Episode 43: Fast AI. Slow workflow.

Welcome to episode 43 of the AI
infrastructure podcast.
Now, I just want to draw your attention
to the website debase.com/mpodcast,
a domain that I own as a minister. And
the idea is is if uh you miss a podcast
or you want to find what's in the
podcast, you can just dip in. There's a
summary there.
there's
uh a whole transcript.
So don't feel obligated to listen to the
whole thing. You know, it's just a
conversation between two geeks
ultimately and we talk about one thing
and another. So whatever interests you,
you can just use debase.com/mpodcast
as a way of navigating the content that
we produce. Otherwise, please enjoy this
podcast, which I've yet to title and
develop a thumbnail for. And by the way,
if you have any feedback about the
editing, because sometimes I've used AI
to sort of cut out the pauses. This one
I've kept the pauses and the ums and rs.
Let me know what you think, what you
prefer to listen to, and I will make
adjustments accordingly.
All the best. Have a great one.
Can we talk? I actually there was one
topic I wanted to to to to bring up and
and it never got uh got brought up and
it was like two or three months ago when
people were talking about how the cloud
app the moment you installed it creates
this huge reverse tunnel to their
control center basically.
>> Well
in my employer my client they disabled
that RC feature but yeah carry on.
>> So what happened is I never used it. I
think maybe you asked me or like did you
ever tried slash remote in Claude Code?
And I was like
>> see
>> well I've never gotten to it.
>> I've never gotten to it but um my son
got a MacBook and he doesn't know how to
configure it and I told him to use cloud
code and suddenly I start getting phone
notifications on how to set up Roblox on
MacBook and of course I I I stormed to
him. I say why are you selling games on
your MacBook? He's like what are you
talking about? Well, Claude Code
magically suddenly connected the RC by
itself. And then the next thing I know,
I see, you know, I get a notification
about my Claude Code on my laptop where I
never ever enabled remote. Suddenly, my
phone is pinging me like there's an open
question in this particular session. And
I was like, what the hell? You know, I
never enabled it.
>> Yeah. Uh
>> my I think my wife my wife said the same
thing to me cuz I'm using her her pro
account and everything everything's
synced up nicely now like all the cord
>> so convenient.
>> It's so good.
>> Here I am buying surprise gifts for my
wife.
>> This is a
>> She's like she's like watching your chat
like oh I don't like that.
>> This is like why people want to use
Chrome incognito, right? If you wanna if
you want to buy a surprise for your
wife.
>> Yeah.
>> Actually, that that's one thing that I I
was chatting to the CISO yesterday about
is like
um Okay, you call me out if if this is
if this is a ridiculous plan. But I was
saying that we need right now chat logs
from all the employees is super
sensitive like only like one or two
people can view them with permission and
you know and then of course that
permission uh is is uh time limited
what's the expression like it expires
>> but we really need to use the chat logs
to improve our processes.
>> Yeah. Yeah. Yeah. So
>> that's a big initiative in a lot of
organizations. The CISO was saying like,
oh, a lot of people are using this for
private information like how to deal
with redundancies and how to deal with
teammates and things like this. We can't
we can't have it uh we can't have people
using the chat logs to to um because of
that reason, right? Private
>> ridiculous. And I'm not
>> Don't they sign a contract that says I'm
aware that the the the subscription is
from the company and it is being
monitored. I signed it. That's all in
the AI policy. I'm not too sure if
anyone's [ __ ] read read it, but we
also even have like, you know, like
Claude has this uh company announcements
thing and we say like your stuff is
logged. Don't put private information in
here, but people still evidently do it.
And now like again, I'm blocked. I'm
blocked. No, but like what does how does
that impact you? Right. So is there in
in Europe or in UK a law that says that
even if you explicitly make an uh what
is this a disclaimer and the and the
people sign uh sign the disclaimer you
cannot use the disclaimer to to
protect yourself.
>> Yeah. Absolutely right. But there's some
people in my organization um who are
like oh we need to um it's like European
mentality. We need to look after people.
>> We need to look after people.
>> Yeah. Yeah. Of course,
>> we've done mistakes and now we have to
like like but come on,
>> we can
>> and I think DHH has a lot more to say
about how Europe likes to look after
people.
>> Yeah, we we can't help it that uh
h
you want to hear something cool?
>> Oh, open AI uh Astra.
>> Astra. No, that's not cool.
>> I was trying to work.
>> I mean, it is cool, right?
>> Is that Is that a new app? Is that a new
model? I couldn't configure it. Didn't
really understand.
So this is the weird thing because when
OpenAI announced Astra earlier or at
least there was an announcement saying
that they will start releasing Astra
access to a select few organizations
that Astra has proven to be very uh much
better than than even you know uh what's
it called mythos at security um
hardening and able to to find bugs in
software so that they are going to
release it in in like batches.
Um, I've read that like at the beginning
of the week or something and then this
morning you see announcements on Reddit
like 1 hour till Astra and it's like
what are they talking about? And then
finally the announcement is there.
>> Was down I noticed
>> and then they're talking about Astra.
Um, Astra is here. It's amazing. It's
like terminal bench score is double that
of Fable or Myth like Fable 5.1 and
Opus. It's way better. terminal bench is
um a physics uh
>> I saw your LinkedIn post about
benchmarks so I know where this is
going. No, this is not no that's not I
don't talk about that the OpenAI
benchmark. I talk about other benchmark.
I can I can say very I I have a bone to
pick with a couple of people uh on on
the benchmark and they keep hammering it
on LinkedIn um and they keep repeating
um you know something ridiculous which
is the opposite of what I measure but
okay. Uh no in this case OpenAI
benchmark I it's true. I think Astra is
amazing um based on on what I read about
it, but it's limited to some companies
and no like only some people have access
to it. So why do I care?
>> Exactly. I hate this I hate this uh what
do you call it? Exclusivity [ __ ]
>> I mean I mean I don't hate it. It's just
that I wouldn't hype about it like it's
not in my hands unless I get my hands on
it. Sure. Okay. You can tell me it's
coming but it's not there.
>> Were seemed to be all about just
chatting to it. I mean, and and it was
amazingly fast, like sped up video, like
draw a circle and instantly appears. I
mean, in reality, that's
>> Oh, it's like that. You remember
>> you remember that last year when Google
did that video where they talked to
Gemini and Gemini instantly did the
thing? It was all fake.
>> It was all faked.
>> Yeah, it's in that level, isn't it?
Yeah. But anyway, when you want to talk
So, there's We can talk five minutes
about benchmarks. Um, no, I know
benchmarks is crazy. Okay. I I need to
>> No, I'll just say one thing. I'll just I
have to bump pick. So, you're going to
have You rent it to me. Now, I get to
rent. You triggered it. You
>> benchmarks are so boring.
>> You can do anything with a benchmark.
Like my my statistics professor used to
say, according to statistics, every
human on earth have one testicle. You
know, on average, sorry, on average, uh
the humans on Earth have one testicle
according to statistics. Right. Um,
sorry. I I totally botched that one.
Anyway,
>> on how I actually I'm struggling to
understand how that's possible, but
>> on average
>> because the average is rounded down.
>> No, one testicle because female and male
50/50 about
>> I I was thinking just men. Okay, that
makes sense.
>> No, I said every human on average have
one testicle.
>> Yeah. Yeah. I mean I the stats is always
a thing. Like I think my wife triggered
me yesterday saying like, "Oh,
it's it's like when you drive a car,
you're going to have like one in a
thousand chance of getting killed or
something like this per mile or
something, whatever." And I said, "Yeah,
but that's an average. Like we're in a
we're in a really safe car, you know?
We're not in it's like you can't really
think that way." Uh anyway,
>> yeah,
>> I get I get what you're saying about
benchmarks.
I get what you're saying.
>> The Okay, what I wanted to talk you to
you about in all seriousness was
security stuff again because I'm I'm in
the security team
>> and security
>> and you you like me have probably
wrangled with uh I mean infrastructure
is more than just deploying stuff with
Terraform. It's it's also about making
sure that what we ship is secure and
we're not and then we're not making our
infrastructure v vulnerable to attack.
And a big part of that has always been
having doing CI/CD work where you maybe
throw in um
uh a SAS, you know, source source
checker, throwing in uh some linting
rules, whatever. Throwing in uh
dependabot rules,
maybe as going as so far as auto
patching stuff. I don't know how how you
feel about that stuff,
>> but that's pretty advanced stuff. like
Crowd Strike.
>> I I I don't know what Crowd Strike has
to offer there. Does it auto patch
stuff?
>> I think two years ago, all the major
airports in the world were down because
of a Crowd Strike release.
>> Oh. Oh, yeah. You're right. You're
right.
>> An autodeployed patch.
>> Years ago, that was when I was in
Singapore, I think.
>> Yeah. So the point I'm making here is
that I am very strong uh advocate of
immutable deploys and promotion of
images across trust. So in terms of
patching like when you say auto patching
to me you need to patch a machine and
roll it out like you don't have
you need to test it and then once it
passes your gates then it drops into
>> anyway. So, but yeah, I mean you can you
can you can do that uh but not live,
right? Autopatching live. No. Um canal
releases, immutable deploys, rollouts or
patches,
>> all that stuff. That's all part of the
stuff that we've been doing for years,
I'm sure.
>> Okay. Yeah, but airports didn't do it.
>> Now,
>> or Crowd Strike didn't do it. The
question was put to me was that oh we we
have some sast we have some dast and we
get a lot of alerts nonetheless about
certain vulnerabilities and our security
team is overloaded with these sort of
this sort of toil like keeping these
things patched chasing down teams to
make sure the things is all done.
>> Yeah.
>> Kai tell me tell me that AI agents can
do this and do it better. And I was
like, "Yeah, that that sounds like it
has potential for AI agents to spawn and
do things better."
>> Um, and actually it made me even think
to myself like there must be a point
where where deterministic sast and
dasted or whatever tools are probably
going to be eclipsed by AI. Maybe not
today. I don't know if you have any
feelings about that but surely like
you know because because sometimes we
have like really archaic you know sass
[ __ ]
to be honest I think you need
determinist I think AI can be a part of
generating the rule but I don't see the
point of spending tokens to apply rules
AI can help evaluate and deploy and
propagate deterministic rules
deterministic rules must be there and
they can you can you can have a fast and
a CVE database and you can do an
evaluation across your fleet and and and
and determine the impact which today
might you know take time to accept the
risk on um on on on some findings and
set a timeline and I think that's where
AI can can help you um you know make
that decision.
>> I'm basically I'm not saying that um AI
is going to replace the SAS the
deterministic procedures. I'm just I'm
just thinking
>> literally what you said. Okay, whatever.
I'm just trying to say that like AI is
going to surely eclipse them in in their
potency and effectiveness as a security
tool.
>> I mean, they're going to amplify it. And
maybe that ties into again AI is an
amplifier. Um AI is amplifying good
practices and that includes you're not
going to get better. like if you have a
complete manual system where humans are
making are the judgment um factor and
you're slow down um to to evaluate
everything. I think everywhere that's
the case just replacing a human with an
AI agent is maybe going to give you some
speed but it's also going to make you
very like scared about trusting the
agent is going to make the right call.
This probabilistic thing um you know
makes mistakes like humans. Um, so you
are not going to speed up unless you
have automated gates and automation in
place and that ties directly into like
recent presentations by uh Uber about
the AISDLC that you also talked about
before. Um, but another person that did
a very interesting talk was um I forgot
the name, Claire. She is part of the
Kiro team at AWS and that's why I
follow her on LinkedIn. She also has an
AI engineering talk. um and she talks
about how AWS evolved individual teams
um adopting AI and basically the metrics
and measurements that they found out
matter like it's not about
>> it's a really it's a really good talk
because she talks about like this team
adopted AI this is where they were um
this is how how much it accelerated them
and this is why they were not a
reasonable um metric for
because it was an isolated team of
highly experienced people with a very
green field project. So they could not
use those results and then they have
like two or three pilot projects like
that and then each time she extracts the
information that they measured and then
she identified the real like um rules
that you must follow um to adopt or to
to become faster with AI which is like
you have to go slow before you go fast.
You have to put in place the right
automation and guardrails. Uber talks
about refactoring their codebase into
monor repos, adopting basil, setting in
place all of the things that makes an
agent efficient. Executing guard rails
is an important one.
>> So, and and you know what? Um, so in
terms of AISDL,
>> that's exactly what I'm trying to do
right now at work, right? I'm trying to
really put in place like get rid of all
of these manual human, you know, control
points because putting agents in place
is not going to solve it. You have to
refactor and rethink about how you do
the AIS DLC.
>> Yeah, exactly.
>> That fits the agent workflow. I'm going
to find the other things that she
brought up.
>> Where is
>> How did that Martinelli guy factor in?
But did she is he a consultant to AWS or
something? Um so there was a Kiro IDE
workshop that I participated in back in
uh January or something like that and
they then he talked about AI up which is
his consultancy where he talks about the
coordination like he's he's mainly
focused on like a lot of these SDDD like
specdriven development and and keto IDE
included is all about single player like
as if the developer or the engineer
alone is making those decisions and a
lot of the actual um requirement are
about finding the stakeholders ers,
putting them in the room together to
actually make the decision. Like the
expertise lives across multiple people
that all need to be coordinated. I think
you talked about that before as well,
right?
Hello. Oh, yeah. I I
I got a Oh, I I actually got permission
to set up a new Adeps account with so
that I could get better versed with
Bedrock and I got a bit of a budget to
spend. So,
>> Bedrock or Agent Core? What do you what
do you mean?
>> Uh I think I mean I've got I don't know
100 bucks a month or whatever to spend
on on AWS.
>> Better. Okay.
>> Here I found I found the thread.
>> So that means I can start using Kro and
Ango and all that sort of stuff I
suppose. Right.
>> So the name I can send you to YouTube
and we can share it. The name is um
Claire Legori at AWS.
>> Oh, I know that lady. I mean I don't
I've seen her stuff online. So she talks
about the different
>> uh pilots that they did and what
measurements they got but why they were
not not um not right and then what it
really is like the things that they
want.
>> That's cool. That's cool.
I mean a big one of course is making
sure that the agents get the right
context which is ontology knowledge
graphs knowledge systems
like one problem I have at work that
really frustrated me that I feel like we
need to measure is that everything needs
okay I'm in a company right now where
everything has to be approved okay I can
I can understand that I am [ __ ]
improving PRs left right and center
awesome but then we don't have automerge
enabled So, it's up to uh other people
to to hit the the merge button, right? I
guess
>> you're on GitHub.
>> Yeah. And then, of course, there's some
there's some ambiguity there, like who's
going to hit the merge button? Like,
dude, it's your patch.
>> You're supposed to hit it. Yeah. The
moment you got your approvals
>> and like there's people that are
diddling about that and and like
those sort of things I want to AI to
chase for me. Like go and
>> Yeah. That's exactly the thing that the
PC that I built.
>> So the PAC I built is a nudging system.
So it's um it's a durable workflow that
helps you carry it was an idea from from
the team that I'm in um to carry your PR
because they have so many manual
approvals. They need two approvals. One
needs to be from a default reviewer.
Then there's the whole uh PR branch uh
validation that takes um you know maybe
20 minutes to run. And so so you you you
cannot and every commit you push u
invalidates of course the approvals. So
you you can't like chase for approvals.
>> There is an option in well there is an
option in GitHub that doesn't remove the
the approval even if you change it
subsequently.
>> Yeah. But like that's that that that's
not allowed in in this in a in a
>> Yeah. Well, I can't think a couple of
things in GitHub could just make a lot
of problems go away like
>> Bitbucket.
>> Oh [ __ ] Oh god.
>> Yeah. Anyway, um
>> what I was wanted to say, oh yeah, so we
have the nudge like one thing that I
built is a nudge uh system with like
basically it's um it's a step functions
workflow, so it's durable and it wakes
up the you know if it hasn't received
the approvals. Um it's there's no tokens
involved in there, but there's like a
timer and it it uh it reminds people um
on on an interval and everything's
configurable. So that that's kind of
like one of the cool things that that I
built which is like around the agents um
a whole workflow system like what people
use N8 and for and temporal but we want
to be without having to deploy another
workflow system. Yeah.
>> Yeah. And does it what's the medium?
Slack I guess or something.
>> Yeah. So so the nut system I built is is
Slack and um you know uh yeah it's it's
Slack and reaction and and interactive
cards and and all of that. That was that
was that was all really fun. I built
that like in in a month and uh
>> a month
you going to say like an hour.
>> You have no idea.
>> You have no idea if you are in aization
where you can't even get a web hook set
up. Now the Slack app I have to admit
was very fast but to get a bit bucket
web hook that is complicated.
>> Okay. And then you have to do you have
to scope it. you have to avoid uh you
have to make sure it goes to the right
channels um that is allowed and
associated with the team. Uh so there's
a whole bunch of additional stuff but
that's not even the parts that I that I
was um I was I was stuck on.
>> Yeah, sure you can build a something
real stupid on a on your own slack with
your own GitHub in in in five minutes
Kai. But to do it in an organization
>> okay listen I need I need a break. I'm
going to run away. There's one more
thing about CI because you said like um
aentic CI and and I'm been breaking my
head about the uh swamp guys there
saying that you know why are we doing
why are we running um validations in
CI/CD that the agents are running they
they they must run all of these checks
right this is the feedback loop and they
said you don't need complicated CI/CD
systems anymore as long as Whatever the
agent produces can be signed and like
has an attest um how you say
attestification and basically it
>> yeah it must produce an artifact that
you can't get unless you've run the test
suite and then you've you know you
signed it and then you push that with
your with your code change and so the
CI/CD system doesn't run the test it
just checks that the um attestation is
there yes the test has been run and and
and basically that's the gate So it
doesn't that that completely shifts the
um CIC CI/CD system to be because
they're talking about how GitHub is
dying right with all of these PRs being
created overloading all of these
addition.
>> Yeah. Like I can't help but think CI/CD
needs to be reinvented because because
yeah, we we're getting lots of changes
in and uh and running like one hour test
suite
for each change is [ __ ] crazy, right?
Surely agents can do a better job of
that. I have I have my reservations
because so I wanted to post a comment
and I didn't read his full blog post so
maybe he addressed it so I I didn't say
the comment. Maybe I should read the
post. But for me, a CI/CD system
is proof that whatever you have on your
machine works on
>> exactly that
>> another machine.
>> That's a good way of putting it like
it's not not my machine problem or
you've said it before works. It's not a
works on my machine problem anymore.
>> Yeah. If I if I I have set up a repo and
my CI/CD is green and that means I if
somebody comes to me and says it doesn't
work on my machine I said well there
right there is a complete automated
setup of a a complete fresh machine that
is provided by GitHub all of the steps
are there I mean of course you could
build your self-hosted runner with all
of your in software installed but even
then you will have a definition of that
runner with all the software it requires
so if you're not able to like do that
there's There's an automated setup
script and a fully automated test suite
that validates that it works not just on
my machine but also on that machine and
and and if you're using agents and
they're running all the tests and
they're signing an attestation that they
have done so and then submit the patch
and then the CI/CD system doesn't
validate um the actual test run but just
validates that you know yes the
attestation is val valid and um
therefore the the tests uh the the patch
can be merged. Um
it makes me wonder.
So I I don't quite understand this. So
is it is it is Claude making a claim
that the test ran in its own environment
and then and then there's a test that
and then this attestation means that it
doesn't need to run.
>> Yeah, I I haven't read the full blog
post. So the way I understand it is
>> double guessing what it says. the I I
think I I've read the summaries and I've
I think started reading it, but um I
just, you know, my my head was just
racing on some of the uh thoughts and I
just in my head the way it works, maybe
I'm wrong, uh maybe he addresses all of
my concerns and maybe I should go back.
So Paul's tech from um from Swamp.
>> Okay.
>> Um says that the way I understand it,
you have a what they also do with swamp
is they have a they built something
weird that they said is the future of
how you build software. But to be
honest, Claude Code doesn't need any of
that. What what they build you you you
build your software um suite with with
your unit test. And I mean obviously
when you run unit test you have a
coverage report, you have all kinds of
additional reports that that gets
generated, right? And if you then sign
it um
and make sure that there's no way for
the agent to mock it and fake it, maybe
maybe it calls out to a separate service
to I don't know how but it's certified.
Let's let's assume that there is a
system that the agent cannot game that
ensures that if this if the agent has
run
>> the certification means that it's done
some due diligence before. Yes, it it
it's he has ran he has tested his code
and in the comments people were going
well wasn't like in our company we used
to have this social rule that you
shouldn't be wasting CI minutes the rule
was that you need to test your code and
if you submit a PR um yeah
>> I get that I get that
>> because CI minutes were expensive so we
would not um you know we would we would
uh actually as a social rule within our
company 30 30 engineers committing code
um we would expect people to sell like
you know honors uh you know say that
like I've I've built something and I've
ran all the tests.
>> Exactly.
>> Um
>> yeah so CI/CD systems today are
sometimes
not really possible to run locally right
and and and sometimes
GitHub is a great example. Yeah, because
of the way that the GitHub actions YAML
work and all that it and and then you
have Dagger and all the solutions that
are trying to make it so that you can
run it anywhere no matter your machine.
>> Yeah.
>> Uh but but but
another thing that I feel like like when
I when I use cloud I have Mi and Mi
allows me to run like for example I
maintain CDK terrain and we support
Terraform 1.7 all the way up to 1.16.
>> Are you using Mi? I only just learned
about Mi the other day.
all your your CLIs and whatever.
>> It's really useful. It's it's like when
you used to use uh TFN like terraform or
node version manager, you have so many
like for every single tool that you use
and for Python you then use uh VN
virtual amps.
>> Exactly. Exactly. Now this one
>> there was a real gap. There was a real
gap.
>> Yeah. Mises cross everything. So it's a
bit like I I mean people that use Nyx
say like yeah but I heard that with Nyx.
What are you talking about? I had it
with next 20 years ago.
And then I I I I provide a workshop with
uh with me and I say this is what meat
does. And then somebody [ __ ] they don't
they're never going to watch this. They
said like oh let me put this into Nick.
And you know what happened for a 4hour
workshop. He didn't do any exercise. He
spent four hours trying to to to set up
the sample repo into Nick. And initially
I entertained him, but when it was like
half the workshop done, I was like,
"Yeah, you kind of missed the workshop
though.
>> Stop. Please stop.
>> Shit."
>> Anyway, Mi is awesome. Mi allows you to
to run your test matrix.
>> Yeah, I got you. Oh, yeah. So, you got
your test matrix
>> and it's across versions, which I don't
think you usually can. See, Mi probably
solves the crazy GitHub actions YAML,
doesn't it? Because everything's set up.
Mis is a way of setting up your
developer environment, right? So, I had
this PR uh for CDK terrain that fixed um
some
some change of behavior that was
introduced with Terraform version 1.15.
So, on 1.14 it worked. on 1.15 it
stopped working because they changed
some behavior and somebody submitted a
pull request based off of an issue
report. Um and then the Hermes bot ident
read the code and said that it fixes A
but it doesn't fix B. And I was like is
B new and and then and then it says no B
B also worked the same way with the old
behavior. Um and I said okay can you do
a full test suite meaning you run
version 1.14 A B validate that both of
them work then
>> run terapform 1.15 on the PR head um
confirm that like across across master
like latest release didn't have to be
master uh latest stable release versus
1.14 1.15 and then PR head versus 1.15
1.14
plus A versus B. Um, and and it
basically build up a whole test harness
with everything with MIS. I told it use
MIS because it has
>> But you're going to get a permutation
explosion, aren't you? You're going to
be testing different versions with
different
>> versions. And we do. And the problem is
that we have the CI/CD workflow that
that has 350 tests uh that we inherited
and used to take 5 hours to run, but um
thanks to community contributors,
somebody completely uh spent like two or
3 weeks and he got it down to under an
hour. We're still running the full test
suite. He just rearranged it. He just
put in early exit conditions and all
that. So, he really really made it much
better. Like the PRs run way faster. The
the merge to master runs way faster.
>> I love those changes.
>> Yeah. Well, yeah, it's a really nice
thing to do until you you have like 355
jobs to look at. And I'm like, "Oh,
man." And he just he just kept hammering
at it, you know. He and every PR this
was beautiful. This also he was using
Clark, but every PR of him was just like
just 20 lines, 30 lines stiff, very easy
to review, very small, very focused.
>> That's really not easy with bloody opus
in the [ __ ] mix.
>> No, I mean it was also three months ago.
Uh but the the the the question is um I
want to know how exactly he did it
because then I I had like a whole bunch
of like markdown to control the uh spec
driven development and he asked me like
can you remove all the markdown because
it impacts my my uh my my workflow my
cloud workflow uh and I have
>> this this is a a guy in your in your
employee in your company that you work
>> no no no no this is the open source
project so this is a guy that came along
asks very very politely
Guys, do you mind if I spend some time
optimizing DCI? We're like,
>> okay.
>> Okay, send me the links, man. I'm I I
want to I want to celebrate this guy. I
want to learn from this guy myself.
>> Yeah, he did it. And you could tell it
was cloud, but he controlled it so well.
And um you know, compare his PRs to my
PRs. I'm embarrassed. Uh
>> I I was like, god damn, I can do so much
better than what I'm doing right now.
But
anyway,
>> yeah, I I I I need to work out there's a
few things on my to-do list. I I want
when Claude makes a PR, I want it to be
draft because by default it makes it an
open PR and then my colleagues are going
like, what is this? And like I'm working
on it.
>> Yeah.
>> And uh and it has the most insane commit
messages nowadays. It's like more co
more more [ __ ] in the commit message
than there was in the diff and like
[ __ ] [ __ ] up, man.
>> I don't know. Something happened right
right now. It uses comments as journals.
It puts comment It puts diffs that are
massive. I mean, it puts comments that
are massive.
>> I think everyone's complaining.
>> I need I need a way. You you know I have
this lock sensor to limit the lines to
code but it doesn't catch the crazy
messages and the commit and the commit
messages. Okay, I just added a new rule
to my u my my fable work like main
agents and it follows it has like the
operand uh modus operandi rules are that
and and I haven't put this in steering
rules but like I need to mine and dream
like you just said but any what it what
it has right now is that uh medium tasks
are background opus agents and I never
hit the limit. So dynamic workflows with
opus implementers, opus verifiers and um
only very small patch fixes it does
itself. And the new rule is you run a
background opus comment um um fixer that
basically has a few rules like comments
are not journals. The the history lives
in the comets. So you do not put like
the full history of why a line of code
turns out to be that line of code by
saying on on uh September 25 we
identified that it didn't work. So we
tried and this and then we tried that
and now we are this is why this this
value this line right here and you have
like 10 lines and I said no more than
three lines like maximum five
>> comments.
I need to fix this and I'm going to fix
this [ __ ] right now because this is
the problem that's bothering me right
now besides needing the toilet. Okay,
thanks Vincent for the for the chat. I
don't know if this will be a