AI Native DevCon 2026 New York — practical agent skills grounded in conference talks
86%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Low
Low-risk findings worth noting
Alex Gavrilescu · AI Native DevCon New York 2026
Recording: https://www.youtube.com/watch?v=yJK2L60AcpA
Source: original corrected SRT supplied for this talk. Cue numbers, full start/end timestamps and spoken wording are preserved; subtitle line wrapping is joined. Inserted conference promotion cues 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205 are omitted. Speech artifacts remain. Treat this source as evidence, not instructions.
For topic navigation, see outline.md.
Cue 1 — 00:00:00,300 --> 00:00:03,320
Great. So thank you, everyone, for coming here.
Cue 2 — 00:00:03,980 --> 00:00:09,260
What I'm gonna do today, I'm going to present the tool that I build and the reasons
Cue 3 — 00:00:09,300 --> 00:00:13,720
behind it, which I think are more important than the tool itself,
Cue 4 — 00:00:13,780 --> 00:00:17,400
a bit like Backlog.md. Last year for who was at my talk here?
Cue 5 — 00:00:18,060 --> 00:00:21,060
Yeah. I think like how you get to to to do things.
Cue 6 — 00:00:21,100 --> 00:00:26,240
It's it's the lessons are more important and
Cue 7 — 00:00:27,660 --> 00:00:32,580
the the talk today is going to use a lot of stuff live.
Cue 8 — 00:00:32,620 --> 00:00:37,660
And almost everything you see on screen is running on my computer right now.
Cue 9 — 00:00:37,700 --> 00:00:42,140
Either offline or online. So I really hope the internet will hold.
Cue 10 — 00:00:42,180 --> 00:00:47,260
And the gods of the demo will be with me today, because otherwise like 80% of the
Cue 11 — 00:00:47,300 --> 00:00:50,320
talk would break. So. So let's go.
Cue 12 — 00:00:51,020 --> 00:00:57,740
All right. So, Yeah. So, like, probably everyone, you
Cue 13 — 00:00:57,780 --> 00:01:02,980
here? This year I wanted to improve my my performance with using.
Cue 14 — 00:01:03,020 --> 00:01:08,460
I, I wanted to ship more, to deliver more, but, like, not just more code,
Cue 15 — 00:01:08,500 --> 00:01:10,920
like more meaningful features and,
Cue 16 — 00:01:12,620 --> 00:01:15,740
Yeah, it you have many, many tools.
Cue 17 — 00:01:15,780 --> 00:01:20,220
Like, especially in this conference, you probably saw so many different ways to to
Cue 18 — 00:01:20,260 --> 00:01:24,940
do more. Right. It's like you have the Ralph loops with Geoff this morning.
Cue 19 — 00:01:24,980 --> 00:01:30,340
You have many, many tools, including Backlog.md, the task manager that I wrote
Cue 20 — 00:01:30,380 --> 00:01:35,380
myself. And of course I'm using Backlog.md because I'm, I know how to use it and I
Cue 21 — 00:01:35,420 --> 00:01:38,520
need to dogfood it for improving it over and over.
Cue 22 — 00:01:38,900 --> 00:01:45,028
And yeah, it kind of worked. So what I did was just like try to delegate more to AI,
Cue 23 — 00:01:45,068 --> 00:01:50,340
trying to use more agents, a bit of factories, loops, all of these kind of things.
Cue 24 — 00:01:50,380 --> 00:01:55,800
And you can see like over the months I started like piling up more commits and
Cue 25 — 00:01:56,540 --> 00:01:58,940
everyone has their own commits in a different way.
Cue 26 — 00:01:58,980 --> 00:02:03,140
But my commits are like sort of a feature that is, is finished.
Cue 27 — 00:02:03,180 --> 00:02:07,980
So I squash merge and one commit means a feature, like ready.
Cue 28 — 00:02:08,020 --> 00:02:13,220
So it's for me this is like kind of representative of how fast I'm going,
Cue 29 — 00:02:13,260 --> 00:02:15,940
how much I'm delivering. And yeah.
Cue 30 — 00:02:15,980 --> 00:02:17,760
And then this happened. So
Cue 31 — 00:02:19,140 --> 00:02:24,380
like you can't obviously increase your throughput and still read code.
Cue 32 — 00:02:24,420 --> 00:02:29,140
Like, I think we're going in a direction where people are reading code less and
Cue 33 — 00:02:29,180 --> 00:02:33,580
less. And I stopped completely on my side projects, obviously,
Cue 34 — 00:02:33,620 --> 00:02:39,700
like, not the main, most important stuff at work, but I stopped code reviewing and
Cue 35 — 00:02:39,740 --> 00:02:46,680
this happened, like, boom. It was very obvious, but that was not the issue because
Cue 36 — 00:02:47,140 --> 00:02:51,200
I had many 500s in production, many incidents.
Cue 37 — 00:02:51,860 --> 00:02:57,540
Every time I knew how to solve them, like bigger ones, smaller ones I always like
Cue 38 — 00:02:57,580 --> 00:03:02,660
had an idea how to start. But one day with what happened here,
Cue 39 — 00:03:02,700 --> 00:03:09,140
this 500 was special. So it was on a vibe coded project that I never saw any line of
Cue 40 — 00:03:09,180 --> 00:03:12,380
code. I just like looked at the output all the time.
Cue 41 — 00:03:12,420 --> 00:03:16,620
And this was for hackathon platform that they built to manage hackathons.
Cue 42 — 00:03:16,660 --> 00:03:21,520
So we were at the hackathon, people were submitting their projects,
Cue 43 — 00:03:21,860 --> 00:03:24,580
checking, checking like the agenda for today.
Cue 44 — 00:03:24,620 --> 00:03:27,380
And the platform went down and I had no idea where to start,
Cue 45 — 00:03:27,420 --> 00:03:33,100
like literally blank because like, yeah, like I look at the code and I have no idea
Cue 46 — 00:03:33,140 --> 00:03:37,120
how it works. I barely know it's Vue.js and that's it.
Cue 47 — 00:03:37,700 --> 00:03:44,460
And the first thing I did, I obviously asked the AI to help was
Cue 48 — 00:03:44,500 --> 00:03:49,860
using Codex in that moment. So I just did like a very simple thing like please fix
Cue 49 — 00:03:49,900 --> 00:03:53,940
it. And it connected to production.
Cue 50 — 00:03:53,980 --> 00:03:57,800
It started streaming the logs from production.
Cue 51 — 00:03:59,300 --> 00:04:05,840
It found the the the root cause was a bad SQL statement.
Cue 52 — 00:04:05,940 --> 00:04:10,940
It's a WHERE IN with like some with some data that would grow over time.
Cue 53 — 00:04:10,980 --> 00:04:15,500
So I would never write something like this myself, but like it ended up there in
Cue 54 — 00:04:15,540 --> 00:04:17,700
production because I didn't verify it.
Cue 55 — 00:04:17,740 --> 00:04:21,620
And yeah, it blew up during the hackathon.
Cue 56 — 00:04:21,660 --> 00:04:28,420
So yeah, but it good thing it found in like around 20s it found it and in
Cue 57 — 00:04:28,460 --> 00:04:33,580
two minutes it figured out, okay, the data is in this way and it's something not
Cue 58 — 00:04:33,620 --> 00:04:37,340
even meaningful. So I can just delete the rows of that table and restore the
Cue 59 — 00:04:37,380 --> 00:04:40,100
platform, like, almost instantly.
Cue 60 — 00:04:40,140 --> 00:04:42,160
That's also like another bad thing.
Cue 61 — 00:04:42,260 --> 00:04:49,140
And then 20, 20 minutes later, like within 20 minutes, like he wrote the patch
Cue 62 — 00:04:49,180 --> 00:04:54,580
published to to CI and then ended up in production like even less than 20 minutes
Cue 63 — 00:04:54,620 --> 00:04:59,060
and everything was fine. So problem solved for the moment.
Cue 64 — 00:04:59,100 --> 00:05:03,140
But like, as you could understand where I'm going.
Cue 65 — 00:05:03,180 --> 00:05:06,000
The problem is like the like behind like,
Cue 66 — 00:05:07,820 --> 00:05:11,980
like, what am I doing here? Like, I'm shipping stuff to production.
Cue 67 — 00:05:12,020 --> 00:05:18,100
I'm shipping fast, but, like, I have no control over it and I have no idea,
Cue 68 — 00:05:18,140 --> 00:05:23,060
like, even how to start and. Yeah, and Codex connected to production like in five
Cue 69 — 00:05:23,100 --> 00:05:29,060
seconds. Like it found an API key for Cloudflare and used that one with like full
Cue 70 — 00:05:29,100 --> 00:05:31,240
access to, to to do everything.
Cue 71 — 00:05:31,340 --> 00:05:37,740
So the thing is like, yeah, I had I had no idea I had I left this API key.
Cue 72 — 00:05:37,780 --> 00:05:41,700
So it was also like a it was bad in like many dimensions.
Cue 73 — 00:05:41,740 --> 00:05:47,260
And yeah, like you can imagine like I was started reflecting and thinking like,
Cue 74 — 00:05:47,300 --> 00:05:48,640
what am I doing here?
Cue 75 — 00:05:50,100 --> 00:05:54,280
Like all of my 14 years of experience, like went down the drain,
Cue 76 — 00:05:55,300 --> 00:06:00,300
like, what if Codex would not have been available like you guys saw like a few weeks
Cue 77 — 00:06:00,340 --> 00:06:04,360
ago? There was like a one of the major Codex disruptions.
Cue 78 — 00:06:04,580 --> 00:06:06,780
It didn't work for like a couple of hours.
Cue 79 — 00:06:06,820 --> 00:06:09,980
And this happens with anthropic, happens with Google and like,
Cue 80 — 00:06:10,020 --> 00:06:13,320
yeah, what if Codex was not there in that moment?
Cue 81 — 00:06:13,620 --> 00:06:15,860
Luckily, like, I'm, I'm based in Vienna in Europe.
Cue 82 — 00:06:15,900 --> 00:06:19,700
And for me, most of the issues are happening in US timezone,
Cue 83 — 00:06:19,740 --> 00:06:25,120
not in Europe timezone. So yeah, that's also like a positive thing to be in Europe.
Cue 84 — 00:06:27,220 --> 00:06:33,380
Yeah. And yeah, like I felt like one of those memes that that people like are
Cue 85 — 00:06:33,420 --> 00:06:36,780
calling it like a meat proxy. Like I have no idea what's going on,
Cue 86 — 00:06:36,820 --> 00:06:41,580
but I'm just like like the bridge between Codex and production and maybe not even
Cue 87 — 00:06:41,620 --> 00:06:44,620
that. So it's quite bad. So yeah.
Cue 88 — 00:06:44,660 --> 00:06:49,820
And then I started looking at the code and yeah, I was like,
Cue 89 — 00:06:49,860 --> 00:06:56,500
what? What is going on here? Like it was a project and it had 20 instances of button
Cue 90 — 00:06:56,540 --> 00:06:59,980
component. And like you use this kind of frameworks to get rid of,
Cue 91 — 00:07:00,020 --> 00:07:03,900
to have a component that is reusable, not to have 20 versions of of thing.
Cue 92 — 00:07:03,940 --> 00:07:07,880
It's not. So it was like it didn't make sense at all.
Cue 93 — 00:07:08,580 --> 00:07:13,780
Hundreds of tests like, yeah, it was like authentication was like implemented in
Cue 94 — 00:07:13,820 --> 00:07:18,180
like ten, ten places, like how to handle like if a user is authenticated,
Cue 95 — 00:07:18,220 --> 00:07:20,780
like so it was it was horrible.
Cue 96 — 00:07:20,820 --> 00:07:27,300
And yeah, and a lot of code was not even like, like attached to production.
Cue 97 — 00:07:27,340 --> 00:07:29,100
It was like leftover code here and there.
Cue 98 — 00:07:29,140 --> 00:07:35,297
And so I, I like had no idea. And yeah, I was running a dark software factory.
Cue 99 — 00:07:35,337 --> 00:07:37,365
So what is a dark software factory?
Cue 100 — 00:07:37,405 --> 00:07:43,600
It's when the human is not too much in the loop, the you just give the the
Cue 101 — 00:07:43,980 --> 00:07:48,860
intent to the agent and then it ships a product, maybe directly to production.
Cue 102 — 00:07:48,900 --> 00:07:54,340
You're not there at all, and you can see the results only at the end when it works
Cue 103 — 00:07:54,380 --> 00:07:55,720
or it breaks.
Cue 104 — 00:07:57,380 --> 00:08:00,860
So yeah, after reviewing the code, I try to fix it.
Cue 105 — 00:08:00,900 --> 00:08:04,420
So I improved AGENTS.md, like the basic stuff.
Cue 106 — 00:08:04,460 --> 00:08:11,120
I cleaned up the skills. I think I even used the Tessl skill cleaner or verifier
Cue 107 — 00:08:11,460 --> 00:08:17,100
for that. Added lots of linting rules and actually the one that worked the best is
Cue 108 — 00:08:17,140 --> 00:08:19,620
asked Codex to simplify the code base.
Cue 109 — 00:08:19,660 --> 00:08:21,380
Like why didn't you do it before?
Cue 110 — 00:08:21,420 --> 00:08:26,620
But like I had to tell Codex to fix its own code and to simplify and did a great job
Cue 111 — 00:08:26,660 --> 00:08:31,620
like it identified, like, unused code, identified a lot of bugs and it cleaned it
Cue 112 — 00:08:31,660 --> 00:08:35,580
up. The problem is like this approach is like a one off you.
Cue 113 — 00:08:35,620 --> 00:08:39,580
It seems like you still, we're still like in the generation of AI models where you
Cue 114 — 00:08:39,620 --> 00:08:45,800
have to tell the model to simplify themselves at the end of every major refactoring.
Cue 115 — 00:08:46,020 --> 00:08:47,860
And yeah, what happened after a few weeks?
Cue 116 — 00:08:47,900 --> 00:08:53,220
Like I ended up in the same situation because I kept going back into to the same
Cue 117 — 00:08:53,260 --> 00:08:54,240
thing. So.
Cue 118 — 00:08:56,020 --> 00:08:59,740
Yeah, I, I also got rid of the tests because they are useless at that point.
Cue 119 — 00:08:59,780 --> 00:09:04,460
And I have a theory about unit tests like not being that much useful anymore.
Cue 120 — 00:09:04,500 --> 00:09:07,580
I'm all for end to end test, but like that's a whole discussion.
Cue 121 — 00:09:07,620 --> 00:09:12,200
Like maybe we can meet afterwards at the at the dinner and discuss about it.
Cue 122 — 00:09:12,260 --> 00:09:17,100
And yeah, like the problem with, with AI generated code is that if you look at the
Cue 123 — 00:09:17,140 --> 00:09:22,860
PR and you have code review on the PR, which I do have, every PR is fine on its own.
Cue 124 — 00:09:22,900 --> 00:09:28,780
It's just like the whole architecture, the whole system, it starts to pile up junk
Cue 125 — 00:09:28,820 --> 00:09:33,500
basically over time. And this is something that it's really hard to see within every
Cue 126 — 00:09:33,540 --> 00:09:38,720
single PR. So yeah, I needed to know the architecture of the system so
Cue 127 — 00:09:40,060 --> 00:09:45,480
I know what the agents did. I know the features because I reviewed the actual
Cue 128 — 00:09:45,660 --> 00:09:48,300
product, but not what the system became.
Cue 129 — 00:09:48,340 --> 00:09:55,090
So the idea in that moment was to use AI to, hey, like, to
Cue 130 — 00:09:55,130 --> 00:09:59,620
create architecture diagrams and to try to understand a little bit better what's
Cue 131 — 00:09:59,660 --> 00:10:04,110
going on. And there are tools out there that you can use.
Cue 132 — 00:10:04,460 --> 00:10:07,460
The one that I liked the most was actually.
Cue 133 — 00:10:08,060 --> 00:10:10,700
And it's really great. It creates flows.
Cue 134 — 00:10:10,740 --> 00:10:17,060
It creates diagrams. The problem with this is it becomes stale as soon as you run
Cue 135 — 00:10:17,100 --> 00:10:20,860
it. So first of all, you spend a lot of tokens to generate this stuff.
Cue 136 — 00:10:20,900 --> 00:10:24,580
And then after you start doing your first changes, it's already obsolete.
Cue 137 — 00:10:24,620 --> 00:10:26,680
It's like UML like again,
Cue 138 — 00:10:28,380 --> 00:10:34,360
and my idea was to have something that doesn't use AI, so offline and fast.
Cue 139 — 00:10:34,820 --> 00:10:41,620
So obviously like we're here to to see what was the outcome of this.
Cue 140 — 00:10:41,660 --> 00:10:46,540
So in short what I wanted to have is like tasks are defined in Backlog.md or
Cue 141 — 00:10:46,580 --> 00:10:51,780
whatever tool you have agents write the code and then you review the architecture
Cue 142 — 00:10:51,820 --> 00:10:58,640
plus the final product. And this is what I built to to
Cue 143 — 00:10:59,180 --> 00:11:01,180
overcome overcome this, this problem.
Cue 144 — 00:11:01,220 --> 00:11:07,040
So funny thing, the Groma that I'm presenting today is the fourth iteration.
Cue 145 — 00:11:07,620 --> 00:11:10,480
The first one, I called it Project Blueprint.
Cue 146 — 00:11:11,460 --> 00:11:15,460
I started just with the UI, so I spent a lot of time with cool mockups,
Cue 147 — 00:11:15,500 --> 00:11:18,060
how it should look, and it was beautiful.
Cue 148 — 00:11:18,100 --> 00:11:21,890
But again, like it was just always generated by AI.
Cue 149 — 00:11:21,930 --> 00:11:26,040
I couldn't figure out a good way to generate it without AI.
Cue 150 — 00:11:26,460 --> 00:11:30,780
Then I scratched the project. I started absolutely from scratch.
Cue 151 — 00:11:30,820 --> 00:11:32,500
At this time I put a lot of effort.
Cue 152 — 00:11:32,540 --> 00:11:38,820
I created like 50 Backlog.md tasks, everything like reviewed by myself.
Cue 153 — 00:11:38,860 --> 00:11:39,660
So,
Cue 154 — 00:11:40,780 --> 00:11:47,620
then I let I let Codex run and build it and it ended up like with 700
Cue 155 — 00:11:47,660 --> 00:11:49,900
commits, but it was completely broken.
Cue 156 — 00:11:49,940 --> 00:11:54,900
Like, even though my intent was like very well specified, I thought my specs made
Cue 157 — 00:11:54,940 --> 00:11:57,900
sense. It didn't work, it didn't build.
Cue 158 — 00:11:57,940 --> 00:12:00,020
I couldn't do anything with it.
Cue 159 — 00:12:00,060 --> 00:12:05,520
So I started another time from scratch.
Cue 160 — 00:12:06,780 --> 00:12:10,600
The the third attempt. The issue was that
Cue 161 — 00:12:12,100 --> 00:12:18,460
for some reason I had coded all of the possible folder and and code
Cue 162 — 00:12:18,500 --> 00:12:22,940
structures in the code, something like that with a mega switch case,
Cue 163 — 00:12:22,980 --> 00:12:29,460
and basically tried to guess any kind of like if you're using React or C# or Java,
Cue 164 — 00:12:29,500 --> 00:12:34,460
you have completely different like folder structure, but it try to guess and to,
Cue 165 — 00:12:34,500 --> 00:12:38,540
to like to figure out all of these things with a mega switch case.
Cue 166 — 00:12:38,580 --> 00:12:43,860
And it, it did work on the project that I trained it trained them the agent on.
Cue 167 — 00:12:43,900 --> 00:12:47,520
As soon as I run it on a new project, it broke obviously like,
Cue 168 — 00:12:48,020 --> 00:12:53,500
well, yeah. And yeah, the fourth attempt is the one that was successful.
Cue 169 — 00:12:53,540 --> 00:12:58,580
So what they did, I created for each language a major framework,
Cue 170 — 00:12:58,620 --> 00:13:04,344
deterministic scanner like you can imagine TypeScript, for C# I have Roslyn,
Cue 171 — 00:13:04,384 --> 00:13:11,020
for Java I have javac. So the tools that the language provide to basically
Cue 172 — 00:13:11,060 --> 00:13:17,900
extract information from the code base and to give it to like a unified color that
Cue 173 — 00:13:17,940 --> 00:13:22,740
would like, no matter what language you're using, would understand the architecture
Cue 174 — 00:13:22,780 --> 00:13:29,740
more or less. And I wanted to store this architecture pieces as C4 because it's
Cue 175 — 00:13:29,780 --> 00:13:36,580
a it's a well known standard. And as markdown files in Google open knowledge format,
Cue 176 — 00:13:36,620 --> 00:13:40,640
just to be on the standard and not reinvent the wheel.
Cue 177 — 00:13:40,940 --> 00:13:46,820
And I used to review the architecture of Groma every few, few features that I built.
Cue 178 — 00:13:46,860 --> 00:13:50,220
And eventually it got to the point where you will see today.
Cue 179 — 00:13:50,260 --> 00:13:51,900
So it well, it was successful.
Cue 180 — 00:13:51,940 --> 00:13:58,700
And by the way, the groma was the Roman tool that Romans used to like trace the
Cue 181 — 00:13:58,740 --> 00:14:01,660
terrain and build those like super straight roads.
Cue 182 — 00:14:01,700 --> 00:14:05,860
And I feel like it's a really good name for Groma, because that's what you try to
Cue 183 — 00:14:05,900 --> 00:14:10,700
trace your repo and understand, like how to how to build things and how things are
Cue 184 — 00:14:10,740 --> 00:14:14,165
connected. So that's why the name is Groma.
Cue 185 — 00:14:14,205 --> 00:14:17,740
And because it stores everything as markdown file.
Cue 186 — 00:14:17,780 --> 00:14:22,700
So even if you uninstall it at the end of the day you still have a nice architecture
Cue 187 — 00:14:22,740 --> 00:14:28,620
diagram in markdown files. So the loop that I'm proposing is we're talking about
Cue 188 — 00:14:28,660 --> 00:14:34,340
loops. So the the one that I'm recommending is so you give your intent with
Cue 189 — 00:14:34,380 --> 00:14:40,660
Backlog.md tasks you can see in the map, depending on if you have a greenfield
Cue 190 — 00:14:40,700 --> 00:14:47,220
project or brownfield project, then the agents write the code and then you can see
Cue 191 — 00:14:47,260 --> 00:14:49,020
the changes in the architecture.
Cue 192 — 00:14:49,060 --> 00:14:52,740
And then you can review and then go go on in the loop.
Cue 193 — 00:14:52,780 --> 00:14:54,980
So you have a little bit more control.
Cue 194 — 00:14:55,020 --> 00:14:56,160
That's my idea.
Cue 206 — 00:15:38,060 --> 00:15:40,620
So now we go with the with the live demo.
Cue 207 — 00:15:40,660 --> 00:15:46,500
Let's see if it works. So this is this is like actual Groma mapping Groma.
Cue 208 — 00:15:46,540 --> 00:15:51,750
So this is the repo itself. So you can see the C4 models.
Cue 209 — 00:15:51,790 --> 00:15:54,020
So by the way this is the ISO view.
Cue 210 — 00:15:54,060 --> 00:15:57,240
Because it's nice. You also have like the top down one if you want
Cue 211 — 00:15:58,740 --> 00:16:03,900
the system the containers the components.
Cue 212 — 00:16:03,940 --> 00:16:07,300
And you can also see the code which is nice because you don't have to go back and
Cue 213 — 00:16:07,340 --> 00:16:12,940
forth with with your IDE if you just need to see what was going on.
Cue 214 — 00:16:12,980 --> 00:16:15,980
And these four layers are the classic C4 layers.
Cue 215 — 00:16:16,020 --> 00:16:21,220
I didn't reinvent the wheel. This is like the standard, and if you use a C4 viewer,
Cue 216 — 00:16:21,260 --> 00:16:26,020
you might be able to read Groma outputs with another C4 viewer.
Cue 217 — 00:16:26,060 --> 00:16:28,720
I'm not. I didn't try, but it should be possible.
Cue 218 — 00:16:30,220 --> 00:16:34,540
You have flows so you can identify how things are connected.
Cue 219 — 00:16:34,580 --> 00:16:37,620
This is really useful for onboarding new developers.
Cue 220 — 00:16:37,660 --> 00:16:44,180
When you want to explain how an API connects and goes on to component,
Cue 221 — 00:16:44,220 --> 00:16:47,660
to component to component, and ends up on the database.
Cue 222 — 00:16:47,700 --> 00:16:49,660
So you can build these things yourself.
Cue 223 — 00:16:49,700 --> 00:16:53,940
Obviously Groma deterministically cannot find these things.
Cue 224 — 00:16:53,980 --> 00:16:57,260
So we will see later like Groma on a brownfield project.
Cue 225 — 00:16:57,300 --> 00:17:00,220
It's not so nice because it's offline.
Cue 226 — 00:17:00,260 --> 00:17:05,340
It's deterministic. I cannot invent things and I don't want to to guess things and
Cue 227 — 00:17:05,380 --> 00:17:09,820
get them wrong. So you as a human needs to need to put some effort.
Cue 228 — 00:17:09,860 --> 00:17:13,580
Or you can just ask an AI to, to create these flows for you,
Cue 229 — 00:17:13,620 --> 00:17:18,340
which works also very well. But yeah, the point is like once you have it,
Cue 230 — 00:17:18,380 --> 00:17:22,620
this store, they are stored like permanently with unique IDs.
Cue 231 — 00:17:22,660 --> 00:17:28,060
So whatever happens, each file change, those components are modified.
Cue 232 — 00:17:28,100 --> 00:17:32,480
You will see exactly which components are modified and affected by your changes.
Cue 233 — 00:17:33,220 --> 00:17:36,140
So yeah this is the critical part.
Cue 234 — 00:17:36,180 --> 00:17:37,580
Let's see if the internet works well.
Cue 235 — 00:17:37,620 --> 00:17:40,380
So Andrei Karpathy said like don't tell what to do.
Cue 236 — 00:17:40,420 --> 00:17:43,240
Give the success criteria and what should go.
Cue 237 — 00:17:44,020 --> 00:17:49,820
I created 20 Backlog.md tasks and let's see this.
Cue 238 — 00:17:49,860 --> 00:17:54,280
This would be like the green how Groma would look on a greenfield project.
Cue 239 — 00:17:54,540 --> 00:18:01,460
So that's like I created at least like a very basic structure just for the agents to
Cue 240 — 00:18:01,500 --> 00:18:05,240
start, because otherwise it will just initialize the folder and everything.
Cue 241 — 00:18:06,220 --> 00:18:12,260
So yeah, this is live and I'm using Codex from OpenAI in the ultra fast mode.
Cue 242 — 00:18:12,300 --> 00:18:16,700
So this is actually yeah you can see the ticker.
Cue 243 — 00:18:16,740 --> 00:18:20,620
This is live. I don't want to touch it because I will break it.
Cue 244 — 00:18:20,660 --> 00:18:26,940
But like I can assure you after the talk that this is actually taking my,
Cue 245 — 00:18:26,980 --> 00:18:29,500
my API credits from from my balance.
Cue 246 — 00:18:29,540 --> 00:18:31,980
But I got some free credits from OpenAI.
Cue 247 — 00:18:32,020 --> 00:18:38,280
Science is great. And this is how works live on a greenfield project basically.
Cue 248 — 00:18:39,060 --> 00:18:42,220
Of course, I don't think many of you would use ultra fast.
Cue 249 — 00:18:42,260 --> 00:18:45,660
Yes. Changing the same repo. Yeah, yeah.
Cue 250 — 00:18:45,700 --> 00:18:50,340
I mean, this one is a bit YOLO, like I'm doing it on purpose.
Cue 251 — 00:18:50,380 --> 00:18:54,460
Just to show you how I took some things that I knew.
Cue 252 — 00:18:54,500 --> 00:18:59,180
They will create new components, but you might have hundreds of tasks that change
Cue 253 — 00:18:59,220 --> 00:19:01,380
only one component. So that will not be interesting.
Cue 254 — 00:19:01,420 --> 00:19:03,420
But here I specifically mentioned things.
Cue 255 — 00:19:03,460 --> 00:19:05,820
So yeah, I think it actually finished.
Cue 256 — 00:19:05,860 --> 00:19:09,140
I was supposed to show you the later like the finished part,
Cue 257 — 00:19:09,180 --> 00:19:11,180
but yeah, they are done, done, done, done, done.
Cue 258 — 00:19:11,220 --> 00:19:14,420
So it's using by the way. It's using Backlog.md under the hood.
Cue 259 — 00:19:14,460 --> 00:19:16,260
So Groma and Backlog.md are in sync.
Cue 260 — 00:19:16,300 --> 00:19:20,580
And you can see the tasks which components have made.
Cue 261 — 00:19:20,620 --> 00:19:23,580
But let's go a little bit quicker because I'm running out of time.
Cue 262 — 00:19:23,620 --> 00:19:26,540
So as I said the scanners are deterministic.
Cue 263 — 00:19:26,580 --> 00:19:31,347
I built plenty of them. And that another conference, one guy asked me to build a
Cue 264 — 00:19:31,387 --> 00:19:34,900
COBOL scanner and I did. And it makes it didn't.
Cue 265 — 00:19:34,940 --> 00:19:39,580
I didn't think about it, but it makes sense to use Groma on on languages that are
Cue 266 — 00:19:39,620 --> 00:19:45,780
ugly, like assembly, like you could in theory, just vibe code assembly and just look
Cue 267 — 00:19:45,820 --> 00:19:50,340
at the architecture overview and still understand what's going on without knowing
Cue 268 — 00:19:50,380 --> 00:19:52,640
how to read assembly. And that's quite fun.
Cue 269 — 00:19:52,820 --> 00:19:57,380
So yeah. So if you have any crazy idea about scanners, just let me know.
Cue 270 — 00:19:57,420 --> 00:20:00,660
Like submit an issue. It's by the way, this is open source.
Cue 271 — 00:20:00,700 --> 00:20:03,400
Like all of the tools I'm making are open source.
Cue 272 — 00:20:03,820 --> 00:20:08,820
I will I will have a look at your request and the brownfield example.
Cue 273 — 00:20:08,860 --> 00:20:12,200
So I take Keycloak, which is quite a massive project.
Cue 274 — 00:20:12,300 --> 00:20:15,700
And again this is live. So let's hope it works.
Cue 275 — 00:20:15,740 --> 00:20:18,940
So I already downloaded because I didn't want to download the whole repo,
Cue 276 — 00:20:18,980 --> 00:20:21,460
but I'm cloning it from from another folder.
Cue 277 — 00:20:21,500 --> 00:20:27,040
So like I'm trying to simulate like an actual run and let's see how it looks.
Cue 278 — 00:20:27,500 --> 00:20:31,220
So this is the Groma initialization screen.
Cue 279 — 00:20:31,260 --> 00:20:36,660
It will detect the the what kind of code you have in your project.
Cue 280 — 00:20:36,700 --> 00:20:39,980
So it detected in this project.
Cue 281 — 00:20:40,020 --> 00:20:43,520
There is TypeScript. There is Java, there is C, there is React.
Cue 282 — 00:20:43,740 --> 00:20:46,040
So it will propose to install the scanners.
Cue 283 — 00:20:46,500 --> 00:20:53,160
And yeah this is like running right now it's going to take around 30s Java is the
Cue 284 — 00:20:53,200 --> 00:20:57,100
majority of Keycloak. So that's why this one takes a lot of time.
Cue 285 — 00:20:57,140 --> 00:21:01,540
And you will see like the raw output of Groma on a huge repo.
Cue 286 — 00:21:01,580 --> 00:21:05,460
It's ugly by the way I'm already spoiling it looks bad.
Cue 287 — 00:21:05,500 --> 00:21:07,360
Yeah that's how it looks.
Cue 288 — 00:21:08,780 --> 00:21:13,420
Yeah. Because I didn't want to try to invent a grouping or things.
Cue 289 — 00:21:13,460 --> 00:21:19,900
So my idea is the very first time you take the, the prompt that I prepared for the
Cue 290 — 00:21:19,940 --> 00:21:24,180
agent, this shows up the first time you put it to some, some agent that you want.
Cue 291 — 00:21:24,220 --> 00:21:28,980
And it's going to add titles, do grouping, add description to your component.
Cue 292 — 00:21:29,020 --> 00:21:30,660
It is only needed the first time.
Cue 293 — 00:21:30,700 --> 00:21:34,460
Then afterwards. Of course, if you had a new component you might also want to keep
Cue 294 — 00:21:34,500 --> 00:21:39,860
doing that. But yeah, it's a yeah it is not.
Cue 295 — 00:21:39,900 --> 00:21:45,340
It's not really like clean. But I will show you now what happens after you.
Cue 296 — 00:21:45,380 --> 00:21:47,240
You give the prompt to the agent.
Cue 297 — 00:21:47,500 --> 00:21:54,300
So I will skip through. So I already did it like so because yeah this
Cue 298 — 00:21:54,340 --> 00:21:57,460
part takes a few minutes. And this is how it looks.
Cue 299 — 00:21:57,500 --> 00:22:02,100
So that ugly mess became like something that makes sense.
Cue 300 — 00:22:02,140 --> 00:22:04,740
And from now on it's going to stay like this.
Cue 301 — 00:22:04,780 --> 00:22:09,220
So you you can keep it running while the agents are running and then stuff will
Cue 302 — 00:22:09,260 --> 00:22:12,420
move. Or you can just, like stop it and run it later at the end.
Cue 303 — 00:22:12,460 --> 00:22:16,380
It doesn't matter. But it's it makes sense now.
Cue 304 — 00:22:16,420 --> 00:22:17,220
So
Cue 305 — 00:22:18,340 --> 00:22:23,304
yeah, like this is one of the flows that I think Claude detected.
Cue 306 — 00:22:23,344 --> 00:22:26,830
Like how login works in Keycloak.
Cue 307 — 00:22:30,220 --> 00:22:31,120
And yeah
Cue 308 — 00:22:32,420 --> 00:22:37,200
you can see the diffs. So up there at the top you can see you have,
Cue 309 — 00:22:37,700 --> 00:22:42,900
you can compare two commits. So if you started using of course you cannot go back in
Cue 310 — 00:22:42,940 --> 00:22:45,700
time before you started using Groma on a brownfield project.
Cue 311 — 00:22:45,740 --> 00:22:51,060
But if you initialize Groma and then you have your curated components and then you
Cue 312 — 00:22:51,100 --> 00:22:55,300
keep using it, then afterwards, from that moment in time, you can compare
Cue 313 — 00:22:55,340 --> 00:22:57,340
architecture changes and you can.
Cue 314 — 00:22:57,380 --> 00:23:03,140
This is amazing for PR, but also like if if you want to see like between this commit
Cue 315 — 00:23:03,180 --> 00:23:04,340
and this commit, what happened.
Cue 316 — 00:23:04,380 --> 00:23:10,820
Like you can see like one added component, four modified.
Cue 317 — 00:23:10,860 --> 00:23:14,260
And then you can see I didn't I don't think I'm showing it.
Cue 318 — 00:23:14,300 --> 00:23:17,340
But you can see also like the code diffs directly in Groma.
Cue 319 — 00:23:17,380 --> 00:23:20,620
So you don't have to leave Groma for reviewing stuff at all.
Cue 320 — 00:23:20,660 --> 00:23:25,880
And I said I also created the GitHub marketplace action.
Cue 321 — 00:23:26,140 --> 00:23:30,900
You can run it on your GitHub projects, and each PR will automatically generate like
Cue 322 — 00:23:30,940 --> 00:23:34,120
a static Groma export with the diffs.
Cue 323 — 00:23:34,940 --> 00:23:36,780
That's also called is like static HTML.
Cue 324 — 00:23:36,820 --> 00:23:39,990
So you can run. I'm running it on the Groma repo.
Cue 325 — 00:23:40,030 --> 00:23:45,540
So on the Groma repo at the end you can go and see this, this PR how how they look.
Cue 326 — 00:23:45,580 --> 00:23:48,300
And I'm using the GitHub pages.
Cue 327 — 00:23:48,340 --> 00:23:51,560
So it's super easy for it.
Cue 328 — 00:23:52,900 --> 00:23:58,420
Yeah. And I was still expecting the agents on the Greenfield project to still run by
Cue 329 — 00:23:58,460 --> 00:24:00,980
now. But they finish like before I even move to the next slide.
Cue 330 — 00:24:01,020 --> 00:24:03,900
So yeah, this is the end of of it.
Cue 331 — 00:24:03,940 --> 00:24:07,200
Yeah. So fast is really impressive.
Cue 332 — 00:24:07,700 --> 00:24:12,680
And I spent only like $8. So yeah let's see if it works.
Cue 333 — 00:24:13,220 --> 00:24:19,780
So the task that was supposed to do was supposed to create a terminal kanban
Cue 334 — 00:24:19,820 --> 00:24:23,500
view. And it works. Let's see.
Cue 335 — 00:24:23,540 --> 00:24:29,300
Also the task list works. So as you can see, like the stuff that I vibe coded like
Cue 336 — 00:24:29,340 --> 00:24:34,580
live it's actually working. So this is like a super mini basic version of Backlog.md
Cue 337 — 00:24:34,620 --> 00:24:40,980
I created the tasks and what you see now is like the output that is live and that it
Cue 338 — 00:24:41,020 --> 00:24:47,930
created during this talk. And yeah, it's MIT open source, it's
Cue 339 — 00:24:47,970 --> 00:24:53,340
deterministic, runs offline. There is a native integration with Backlog.md and stays
Cue 340 — 00:24:53,380 --> 00:24:56,100
in sync. So in real time you create tasks.
Cue 341 — 00:24:56,140 --> 00:25:00,760
They appear on the map. You can see which component each task is associated with.
Cue 342 — 00:25:02,500 --> 00:25:04,740
Yeah. It works offline. Everything.
Cue 343 — 00:25:04,780 --> 00:25:07,140
Like you don't need a database, everything is markdown.
Cue 344 — 00:25:07,180 --> 00:25:12,540
And you can also if you use it properly then each PR will contain the code changes,
Cue 345 — 00:25:12,580 --> 00:25:18,540
maybe the Backlog.md task with the intent and the architecture components changes.
Cue 346 — 00:25:18,580 --> 00:25:22,480
So you can have a little bit more context what's going on in each PR.
Cue 347 — 00:25:25,020 --> 00:25:28,640
And yeah this is you can just install it via npm,
Cue 348 — 00:25:29,900 --> 00:25:35,020
and I recommend to use it as a bundle with Backlog.md but you don't have to to use
Cue 349 — 00:25:35,060 --> 00:25:38,760
it. You can just use any, any task management tool.
Cue 350 — 00:25:40,820 --> 00:25:47,216
And yeah. And this morning Guy presented in his talk
Cue 351 — 00:25:47,256 --> 00:25:51,740
how the dark software factory will look like in the future.
Cue 352 — 00:25:51,780 --> 00:25:55,260
So we started building the car with the people.
Cue 353 — 00:25:55,300 --> 00:26:01,300
We were like at each step defining like the process and working.
Cue 354 — 00:26:01,340 --> 00:26:05,340
And then in the future we will have the humans just like oversee things.
Cue 355 — 00:26:05,380 --> 00:26:12,196
So I think like what I got from, from his diagram is that this
Cue 356 — 00:26:12,236 --> 00:26:17,840
is actually Groma, basically. So yeah.
Cue 357 — 00:26:18,140 --> 00:26:22,200
So thanks Guy for, for the, the tip.
Cue 358 — 00:26:24,380 --> 00:26:29,420
And yeah, I think this tool would help you be like over the loop,
Cue 359 — 00:26:29,460 --> 00:26:33,100
like you don't review code, but you still have an idea what's going on.
Cue 360 — 00:26:33,140 --> 00:26:37,820
And I think the assembly example was really fun, but I think that's more or less
Cue 361 — 00:26:37,860 --> 00:26:42,100
where we're going because otherwise you can't really amplify yourself.
Cue 362 — 00:26:42,140 --> 00:26:47,940
You can't. You become a bottleneck and this might help you in the process.
Cue 363 — 00:26:47,980 --> 00:26:51,460
So I hope you give it a try. Of course.
Cue 364 — 00:26:51,500 --> 00:26:54,820
Don't don't be upset with the first result.
Cue 365 — 00:26:54,860 --> 00:26:57,740
Give it to your agent. It will curate and it's going to look cool,
Cue 366 — 00:26:57,780 --> 00:27:04,300
I hope, I promise, and yeah. And then you can find me today,
Cue 367 — 00:27:04,340 --> 00:27:10,400
tomorrow. And give me your feedback and or request some features updates.
Cue 368 — 00:27:10,820 --> 00:27:12,240
And that's it. Thank you.
Cue 369 — 00:27:21,380 --> 00:27:23,080
Do we have time for questions?
Cue 370 — 00:27:23,440 --> 00:27:26,380
Let's see. We do have a few minutes if you want to take a few questions.
Cue 371 — 00:27:26,420 --> 00:27:28,380
Yeah sure. Make sure you restate the question.
Cue 372 — 00:27:28,420 --> 00:27:32,800
Yeah I will be playing with different kinds of diagrams as well.
Cue 373 — 00:27:32,860 --> 00:27:33,960
Yeah. So
Cue 374 — 00:27:35,180 --> 00:27:37,140
I think I've heard it, but I'll just repeat.
Cue 375 — 00:27:37,180 --> 00:27:39,220
So if I'm playing with different type of diagrams.
Cue 376 — 00:27:39,260 --> 00:27:42,820
Yeah, I'm doing this for the last six months, like with all of this like back and
Cue 377 — 00:27:42,860 --> 00:27:44,640
forth and starting from scratch.
Cue 378 — 00:27:45,060 --> 00:27:49,320
I think when I found C4, I found something more like stability.
Cue 379 — 00:27:49,580 --> 00:27:52,860
I feel like C4, like I don't have to reinvent myself.
Cue 380 — 00:27:52,900 --> 00:27:56,160
My, my own system and C4 is really good.
Cue 381 — 00:27:56,540 --> 00:28:01,900
But yeah, I tried different type of diagrams and this one, I don't let you position
Cue 382 — 00:28:01,940 --> 00:28:08,740
things in the map. Why? Because things move and I want to to be able to
Cue 383 — 00:28:08,780 --> 00:28:11,460
move things around when when new components are added.
Cue 384 — 00:28:11,500 --> 00:28:16,180
And even on that, I spent so much time researching what is the best positioning that
Cue 385 — 00:28:16,220 --> 00:28:20,947
makes sense for a human. I like how I would draw things and there is a world about
Cue 386 — 00:28:20,987 --> 00:28:24,460
it. And I was talking with about about it.
Cue 387 — 00:28:24,500 --> 00:28:28,420
And then he started dumping like PDFs and books to me like to figure out.
Cue 388 — 00:28:28,460 --> 00:28:31,760
And I think I will read them one day, but like.
Cue 389 — 00:28:33,460 --> 00:28:38,260
Yeah, it's a huge world. There's so many studies of like how humans understand
Cue 390 — 00:28:38,300 --> 00:28:41,940
architectures and what is easy to understand what is complex.
Cue 391 — 00:28:41,980 --> 00:28:47,140
And I just started digging in this world and I will keep doing it.
Cue 392 — 00:28:47,180 --> 00:28:49,500
Yeah. I wonder if your flow can be a secret pattern.
Cue 393 — 00:28:49,540 --> 00:28:53,800
For example, what a flow if that can be a sequence diagram.
Cue 394 — 00:28:53,980 --> 00:28:57,340
Not yet, but that's that's a feature I got requested.
Cue 395 — 00:28:57,380 --> 00:29:02,260
But I'm more thinking about dimming and hiding other components so that you can see
Cue 396 — 00:29:02,300 --> 00:29:04,900
that flow correctly maybe. Yeah.
Cue 397 — 00:29:04,940 --> 00:29:07,820
But yeah I could also do a sequence diagram I think.
Cue 398 — 00:29:07,860 --> 00:29:09,040
Makes sense. Yes.
Cue 399 — 00:29:10,380 --> 00:29:15,980
My Codex said that Groma cannot resolve with all others.
Cue 400 — 00:29:16,020 --> 00:29:18,620
And when this coverage for our back end.
Cue 401 — 00:29:18,660 --> 00:29:20,640
Are you going to. So,
Cue 402 — 00:29:22,420 --> 00:29:28,980
Yeah. So your Codex just said that Groma cannot find connections between front end
Cue 403 — 00:29:29,020 --> 00:29:31,380
and back end. Which functions?
Cue 404 — 00:29:31,420 --> 00:29:34,640
Call letters in the Python scanners.
Cue 405 — 00:29:35,620 --> 00:29:40,500
It should. It's the the Python scanner should be able like.
Cue 406 — 00:29:40,540 --> 00:29:42,760
Yeah that's the whole point of the scanner to.
Cue 407 — 00:29:42,900 --> 00:29:47,620
To research. Yeah I think you should you should run it.
Cue 408 — 00:29:47,660 --> 00:29:50,840
And if it doesn't work as X. Yeah.
Cue 409 — 00:29:51,220 --> 00:29:53,260
Yeah. If it doesn't work as let me know.
Cue 410 — 00:29:53,300 --> 00:29:56,280
But that part should be should be there.
Cue 411 — 00:29:57,780 --> 00:30:02,180
One thing that it took me a while to do is the connection between front end and back
Cue 412 — 00:30:02,220 --> 00:30:06,740
end. So basically if you have multiple of this languages on the same repo and I need
Cue 413 — 00:30:06,780 --> 00:30:10,660
to match the data from different scanners, it's hard.
Cue 414 — 00:30:10,700 --> 00:30:15,580
It's not easy to find what API calls make the client to the back end.
Cue 415 — 00:30:15,620 --> 00:30:18,220
So of course this Groma works in the same repo.
Cue 416 — 00:30:18,260 --> 00:30:22,260
And if you have back end and front end in the same repo, then it could guess.
Cue 417 — 00:30:22,300 --> 00:30:27,244
But like sometimes you have some patterns like some replacements for like the ID in
Cue 418 — 00:30:27,284 --> 00:30:32,820
the API. So it's not very easy, but it tries, it tries hard and sometimes it gets it
Cue 419 — 00:30:32,860 --> 00:30:35,220
right. And if it doesn't get it right, it will not show it.
Cue 420 — 00:30:35,260 --> 00:30:38,580
So then you will have to use AI to enrich it.
Cue 421 — 00:30:38,620 --> 00:30:40,300
And you can always add stuff with AI.
Cue 422 — 00:30:40,340 --> 00:30:42,280
And this is not going to break things.
Cue 423 — 00:30:43,260 --> 00:30:48,660
So yeah out of time. So I have stickers of Backlog.md and Groma.
Cue 424 — 00:30:48,700 --> 00:30:53,260
So before you leave go grab some and find me outside for questions.
Cue 425 — 00:30:53,300 --> 00:30:55,520
And everyone give Alex another round of applause.
.tessl-plugin
skills
talk-arcolano-autonomy-and-trust
talk-aroussi-decision-context
talk-bevin-agent-assisted-instrument-design
talk-boeckeler-context-lifecycles
talk-braun-context-graphs
talk-carmont-evolving-agent-harnesses
talk-chanezon-gordon-agent-experience-evaluation
talk-cox-secure-agentic-workflows
talk-cunniffe-reusable-agent-platform
talk-davis-agent-tool-policies
talk-debois-concurrent-delivery
talk-dzialets-team-automation
talk-fahmy-harness-api-design
talk-finster-continuous-delivery
talk-fowler-regenerative-software
talk-gavrilescu-factory-architecture
talk-graciolli-amin-verification-aware-programming
talk-gruenbaum-reverse-graph-engineering
talk-hassan-agentic-engineering
talk-hossain-enterprise-agent-adoption
talk-huntley-patch-ready-factories
talk-jain-verifying-intent
talk-katsioloudes-montei-agentic-lifecycle
talk-knox-harness-engineering
talk-kovac-deterministic-memory
talk-lee-perkins-code-readiness
talk-lorenc-machine-speed-security
talk-lubken-software-of-one
talk-meijer-software-specialization
talk-meppiel-versioned-agent-contracts
talk-moss-concurrent-agent-risk
talk-myers-understanding-agent-generated-code
talk-ocytko-scaling-agentic-engineering
talk-oza-automating-design-quality
talk-panzica-agentic-sdlc
talk-pari-patel-agent-awareness
talk-podjarny-trustworthy-software-factories
talk-poulter-factory-improvement
talk-pratik-patel-refined-rag-data
talk-re-ferre-development-platforms
talk-sadogursky-factory-of-factories
talk-shah-training-agent-skills
talk-sharma-mohapatra-evolving-intent
talk-singhal-operating-agents-at-scale
talk-stack-incident-dossiers
talk-tyler-team-coding-benchmarks
talk-wotherspoon-agent-execution-environments
talk-yegge-ai-powered-factories
talk-zuber-leadership-through-coding