YT YouTube2Text

YouTube2Text → Transcripts → Agentic Engineering, explained by a 10x developer

Agentic Engineering, explained by a 10x developer — Transcript

by David Ondrej · 9,180 words · 1,269 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00I think the models are dead. Those
0:01contribution graphs, they're not worth
0:03anything anymore. Code is going to get
0:05so cheap. You tell me somebody reports a
0:07bug and then you park it somewhere until
0:10somebody thinks they have time to look
0:12at it and then what? Then they spawn an
0:15agent. The agents are already running
0:17the tests for you.
0:18>> You can get a lot of speed by leaning
0:19into this. People say, "Oh, Amp is taste
0:22like they say this is polished." But the
0:24fact is that 99% is written by AI.
0:28>> This is Furstenball, a founding engineer
0:30at Amp and the author of the famous Go
0:33books. In this podcast, we discuss why
0:35everything needs a CLI, whether computer
0:37science fundamentals still matter, and
0:40why 99% of slope comes from humans and
0:43not AI. If you want to be on the
0:45frontier of AI coding, watch until the
0:47end. This is the David Andre podcast.
0:49Enjoy. You know, the people who are
0:51still in the 2024-2025
0:53era of AI, how do they embrace the new
0:56way?
0:56>> So, one thing we've been doing from the
0:57start with Amp is we cannot stand still.
1:00The last 1 and 1/2 years has proven us
1:03correct. We could have made a lot more
1:06money and gotten a lot more users if we
1:09had stuck to a way of working that was
1:11maybe prevalent in early 2025. Oh, you
1:15have a single agent in a sidebar in VS
1:17Code and enterprise customers, they want
1:19like a proper permission system and you
1:22only run these commands, you know, like
1:24all of that stuff. And I want
1:25attribution. I want every line of code
1:28to be attributable to one instance. It's
1:30really important and if we had built
1:32this, I think we would have gotten one
1:34year of enterprise contracts, made a
1:36bunch of money with it, but then after a
1:38year, the competition would have been
1:39somewhere else. So, with Amp from the
1:41start, we said we have to optimize the
1:44company, the team, the codebase, and the
1:46product to be able to move fast, to shed
1:49weight when we think the frontier has
1:52moved on again. So, if you look at our,
1:54you know, ampcode.com/news,
1:57the meme is we constantly killing
1:58features because we think they're no
2:00longer necessary. And we've been doing
2:02this a lot. Like the ultimate thing
2:04amcode.com killed, you know, deleted
2:06itself, whatever. But they're right. And
2:09to give you a concrete example, first we
2:11had a VS Code extension and a twee,
2:13right? That was par for the course last
2:14year. Remember? Like we all have VS Code
2:16extensions. And then we said, "No." Like
2:19the app the VS Code extension, that's
2:20dead. Like it doesn't make sense to have
2:22this next to the editor. Who has the
2:23editor open anymore? Like yes, the twee
2:26is not for everybody, but like let's not
2:27get stuck to the editor. But if we would
2:30have stayed, we would have kept around
2:32users who really like to stay in their
2:34comfort zone. And then the question is,
2:37"Do you build up an echo chamber for
2:38yourself?" So we said, "Let's going to
2:39kill it." But um I think most of the
2:41time they appreciate us pushing. And we
2:44have a lot of customers that actually
2:46like to use Amp because they appreciate
2:48us doing this. They they ask us. They're
2:50like, "Guys, what's the matter? Like
2:52what do you guys use? What's the model?
2:53How do you prompt, you know?"
2:55>> Because it's the research, right? Like
2:56you're doing the research, not just the
2:57software.
2:58>> Exactly. Exactly. Like we the Amp
3:00Frontier Corporation, it's the official
3:02company name. And one of our things is
3:04we got to stay on the frontier. And then
3:07people pay us money so that they don't
3:09have to make the choice what goes into
3:11high or ultra or medium or what thinking
3:14levels. It's a waste of time for most
3:16people to like go around like, "Oh, I'm
3:19going to use GLM-5, too." Like you
3:21should think about what do you want to
3:22do? Where does the information live? How
3:26agent-friendly is your code base?
3:28And then concentrate on getting stuff
3:30done, but don't get bogged down in the
3:31details.
3:32>> Yeah, I have a lot of notes about
3:34everything you just said. I guess let's
3:35start from the backlog. In the, you
3:37know, olden days everything would go to
3:38a backlog, but now you can just like
3:40launch a new agent, new session, and it
3:42can be fixed, you know, in 10 minutes.
3:43So instead of like you stashing it
3:45somewhere for somebody else or for
3:46yourself, and then you have to remember
3:48it a couple of weeks, you can just like
3:49do it when you have that context, when
3:51you have that energy. The agent works on
3:52it and and it's done. That's I guess one
3:54of the things, but I think the bigger
3:56threat is like each generation of
3:57models, so much changes that we need to
3:59figure out like how do we work, you
4:01know, what is the workflows, what what
4:02is the user interface look like.
4:04>> I 100% agree with you.
4:05>> It feels like I paid you to say this.
4:07>> We now have powerful models like Fable 5
4:09or GPT 5.6 all. And when you plug these
4:12models into agent harnesses like Claude
4:14Code, Codex, or Cursor, they can build
4:17full stack apps with no issues. But the
4:19moment these agents touch your database,
4:21that's where things get dangerous. They
4:23often skip RLS policies, they
4:25hallucinate CLI commands that don't
4:27exist, and they rely on training data
4:29that's months or years out of date. This
4:31is why Supabase released Agent Skills.
4:34This is a free open source skill that
4:36makes any agent good at using Supabase.
4:39One command installs it and it works
4:40right away with Claude Code, Codex,
4:42Cursor, Pi, and basically any other
4:44agent out there. So instead of guessing,
4:46your agent now checks the official
4:48Supabase docs before doing anything. And
4:50it follows a built-in security
4:52checklist. RLS on every exposed schema,
4:54security invoker on views, and other
4:56best practices just like a professional
4:59back-end engineer would. Supabase is
5:01betting on a world where agents write
5:03all of your code. And this is them
5:05setting the foundations for it. If your
5:07agents are touching the database, make
5:09sure to install this Supabase skill
5:11first. The link is going to be in the
5:12description and it's completely free to
5:14get started. Thank you, Supabase, for
5:16sponsoring this video.
5:17>> I said last week, I think the models are
5:19dead in the sense that Fable is pretty
5:22good. GPT 5.6 is pretty good. But we're
5:24now at the level where I don't have to
5:26like watch a model and then like, "Oh,
5:29don't change this file. Oh, don't
5:31change" Like I think there's diminishing
5:33returns now to like this close-up
5:35managing of what a single model does. I
5:38think you could trust them a lot more
5:39than people do. And I think if people
5:41have like issues with "Oh, it doesn't
5:43use camel case" or whatever it is, I
5:44think they need to take a step back and
5:46rethink engineering because the way I
5:49phrased it for a couple weeks now is you
5:51will get a button with which you can
5:53spawn a John Carmack that writes the
5:56code for you. The John Carmack it
5:58doesn't speak unless spoken to, so you
6:00need to wake it up or have another John
6:02Carmack wake it up, right? That's a
6:03peculiar thing. Like he doesn't walk
6:05around discovering new things, but it
6:07won't make mistakes of syntax. It won't
6:09make expression mistakes. Like it will
6:11do the thing that you ask it to do. But
6:13then you got to ask yourself what do I
6:16want to do? And a lot of software
6:18development was consumed with how to do
6:22things. Like how do I write this? What's
6:24the best way to structure this? What's
6:25the best way to write this? How do I
6:27approach this? How do I approach this?
6:28Like a lot of stuff in software
6:30development followed from this and I
6:31think a lot of this stuff is not useful
6:33anymore. And I think even stuff like I
6:36have a custom slash command. This is the
6:37stuff that I invoke when I review or
6:39this is the flow that I do when like the
6:41models are pretty good, man. Like if you
6:43ask Fable to write
6:45>> Yeah. Like yeah, spawn a different agent
6:46and say review it, you know, whatever,
6:48but man
6:50>> diminishing returns. Diminishing
6:52returns. The most important, you know,
6:54number in this equation of how good are
6:56the results is the information that you
6:58put in. It's not about like details.
7:00It's like, okay, is this a gnarly
7:02problem? Like do you want like the John
7:04Carmack or do you want John Carmack's
7:05little brother, you know? It's insane to
7:07me that people take coding agents and
7:09then bolt them onto this old way of
7:12working where you tell me somebody
7:13reports a bug and then you park it
7:16somewhere until somebody thinks they
7:18have time to look at it and then what?
7:21Then they spawn an agent? Why don't you
7:23optimistically spawn these agents and
7:25have them parked somewhere and then go
7:27through them and look at the bug fixes
7:28and see if they're good or not. You
7:30don't need to estimate whether this is
7:32worth doing if it's all being done in
7:34the background while you sleep. Like
7:35let's just do it. A lot of stuff has
7:37changed. Old assumptions are no longer
7:39true. The agent is already a-sync. It's
7:42already running the tests. It's already
7:44in a sandbox with like isolated state,
7:47right? It's not like there's some stuff
7:49lying around. It's pretty good at doing
7:51this. So, if you ask it to run all of
7:53the tests and it runs them, you're
7:55telling me then going to push it
7:56somewhere else and then have it run
7:58again for 10 minutes because you have a
8:00CI system somewhere else that does
8:02basically the same thing. Like, the
8:03agents are already running the tests for
8:05you. Like, you got to rethink a lot of
8:08stuff from first principles, is what I'm
8:10saying. You probably should just pick a
8:12good model and focus on getting stuff
8:14done versus tweaking around other
8:16models, you know? If obviously if you do
8:17it in your free time, whatever, for fun,
8:19go go knock yourself out. But, um I
8:21don't think that's where the performance
8:22lies, you know? I don't think that's
8:24where the good results come from.
8:25>> Yeah, I mean, ultimately it's like
8:27knowing what to build and, you know,
8:28then building it. Yeah. I think like I
8:30have this controversial opinion, and I
8:31think it's becoming less controversial,
8:33that like, you know, people say AI slop,
8:35but I think most of slop comes from
8:37humans not having good product sense.
8:39You know.
8:40>> Yeah. They just with AI they can just
8:41build these trash products faster.
8:43>> Yeah. Maybe it's bragging, but look,
8:45people say oh Amp is taste like they say
8:47this is polished. Whatever, they like
8:49the key bindings, they like a bunch of
8:51stuff. Oh, they say you guys have good
8:53taste. You polish, you craft, whatever
8:55it is, right? But, the fact is that from
8:58what you see here, 99% is written by AI.
9:02Like, nobody in our team writes code by
9:05hand anymore that much. I I did a poll
9:07internally last week or 2 weeks ago and
9:09I said, "How much code is written, you
9:11know, how what do you guys think?" And I
9:13gave three options and one option was 99
9:15plus, the other was between 90 and 99,
9:19and then I was like below 90. I didn't
9:21even anticipate that somebody would go
9:22lower than 80%. And then Camden, who's
9:25in our team and who's great, he's like,
9:27"I think I still write a bunch of code
9:29by hand." And then I'm like, "Camden,
9:31really? Like, how much code do you write
9:32by hand?" And he goes, "Ah, you're
9:34right. It's probably 95%, you know?" So,
9:38I I agree with you. I think I put a lot
9:40of care into the stuff that I build, but
9:42it's it's the ideas, you know, it's it's
9:44like do I think this little icon is
9:47misplaced? Do I think this icon should
9:49have a tool tip? Do I think we should
9:50have a cool animation? That stuff here
9:53where you can, you know, you spawn amp
9:55and it shows you this animated orb, you
9:57know, like what AI helped me to do here
10:00was it generated 15 different versions
10:04of these orbs in different colors that I
10:06can look through.
10:07>> Right? It made this faster for me. I
10:09said like give me a version with like
10:11braille characters. Give me a version
10:12with these characters. Give me five
10:14different versions. Give me 18 different
10:16colors. And then I went through them and
10:17picked the one that I like best. And it
10:19seems to resonate with people. So, I
10:21used AI to express what I think is good
10:24faster, you know, without doing a bunch
10:26of work. If I had to write this all by
10:28myself, never would have done it. But
10:30now I did and people love it and they
10:31think it's it's I don't want to say it's
10:34human, but it resonates with people. And
10:36they think it's high taste. And to your
10:38point, it's not slop. It's not it's not
10:40the enshittification of everything. I
10:43use AI or we use AI to accelerate it,
10:46right? People like our news posts and
10:48say we have like good background images
10:50in the news posts and what not, right?
10:52That's me and Tim and Brad going turn by
10:54turn by turn in midjourney. That's me
10:57reading Moby Dick and thinking about
10:59orbs and then trying to combine in my
11:00head Moby Dick and orb. What else could
11:02we do? Let's mix this up. Let's ask
11:04ChatGPT for 10 different variations of
11:06this midjourney prompt. Let's pick this
11:07one. Let's remix this one, you know? AI
11:09helps me to go fast, but I still put the
11:11effort in and I don't think it's slop.
11:13So, yeah, I agree with you. If it's
11:15slop, it's because there's a lack of
11:17ideas and there's a lack of input.
11:19There's a lack of knowing what you want
11:21to exist in the world or a lack of
11:24playfulness, whatever you want to call
11:26it. Creativity maybe? I agree I I agree.
11:29So, for people like who want to start a
11:31new project, you know, let's say they
11:32want to build an AI startup this year,
11:34given the state of AI agents, how how
11:36would you advise them to do that?
11:38>> Start thinking about what you want the
11:40business to be. Like, what's your
11:42market? Who's your customer? How do you
11:44serve them? How do you generate value?
11:46How do you sell? Let's turn it around.
11:48Let's say somebody's really interested
11:51in hiring sales people and wants to
11:52create an AI startup. You know, how how
11:55should they start? You know, it's the
11:56same thing like why would you hire sales
11:58people if you don't know what the
12:00company is. It's the same thing for
12:02software now where software as the thing
12:05itself, like that's become less
12:07valuable. You know, we said this last
12:09year in our podcast, we said like open
12:11source is going to take a hit. Like
12:12those contribution graphs, then they're
12:14not worth anything anymore. You know,
12:16like code is going to get so cheap. Now
12:18it's it's much more about what is the
12:21software doing, how much value is it
12:23providing. And that sounds boring as
12:26hell, but that's just the truth of it.
12:27And the interesting thing is that
12:29actually in the last 20, 30 years, most
12:32programmers were in that lucky situation
12:34where the thing that they really care
12:35about and, you know, treat as their
12:37craft or as their hobby overlapped with
12:40what's valuable for the business. You
12:42know, like the guy who really likes
12:44writing Haskell on the weekend, turns
12:46out he's also the guy who can build like
12:48a good finance back-end because he
12:50really cares about types and whatnot,
12:51you know, and is really good at modeling
12:53this stuff.
12:54But now, Fable is also pretty good at
12:56that stuff. And people say, "Well, I got
12:59domain experience." And then it's like,
13:01man,
13:02I've seen these models. This was GPT-55,
13:04you know, where you ask it to write a
13:06database migration and it just says
13:09like, "Oh, wait, we cannot drop the
13:10column yet. We got to add the new
13:11column. We got to deploy this. We got to
13:13write to the new column. Wait until all
13:15the other code is gone. Only then do we
13:16drop the old column." Like that's stuff
13:18that a senior engineer taught me in like
13:19my first two or three years as an
13:21engineer. And I think that was a
13:23hard-won lesson that nobody really
13:25spelled out to me like that. And it
13:26seems trivial, but man, it's something
13:28the model does in like half a minute.
13:31And like a lot of stuff is just faster
13:33and and easy to get done and thus
13:35cheaper. So, the question is more what
13:38do you want to get done? Does that
13:39answer your question? Is that Is this Is
13:41this Am I cheating by saying
13:44>> [laughter]
13:45>> No, no, no. Like that I think that
13:46answers like the product side of it. So,
13:48let me ask the second side. It's like do
13:50you build for the human or do you build
13:52for the agent?
13:52>> Oh, it's
13:53>> Because the orbs, for example, orbs you
13:55build for the agents, right? Even you
13:56say like the humans, you know, they
13:57might want a local blah blah blah. It's
13:59like no, the agent wants a full
14:00computer. He wants to do anything on
14:02that computer. How do you think about
14:03product like that?
14:04>> I think we're all finding out. I think
14:06it's it's kind of changing, right? I
14:08think all of software is trying to find
14:10out what is the future of software. Like
14:12there's been a few ideas like everything
14:14will be an API or everything will be a
14:16CLI and then the agents will do your
14:18shopping and whatnot. And maybe that's
14:21what happens. I do think like a lot of
14:24old software will die out. The question
14:27is how fast, you know, like can you get
14:29like the 70-year-old people to use an
14:30agent to book a flight for them? I don't
14:32know. But like what I think like if I
14:35had to start and build for the frontier
14:37a new product, right? I would make it so
14:39that no human has to fill out forms
14:42again. I think that's dumb. Like that
14:44stuff is dumb. Like everything that a
14:46human can do, they should be able to ask
14:48an agent to do on the website, right?
14:50That would be amazing. The APIs are
14:52maybe not there yet, but
14:54it would be good to be able to bring
14:55your own agent. Like most people don't
14:58want to use your shitty built-in agent
15:00maybe, you know, like it would be nice
15:02if I can use Amp to do my shopping for
15:04me on another side or, you know,
15:06whatever. I don't know if that's a pipe
15:07dream, but that's how I think about it.
15:09I think a lot of stuff is stuck in the
15:11old assumption. To give you give you a
15:13funny example, on Sunday, it's Tuesday
15:16now, on Sunday I'm here in a local club
15:18and we had like a little festival and I
15:19built like for fun in an orb from my
15:22phone, I built like a little uh shopping
15:24thing where people can order food and
15:25whatnot and then the kitchen gets like
15:27the orders and they can check them off
15:29and whatnot because that's what we did
15:30with paper on Sunday. So, somebody was
15:33like, "Can't you make an app for this?"
15:34I'm like, "Actually, I can." So, I built
15:36the app, right? I took a photo of like
15:38the menu front and back, took two
15:40photos, sent it to the thing, and it
15:42built it, works, did it in the orb. I
15:45could share the URL, people could go to
15:46the URL. It worked. It was crazy. The
15:49interesting and to me, maybe only to me,
15:51you tell me, but the philosophically
15:53interesting thing to me was it built an
15:55admin interface where somebody could now
15:59go and look through all of the products
16:01that are on the menu and adjust the
16:02prices manually, right? Or adjust how
16:04they're spelled and stuff. And I looked
16:06at this and I'm like,
16:08"I'm never going to do that, dude. Like,
16:10I'm not going to go in there and use the
16:12form. What are you talking about? I'm
16:14going to use you, Amp, in that code base
16:16again. I'm going to send you another
16:18photo if the prices change and I'm going
16:20to say, 'Fix the pricing, you know?'"
16:22Like, a lot of
16:24setting stuff, a lot of admin stuff, a
16:26lot of stuff that a human did so that no
16:29code had to be changed is now probably
16:31easier by changing the code, you know?
16:33Imagine a WordPress back end, right?
16:36Where you you build a WordPress back end
16:38so that everybody in your company,
16:39marketing and whatnot, could add a blog
16:41post and and stuff like this and have a
16:43spell checker and stuff. If you had an
16:45agent in your code base, you could say,
16:47"Here's the draft for my blog post.
16:49Please add this header image and publish
16:51this blog post and do a spell check."
16:53That's one sentence and it goes and does
16:55it and you don't have to click around
16:56with a bunch of stuff. So, if you start
16:58a new company and you start a new code
17:00base, you have to think about what is
17:02the role of code in your product. Don't
17:04fall in the trap of thinking the old way
17:07of, you know, shipping software is still
17:09valuable because it might not be like it
17:11the the calculus has changed, you know?
17:13You can do a lot of stuff, even stuff
17:15that seems unfeasible or seemed
17:17unfeasible, like having five different
17:19copies of stuff, having five different
17:20versions being built, have you know, all
17:22of that stuff I think you need to
17:24rethink. And I think the whole, you
17:26know, software industry is is is
17:28figuring this out at the same time. It's
17:30very interesting. I think, you know, we
17:32had an internal document, I think it was
17:34called software's dead, now what? You
17:37know, because it was about this very
17:38same thing where, look man, like the
17:41models are getting to this point where
17:43you can wish things into existence as
17:45long as you have the wishes, as long as
17:47you have a use case for what you want to
17:49build. And now, we're thinking the
17:51future of software might be that, well,
17:53you don't want to rebuild everything all
17:55the time, but there's this element of, I
17:58want to take bits and pieces and then
18:00remix them, and that's easy to do.
18:03Somebody wrote a blog post, I wish I
18:04could remember the name, they were
18:06saying the Emacsification of software.
18:09So, in Emacs, which is this text editor
18:12and you configure it by writing Elisp,
18:14there's this thing where everybody has
18:16their own config. Like they they have
18:18like their dot files and everybody has a
18:20collection of Lisp files and they
18:21collected it over many years. And then,
18:23say if somebody has a fuzzy finder
18:25plugin, what's common in the Emacs
18:27community is they would fork it and
18:29change it and like rewrite the code and
18:32then not contribute it back because they
18:34customized it so much. This is exactly
18:36what's happening right now, too. Like
18:38there's this great diff viewer called
18:40hunk, and I wish I could remember the
18:42author, it's a nice piece of software.
18:44It's to view diffs and I I actually
18:46stopped using it, but when I used it, I
18:49forked it, pointed amp at it, and said,
18:52"Please add the gruvbox dark hard theme.
18:55Please add the ability to check off
18:57files in the sidebar. Please do this,
18:59then compile it and put it in my bin
19:01folder in my home directory." And it did
19:03it, and I had custom software, and there
19:05was no reason for me to contribute back
19:07upstream. Like I didn't want to
19:09generalize it. I I think it might be
19:10valuable to others. I didn't spend any
19:12time on it really. Like it was just 2
19:14minutes of the agent running in the
19:16background.
19:17And like what happened there? Like I
19:19took a piece of software and remixed it
19:21for my own use case and now I use it,
19:23you know? And I think we might see a lot
19:26more of this. Like obviously on the
19:28other end of the spectrum there's the
19:29skepticism that, you know, like
19:31non-technical people are not going to
19:32vibe code whatever their apps and stuff
19:35like this. But I do think there's going
19:36to be a lot of software. I do think
19:38there's probably going to be a lot of
19:40internal software and I think for these
19:42use cases you had internal software
19:45being developed before. But now, the
19:47example from Sunday, right? It's a
19:49little club, it's 20 people. We had this
19:51process of you give up your order here,
19:53somebody writes it down on a piece of
19:54paper, then it goes over here. I could
19:56encode this in software, I don't know,
19:59in three times 5 minutes over the course
20:01of what was it? 2 hours? Like I pulled
20:03out my phone, I wrote my prompt, and
20:05then said like now we need another
20:06thing.
20:08And then I showed it around like that's
20:09crazy. Like if people if people start to
20:11realize that they can do this in
20:13companies, it will replace like what
20:15previously was maybe like that one Excel
20:17file or that one wiki page or that one,
20:21you know, hacked together whatever
20:23thing. Like I think a lot of that's
20:25going to happen. And and maybe that
20:27blast radius of this movement is going
20:29to get bigger so that people will attack
20:31bigger things. That they're going to say
20:33well, we want to buy this software,
20:35Riverside. I don't know I'm picking
20:37Riverside. We're going to pick
20:38Riverside, but actually I don't care
20:39about audio only. I only care about
20:42videos. Whatever it is. Or the other way
20:44around, right? I only remove everything
20:47in the app. I don't want to see this. I
20:49don't want to deal with it. Give me a
20:50custom version of this. And I think, you
20:53know, the agents will be able to do this
20:55very soon. And and the question is what
20:57happens to software then, you know?
20:59>> Yeah, I mean there's going to be
21:00explosion of it. It's going to be a game
21:02of like who has who knows what to build
21:04and who has the right setup. I think if
21:06you boil it down it's two variables.
21:07Knowing how to use these agents and then
21:09just having the budget to do it.
21:10>> Yeah.
21:10>> If you can afford to spend the tokens
21:12>> Yeah.
21:12>> and you can use
21:13>> Yeah.
21:13>> the frontier models, I I think you're
21:14you're going to crush in the future.
21:15>> I
21:16>> You cannot take a programmer who doesn't
21:18use AI. He's going to get crushed by a
21:20mediocre programmer with AI. Unless you
21:21say like, "Okay, it's something like
21:23super technical and you know, there's
21:24like only a few people with that
21:25knowledge." But like in general, you
21:26know, if somebody's like a better
21:28programmer, but you give this guy a
21:29ton of AI,
21:30you can extrapolate where the models
21:32will be in 6 to 12 months. It's just all
21:34about like knowing how to use these
21:35agents.
21:35>> Yeah, I agree. I think the most
21:37important skills will be like first
21:39principles thinking, like actual like
21:42thinking, not like technical detailed
21:44stuff and whatnot. Like you actually
21:46have to think through problem-solving
21:48thing of breaking down business problems
21:50into solutions. Again, dumb example
21:52Sunday. The first solution somebody
21:54says, like, "Hey Torsten, I heard you
21:55can build apps now with AI." And I'm
21:57like, "Yeah, I can." And they were
21:58saying, "Look, right now we have these
22:00orders in our kitchen and somebody makes
22:02a mark on a piece of paper and then
22:03hands a piece of paper over. Can't you
22:06build us an app so that somebody on a
22:08tablet hits this thing and then prints
22:10out the piece of paper and then hands it
22:11over?" And I, as somebody who's been in
22:13software for a while, I was like, "Okay,
22:16I can build this. I can get a printer.
22:17Like it's not, you know, we can do this.
22:20But why do you need a printer? Like why
22:22not have a second tablet instead of
22:24having a piece of paper? You know, and
22:26and this kind of thinking of what
22:28actually is the process, what is the
22:30workflow, what problems are being solved
22:33through this workflow, and how can we
22:34replace this with software? I think
22:35that's going to be incredibly important.
22:38>> So then it's more like everybody's
22:39architect and you just need like to be
22:41aware of like the different solutions
22:42from different industries, different
22:43businesses
22:44to like have the right idea, I guess.
22:46>> Yeah, I Look, like I think what I do is
22:50still engineering, right? When I'm now
22:52going to look at should we port it like
22:53this to the TUI or to the CLI or
22:55something, right? But the mistake that I
22:58see people make online is they compare
23:01They always use like the 1% as an
23:04example. They say, "Look, a Mitchell
23:07Hashimoto reads the code and, you know,
23:09like AI is not good enough to solve this
23:12issue for him." And Mitchell is one of
23:13the best programmers in the world. He's
23:16working on like beautifully graph, like
23:18it's incredible software. It's it's
23:21technically challenging. It's really,
23:22really good. But look around what most
23:25of the software developers around the
23:27world that will look at most of the
23:28software around the world. It is not
23:31building a GPU accelerated terminal
23:33emulator, you know? I think people they
23:36underestimate like how much software
23:37there is where you can solve problems
23:40without being overly technical. Where
23:43it's okay to just use whatever MySQL or
23:46something something and have a crowd app
23:48and whatnot. That's good enough, you
23:49know?
23:50>> All right, first of so, can you show us
23:52your agentic engineering setup?
23:53>> Yes. So, this is Amp, ampcode.com. It's
23:56installed as a PWA, which
24:00>> Okay.
24:01>> See, I think yeah, I think that's
24:03actually what stops the adoption of the
24:05technology. It's
24:07It's not easy to pronounce, a PWA. But,
24:09um yeah, ampcode.com. And Amp is a
24:12coding agent. We've launched it last
24:14May. One of the first ones, right? After
24:16Aider, then Cloud Coden, and then right
24:18after Cloud Coden, we came out with
24:20multimodal. So, we support GPT models,
24:22we support Anthropic models, we support
24:24GLM. Since 2 weeks, we have the dial.
24:27So, here you can select like how how
24:29much effort could you want to put in?
24:30Low, GLM 52 with the Oracle, which is a
24:33sub-agent we have. I often go with
24:36medium actually, 56, solve medium, and
24:38Oracle is solve high. So, Oracle is a
24:40sub-agent that is a reviewer, you know,
24:43gives advice, I guess. Then high, GPT
24:4556, extra high, ultra, uh means, you
24:48know.
24:49>> So, from your experience, 5.6 is a
24:51better reviewer than Fable?
24:53>> Um it depends, right? Like here in high,
24:55we have Fable as the Oracle. Like in
24:58that case, you know, it reviews. In this
25:01case, we switched it. Honestly, like
25:03it's nuanced, right? Like between high
25:05and ultra, but in this case, medium is,
25:08you know, it says here balanced
25:09intelligence speed and cost for most
25:12tasks, right? So, I think in that case,
25:14it's fine. You don't need fable in the
25:16mix. We can always then ask for fable.
25:19So, I don't start a lot of agents in the
25:22twee anymore. We've had this feature
25:24orbs since 2 weeks, so I can, you know,
25:27give me an overview of this repo. So,
25:29this will spawn a sandbox remote
25:32machine, which is We call this an orb.
25:34It's tied to this conversation, which
25:37is, I think, a little bit of a game
25:39changer if we if you want to take this
25:40in a philosophical direction about, you
25:42know, the future of software. The start
25:44is This is now running an agent in a
25:47checkout of this repo in a remote
25:50sandbox. Once the agent goes idle, the
25:52sandbox goes to sleep. Once you start
25:54typing, it goes back up again. If I open
25:57this on my phone, basically it's
25:59streaming on my phone, too. You know,
26:01it's multi-device. I can even open this
26:04in the twee. Okay? And that is actually
26:07how I do most of my coding in the last
26:10few weeks.
26:11>> So, the way I understand it is like when
26:12you give it its own any system
26:14basically, like you know, some people
26:15deploy agents on VPS. When you do that,
26:17the agent can like take way more risk,
26:19right? Because it cannot like mess up
26:20your computer.
26:21>> yeah.
26:21>> So, is that the main benefit you see?
26:22>> If I start amp any client in amp, and
26:26then if I go here, you know, here this
26:28is just I started this just now. I said,
26:30"Hello." This is in my twee. It's on the
26:32second screen here. It's the amp twee. I
26:34can now remote control this and say, uh,
26:36you know,
26:38whatever. I can see it on the other
26:40screen, right? But now, I can also say I
26:42want to start a thread here in this twee
26:45instance. Like in this folder, my one
26:47amp one checkout on my MacBook Pro,
26:49right? And it starts a new thread and
26:51say, you know, "Hi cool let's use low
26:53for this. Don't waste any tokens and
26:55start it, right? And this is now
26:57creating a thread in my tool we on my
27:00machine. You can imagine you can start
27:02amp instances on your Raspberry Pi, on
27:04your dev cloud box, on your VPS and
27:06whatnot. The reason I'm mentioning this
27:08is that if you look at this, it's the
27:10same machine. So if these two threads,
27:13these two conversations make a change,
27:15we don't do any work tree or checkout
27:17management right now. So they will bump
27:19into each other this change. We could
27:21add work tree support, but I think
27:23philosophically now where we're going,
27:25we believe that these models are so good
27:28at being in these headless boxes that
27:31there's very little reason to stay local
27:34and mess around with like local setups.
27:36And it's really powerful to have a
27:39separate orb
27:40per change. This for example, I worked
27:42on this like two weeks ago, goes to
27:44multi-account switching feature. If I
27:46open this, you know, like this will come
27:48back up if I want to make a change and
27:50then I can see the diff, right? But it's
27:53now parked, that means the computation
27:55that made a change, the context that
27:56produced the change and the agent are
27:58one behind one URL. I can share this URL
28:01with my team and they can now access
28:03this. As of couple days ago, we also
28:06have multiplayer support which by the
28:07time this episode comes out, I think it
28:09will be released where you can now
28:11access these orbs where other people in
28:13your team are working on. The fact that
28:15you can package up the thread, the
28:17agent, the computation, the change in
28:19one URL and you don't have it on your
28:21machine, the thread somewhere else and
28:23the change pushed up to GitHub in a
28:24branch, I think it changes things a lot.
28:27Like we use it a lot and I think like 10
28:29years ago, there was this big push for
28:31cloud IDEs. I don't know how much, you
28:33know, Cloud 9 and whatnot and all of
28:35that stuff. And back then, cloud IDEs
28:37were seen as this big solution of, you
28:39know, you can have a Chromebook do all
28:41your dev on a bigger machine and blah
28:43blah blah all of that stuff. But then
28:45people bumped into these issues of,
28:47well, the latency or I don't have my key
28:50bindings or I cannot SSH into it. I
28:52cannot do this. I cannot do that. I
28:54cannot you know, like all of that stuff
28:56and I think if you look back at the last
28:596 months of how people write code or let
29:02code be written, right?
29:03>> Yeah.
29:04>> The arguments are gone, man. Like who
29:05cares about latency when you have to
29:07wait for tokens per second? Who cares
29:09about language servers anymore? I don't
29:11use language servers anymore. I don't
29:13use my editor anymore. I don't use crazy
29:15key bindings anymore. So a lot of those
29:17arguments don't hold and when you go
29:19remote, you now async by default and you
29:22can do a lot of other stuff. Like just
29:23to drive the point home, somebody on
29:25Twitter was like, "Hey, can the agents
29:26and orbs even use like a rust
29:28cross-platform desktop framework, you
29:31know?" And I ask Puck, which is a thing
29:34we released yesterday, which is like the
29:36meta agent. I said like, "Create a
29:38project where you can prove like
29:40somebody on Twitter asked this, create a
29:41project here, rust desktop testing, you
29:44know, pick a rust desktop framework, ask
29:45the agent to take a screenshot." So it
29:47did. We ran into a bug that I fixed in a
29:48separate thread, which we can look at in
29:50a second. It sent a message to this
29:52other agent. Like Puck is the meta
29:54agent. It's this It's this thing which
29:56you launched um yesterday and people
29:59love it, but this Puck can control your
30:02other agents. So it spawned an orb. It
30:05sent the message and it says,
30:07"Investigate whether an orb amp orb can
30:08support cross-platform desktop app
30:10development in this repo, right? Pick an
30:12appropriate blah blah blah." It's a
30:13prompt, okay? Starts working. I'm doing
30:15some other thing and then it yeah, it
30:17produces a screenshot. And I mean, look
30:19like it worked for 16 minutes. From a
30:22dev perspective, whether the latency is
30:25say 2 minutes or 16 minutes does not
30:27matter. You're already somewhere else.
30:29>> Cuz you delegate it.
30:30>> Yeah, exactly. And with 16 minutes, you
30:33know, they can do a lot. The point I'm
30:34making here is I think people haven't
30:37seen what these agents can do now. Like
30:40in these headless boxes. Like this is
30:42crazy. Yeah, so we have Puck and I have
30:45this website ampco.com/whatthepuck.
30:48In it, you know, we have like different
30:50instances of Puck, you know, different
30:52renders of it here. That that's what
30:54they look like. So,
30:55and okay, so this is a fun thing, right?
30:57Let's not pretend that this is serious
30:59stuff, right? But look, like I ask it
31:01create more variations of this, okay?
31:04And then in amp we have a sub agent
31:06called the painter which can create
31:08images. And what it did was if you look
31:11through it, let's see best version. I
31:12said like, "No, no, no, do this." And it
31:14uses the painter to create a new variant
31:17of the Puck. So, this is the lava lamp,
31:19this is the bubble wrap, blah, blah,
31:21blah, you get it.
31:21>> Yeah.
31:22>> And my point is if you had told me 6
31:25weeks ago you're creating images and
31:28like previewing them and reviewing them
31:30and then cutting them or slicing them
31:32together or putting them in homepage or
31:33removing the background, I would have
31:34said, "I kind of want my hands on." You
31:36know, like I might open Pixelmator or
31:39Figma or whatever it is. But man, look,
31:41like I didn't go, you know, it added
31:44them and then I said, "We need to update
31:46the copy on the page." And I said, "Show
31:48me a screenshot." It comes back and it
31:49updated the copy. Let's do it live. Like
31:52I think we can ship this boom to
31:54production, okay? So, now it's sending
31:56this prompt and it's going to ship it to
31:57production. And it showed me the
31:59screenshot and Quinn, CEO of amp, he had
32:02a good tweet I think yesterday or today,
32:04which is the same point that I just made
32:06of you're async anyway, so you can ask
32:09the agent to give you proof. It's not
32:11like you're waiting the next 30 seconds,
32:14you're waiting maybe a minute. So, you
32:15ask it and I we can look at some other
32:18things. I'm like, "No, test this again
32:20against the dev server. Like test it
32:21again."
32:22>> Yeah, launch 50 50 different tests to
32:24see if it works, you know, benchmark it.
32:26>> Exactly. Show me the screenshot. Show me
32:28the screenshot in dark mode and light
32:30mode, whatever, because it's going to
32:31take 10 minutes anyway and once I'm
32:32back, I'm sure it works and then I'm
32:33going to hit the ship it button. The
32:35whole game I think changed. Here, what's
32:38another example? I wanted to show you
32:39this. So, Puck is right now it's only
32:42this thing in the web, right? It's one
32:43conversation. So, then I ask it I think
32:45this was even Fable. Here, this is how I
32:47prompt, by the way. I don't use any
32:49custom slash commands. I don't use any
32:50skills, really. I don't have any MCP
32:53servers. The number one thing that's
32:55important for me when writing prompts is
32:58where does the information come from of
33:00that the agent has access to to
33:02interpret what I mean. So, that means an
33:04agent has two sources where it can get
33:06information from, right? Training data
33:08and the context window. That's it.
33:09>> Yeah.
33:10>> That's what boils down to. Training
33:11data, as we all know, is the model
33:15observable internet public knowledge,
33:17whatever, and it's kind of lossy in a
33:20sense, but the way I like to think about
33:23it is the model is a senior engineer
33:25who's seen it all, who's had work in all
33:29of the operating systems, back and front
33:30end, client, all the different
33:32languages. It knows how to invoke an
33:34anonymous function in scheme, in common
33:36lisp, in Lua, in JavaScript, whatever.
33:37It knows it all, but the mental model I
33:40have is imagine that senior engineer and
33:43somebody snatches you off the street,
33:45puts a hood over your head, drives you
33:47somewhere, and suddenly pulls the hood
33:48off your head, and in front of you
33:50there's only a desk with a computer on
33:52it. On the computer there's a code base,
33:55there's a browser, there's a terminal,
33:58and there's a text editor. That's what
34:00you have access to. And then, magically
34:02from somewhere, somebody hands you a
34:04prompt and says, "Implement this or fix
34:07this feature." That's roughly the
34:08information that the model has. It can
34:10look in the code base, it can look on
34:12the internet, it can use the terminal
34:15and run commands, and it's very
34:17experienced, but if your prompt does not
34:20contain enough information, or your code
34:22base doesn't contain enough information,
34:24or your agent MD doesn't contain
34:25information, it can't turn a shitty
34:27prompt into a good prompt. You know,
34:29like here for example, it's about
34:30porting Puck to our CLI. I could have
34:33written port puck to the CLI. That's I
34:35think that's a shitty prompt. I have
34:37more information in my head about how I
34:38think it should work. So, what I do is I
34:40just put that information in, right? I
34:42say, "Look at how it's implemented in
34:44the web UI." So, I set the standard. I
34:46pointed at the information. I say, "This
34:48is what you should take as the standard.
34:50I want to port this to our CLI." And
34:53then I say, "I think we should have a
34:55puck open command. We have a command
34:56palette in our CLI that opens the puck
34:59for it." And then, you know, that shows
35:00up in the sidebar. We have a sidebar in
35:02the TUI. And then I think we should have
35:03a puck clear thing, blah blah. So, then
35:05I'm kind of like riffing. Maybe we need
35:07a new modality for the TUI, right? This
35:09is how I would talk to a senior
35:10engineer, right? Like, if I were to talk
35:13to a fellow senior engineer and say,
35:14"Hey, could you port puck to the TUI?"
35:17This is what I would tell them, right?
35:18This is the slack message I would send.
35:20And then I say, "Obviously, this is not
35:22what I would tell to a human being
35:24without maybe some flowery words." But
35:26instead I say, "I What I want you to do
35:28is you should research how it's
35:30implemented. All right, research how we
35:32communicate, what it does. So, look at
35:34the news post, right? We released this
35:35yesterday. Document how it works in the
35:37doc." And I don't care really about the
35:39doc, but I think it's the same for
35:41humans and models. I want you to sit
35:43down and think about it. Like,
35:45compile what you learn. I don't know if
35:48this is made up stuff, but I do think it
35:51helps. And then I say, "Come up with a
35:53good idea for how to translate it. Then
35:54use sub-agents, you know, fabulous
35:56expensive. It's so expensive that it
35:58scares me. So, I'm like, use sub-agents
36:00for implementing this, which I think is
36:02GPT models right now. And then present
36:04me with the results." And then it went
36:05off and did it. This was an hour ago, I
36:07think, right? It's like, "Hey, I ported
36:09it to a design. Um I used this. It's
36:12bound to the blah blah blah blah blah.
36:14Puck So, kind of like what I said, and
36:16it appears in the sidebar, right? And
36:18then he has the reference. Okay, cool.
36:21Uh maybe we can reuse this. Storybook I
36:23asked it to create storybooks. We have
36:24storybooks for the CLI. And then I said,
36:27"Can you test it against the local dev
36:28server?" And then I did something else.
36:30And then I think it went on for
36:32I don't know how long, and it found a
36:34bunch of bugs, and I don't care. So, now
36:36I need to read through this, but it says
36:38like it did an end-to-end thing. Uh the
36:40code 1,200 or 300 lines verification,
36:43ran all of the tests. So, now, you know,
36:45either I go in, we have like access to
36:48the terminal in the orb. I could now
36:49test the CLI and see whether Here there
36:52it is. Puck shows up in the sidebar.
36:54This is, by the way, the app CLI. This
36:56is Ghost A, you know, web assembly Ghost
36:58A version, so a bunch of stuff works,
37:00even the mouse clicks work. But, uh even
37:02this works, and I think I have an Easter
37:05egg here, so there you go. If you click
37:06five times, let's see if that works.
37:08Boom, the orb explodes. So, yeah, here,
37:10Puck seems to work. And then I would I
37:12don't, you know, okay, but it seems to
37:14work, right? That's pretty cool. So,
37:16again, coming back to the question, you
37:18didn't ask question, I posted. How
37:19important is the local dev env? Like,
37:21look at what I'm doing, right? Like,
37:22this ran, I confirmed it works. I'm
37:24going to do spot checks of how it, you
37:27know, how how it implemented it, but at
37:30the end of the day, if it did end-to-end
37:32tests, and I read through it, and I kind
37:34of agree with the say architectural
37:36choices or bigger
37:38>> Software design choices.
37:39>> Yeah, software design choices. What do I
37:41need my local dev env? You know, like,
37:43this is it. Like, I I'm testing it, I
37:44have access to the terminal.
37:47It's pretty good, you know? And if here,
37:49like, this is So, I'm going to Laracon
37:51next week, and giving a talk. So, I was
37:54kind of playing around with uh Laravel,
37:56which I've never played around with.
37:57I've been a Rails programmer before,
38:00right? And I think it's inspired by it.
38:03So, here I had it build in an orb just
38:05to try it. It set up a fresh Laravel
38:07thing. And in orbs, we have portals,
38:09which, you know, you can spawn anything.
38:12something running on port 8000, and now
38:14I can access it here, and I say, "Hello,
38:18uh from our podcast." Yeah, whatever.
38:21And I just want to test it whether this
38:23works with like proper state. I think it
38:25had like some front-end tests, but um
38:27Oh, there it is. It did It did work. So,
38:29this is a full Laravel app running in
38:31this, and now I can even share this URL.
38:34I can open it in browser and share with
38:36somebody else, right? I can test it. I
38:38can look at the files in the orb. I can
38:40see all of the changes it made. Question
38:43is, what more do you need? And now I can
38:45basically say, "Okay, take this, share
38:47it with my teammates." I can turn
38:49multiplayer on, and then it would show
38:51up here, and then say a team they could
38:53take over, and say, "Yeah, actually this
38:55sign-up screen looks bad." And they can
38:57ship it, which I think is pretty neat.
39:00>> Yeah, especially when you have like less
39:01technical people on the team they can
39:02start something off, and then somebody
39:04can pick it up and do the last 10%
39:06that's really hard.
39:07>> Yeah. And I mean, you know this like
39:09preview URLs in PRs were a big thing.
39:12Like if you think back 5 years ago where
39:14people would, you know, "Oh, I want to
39:16change something in our blog." So, yeah,
39:18download GitHub Desktop because you
39:20don't know how to use GitHub on the CLI,
39:22then do this, create a pull request,
39:25then you get a Vercel or whatever
39:26preview URL, and that was cool. But
39:29obviously here now you have the context
39:31of the change, like what that person
39:33typed into the agent, you have the
39:35change itself, you have the computation,
39:37you can check whether it runs, you don't
39:39have to check it out locally, and then
39:41you can take it over the finish line if
39:42you want to. And you have an agent to do
39:44all of the changes that you want. It's
39:46pretty sweet. I don't know, that's kind
39:48of how I do most of it now. For certain
39:51things I might still
39:53um
39:54I actually don't know. Like it's become
39:55less and less. I still use like a local
39:57tool we for if I need like some
39:59credentials and I can't be bothered to
40:01secure them really You know, like I
40:02don't want to paste credentials
40:04somewhere and be have like better OIDC
40:06support now for credentials and whatnot,
40:08but I don't know. I don't know.
40:09Sometimes out of habit I open the local
40:11CLI, but I've been doing a lot like
40:13this. Like this is actually how I work.
40:15I just had a conversation I'll go in
40:18Slack with the others and I was saying
40:20it changed so much just in the last four
40:22weeks. Like we released a bunch of stuff
40:24and now our shipping velocity went up.
40:27Everybody is like Tim, one of our
40:29designers who codes a lot, he said he
40:31never fixed so many paper cuts and
40:34little polish things as in the last two
40:36weeks because every time he finds
40:39something, he just takes a screenshot
40:41basically and sends an orb and then he
40:43has like this orb. Then he does like
40:45spot checks and he's like go and merge
40:47it, you know? And now since last Friday
40:51we have agent to agent communication. So
40:54I can tell an agent, for example if it
40:56says like this, I found another bug,
40:57then I can say launch another orb to fix
41:00this bug, you know? But you keep doing
41:02your work. And then it spawns another
41:04thing in another checkout on another
41:06branch that fixes the bug and then I can
41:08come back to it. My predictions are I
41:10think the local development is going
41:11away. I think the models are dead in the
41:13sense that the distinction between
41:15different models will matter less in the
41:17future. I think the models will get to a
41:19point where it's a John Carmack on a
41:20button. I think it's not clear what
41:23software will survive. Like it's not
41:25clear what software will be worth
41:27building in the future. I think right
41:29now people are saying it's moving all
41:31down to infra. I think everybody moving
41:33to infra is going to kill the margins in
41:36infra which means there's now 15
41:38different sandbox providers, you know,
41:40that eating away each other's margins
41:41which is also not a sustainable
41:43business. So the question is what's left
41:45and I think we're all try we're all
41:47figuring this out right now, but we're
41:49all moving up to a higher level of
41:51building software. I think there's going
41:53to be a lot to build. The question is
41:54just what is that going to look like?
41:56>> I feel that's an amazing point to end it
41:57on. Thank you for listening so much.
41:59Where should people go?
42:00>> Yeah, I'm I'm still active a lot on X,
42:02Torsten Ball. I write a weekly
42:04newsletter called register spill where
42:06I, you know, spill what's in my
42:08registers. I've been sharing a bunch of
42:10links. I'm now having fun with it
42:11recording audio notes and whatnot. Go to
42:14anchor.fm, use amp. We now have
42:16subscriptions, we have orbs, we have
42:18multiplayer coming, H2H communication.
42:21We have a bunch of stuff. Check that
42:22out. Let me know what you think of pug.
42:24And yeah, find me online.
42:26>> All right, I'm going to link all of that
42:27below the video. Once again, thank you
42:29for your time and thank you for joining
42:31this podcast.
42:31>> Thank you.

About this transcript

This page contains the full transcript of Agentic Engineering, explained by a 10x developer by David Ondrej, generated from the public captions YouTube serves with the video. The transcript has 9,180 words across 1,269 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.