hour reading. So you can do it async, you can do during the week, you can do during the weekend, and, uh, then you have challenges that you spend some time on. And if you want to get involved in the storyline, also you can spend some time on that. And then on top of that, you have, uh, each week, um, you will have a, uh, live session with one of the authors. So you will be able to ask all of those great questions. To be honest, I would love to answer all of them, uh, because there are really good questions, uh, on the chat right now. But I'll allow myself to basically copy, uh, those questions and, uh, and pass them to, uh, to the live sessions and the Discord so, uh, that we make sure that we, uh, try to answer them, because we are already 10 minutes past, past time. So maybe there is, um, some closing thoughts that you guys want to share, um, about Mega and your weeks, because basically we wanted to make it about, about your weeks and to maybe you, uh, to introduce yourself, uh, to, uh, to Mega students. Uh, so if you have anything else to add, um, please, uh, do now, and I will, um, only quote, um, Jonas or Jonas here. Um, "I'm very happy with this intro, setting the tone for what we'll go through the coming weeks. Uh, great to see all instructors and excited." Thanks so much, uh, Jonas, and thank you all for, uh, coming and being here with us, meeting one another and the instructors as well. As I mentioned, we also invited, uh, people from outside that were hesitant about joining Mega. Maybe it, it kind of gave you the, uh, the, the vibe of, um, of the weeks. Um, with, uh, all of, all of you who are already inside, we will meet on Discord that we are already there. Any final thoughts from you, Angie, Kent, John, uh, Joel, Theo?
John 01:35:31
Buckle up, I guess.
Greg 01:35:33
Very excited. Yeah, it's gonna be a ride, for sure. Um, there is a lot of content, uh, waiting for you already. The first lesson is up and the challenge as well. Uh, and I've-- and I saw that you guys already started to, to... There's already two and a half thousand people on the platform. So, uh, yeah. Joel?
Theo 01:35:49
Come in with an open mind and an empty context window. The goal is to help you reset and rethink where things are at. The less you apply the way things, the way you do things today and the more you open up to new ways of doing things and thinking about things, the, the more useful the material will be.
Joel 01:36:07
Yeah. Burning tokens is easy. Burning them well is a challenge. And I think when you come out of this, you are going to, uh, be way more equipped to, like, spend your tokens for building things well. Um, and that's a, a great skill that you'll be able to carry forward through, um, the next decade at least.
Greg 01:36:23
Yeah. Thank you all. Thank you, guys. Thanks so much for being here today and enjoy Mega.
Joel 01:36:28
Cheers.
Theo 01:36:28
See y'all.
Angie 01:36:30
Bye.
Angie 01:36:35
[outro music]