Episode 34: Claude Dynamic Workflows 🤑

Morning.
>> Good morning.
>> Yeah, I'm I'm back at work after a week
holiday from Oh, you got You grew a
beard.
Oh, yeah, again.
>> uh no, I just I'm I haven't shaved for a
while.
>> Since you've been
uh
I don't know. You
You didn't giving me a lot of AI
updates, so I guess you're you're deep
in it.
>> You were You were like on holiday, I'm
addicted, and so I I started to No, you
said like let's let AI get get like some
rest. So, I started to to be like,
"Well, haven't you seen the news?
There's so much happening."
>> [laughter]
>> So, what what did I miss? I was hoping
you could brief me before I go to this
AI conference in London.
>> Yes, so
>> What's the hardest thing?
>> The hardest thing, to be honest, is the
um is the actual dynamic workflows in
Claude Code. They seem to work.
>> Okay, so
I've been using workflows I mean, I've
been wanting workflows, and I guess I've
been getting my workflows
Basically, I was using that NA N8
N
I don't know how to pronounce the
project's name.
I using NA N for a little bit, but then
I ended up just converting my
automations to system D.
But,
what do you use Cloud
workflows for?
>> Okay, um
>> there? What I
>> Can you hear me?
>> Yeah, yeah.
>> Yes. Yeah, sorry. I got a message and I
got distracted. So, um basically, what
happened
um
they launched Opus 4.8 on I think
Vietnam it was Friday morning.
And with it, in Claude Code, they
released dynamic workflows, which is a
tool that can execute uh JavaScript file
that um
the JavaScript script
>> that on a schedule, all right, I'm
assuming.
>> No, no, not on a schedule. Dynamic
determinate dynamic workflows
is lets the agent take your context then
build a graph of agents and each agent's
prompt and then
execute it. So, the tool just runs
a deterministic workflow that gets
dynamically generated depending on the
tasks at hand. So, I with the spec
driven development workflow, like for
example, if if we put this in terms of
spec driven development workflow and
our, you know, we are familiar with Open
Spec or Spec It, specifically Spec It,
they you have the specify phase where
you have the user stories, then you have
the
planning and research phase where you do
some tech spikes, you identify exactly
what is the libraries you're going to
use, what's the tech stack, how is the
technical implementation, and then
usually you generate
a task graph. Like first we have this
phase of like foundational setup and
then we have another phase of
implementing the the necessary
functionality modules required for this
first user story, and then we have the
second user story, and then we have
priorities and things that we can do in
parallel and all that, right? And and
then we go into an implementation phase
which goes through this preset
you know, graph of tasks to execute, and
you basically have a prompt that loads
it, right? So, you can get rid of that
task breakdown and implementation step
completely now. You can just go straight
Yeah, you don't need to do the task
breakdown anymore because you're using
Opus. Before
you would do the task breakdown and you
would review the task and you say, "Yes,
I agree. User story one is this module
that needs to be built first. The
priority is correct. We can skip the
second part because that part is like
nice to have." So, you would have this
like ability to decide what you want or
what you don't want on that on that
implementation like multi multi phase
plan. Um so I guess if you still want
that, you you still get that because the
the dynamic workflow will present to you
exactly the phases and the the the the
workflow that's going to get done.
>> controls Does it also generate the tasks
and the order of the tasks, I guess,
right?
>> Yes, it it it So in my case, because I
give it I use it with Spec kit, it
creates phase one tasks like phase two
tasks. And each task is basically an
agent that goes and does does a task. So
in the past I was trying to run parallel
agents that didn't work. But now today
and a lot of people like working with
Git work
Git work trees or like using agent mail
or agent communication to lock files. If
you don't use work trees, that was all
things that people were trying to do it
on a single box to go and build with
multiple agents and then have the
coordination and all of that, right? Now
with dynamic workflows, it's all done
for you.
The the the master agent does
>> The The This is a cloud-only feature,
right? So what are you going to support
>> Yeah, and if you're on enterprise, you
need to ask your admin to enable it
because it by default dynamic workflows
is disabled. It The main The main feed
feedback
from most people was that this is
extremely expensive. Like in 20 minutes,
it burns the equivalent of what would
normally take me 4 hours to
to to use because it spins up like 12
agents to do the work. And the orchestra
And then also merge agents. But you
don't need to use it just for that,
right? I also do it for like
adversarial reviews.
And
I saw this very funny post from Antal
Babenko about like you need three review
agents to to come to a consensus. So you
always need to spin up three of them and
I was like Yeah, I mean I I I'm okay
with one or two.
Um but it's funny because you can you
ask new ways to use agents.
Well, it's more like
the way Yeah, the way that you were
using agents or that you everyone was
building like their their own agent
orchestration tooling.
It's kind of like all been taken care of
by this workflows solution. It's like I
found it amazing because the workflows
are generated and then they are
deterministic. So you can save them.
It's almost like swamp but built into
cloud, right? It has its own little
schema of of building a
workflow and then you can save those
workflows and you can reuse them and you
can share them and and you can generate
adversarial. So it will spin up three
agents to do an adversarial review and
then it will merge all of their findings
into a single report.
They find different things. I like I
have three Opus 4.8 review agents
running with a with a blind context
validating the the constitution,
validating the
the user stories and the definition of
tons that are derived from the
acceptance criteria.
To me they're like pretty much the same
thing and on or from the functional
requirements you you get that.
>> Obviously it works well cuz I think I
can sense your excitement.
You're getting stuff done or
>> Yes,
I built quite a few agents.
>> But the quality work.
>> It is probably because these are Opus
4.8 like high effort agents. You do need
to be careful like make sure that you
you confirm the model and you confirm
the effort like by default it will
inherit. So So if you're doing an
orchestration and at like 4.8 extra high
effort and then you you trigger a
workflow, it's got if you don't tell the
orchestrator to create a workflow where
the model is defined for each workflow
step, it's going to inherit the model
and the effort and then it's going to
burn like crazy. So you have to be
mindful of that.
>> So you're you're on the max plan. Have
you managed to exhaust your tokens?
>> I'm I'm getting there. Like I was
it was resetting in 5 days and now I'm
actually starting like on I'm on on the
marker. I use this codex bar which gives
me an indicator on on how much I I
should be at if I I want to
I'm exhausted.
>> Basically your your life is basically
um
proportional to how much your token
spend is is like you slow down when it
when you come to the end. Or anyway, I'm
getting
Oh.
>> Sorry, um hello.
Hello.
>> Yeah, you have you got you got some
people around you?
>> Yeah, my my bike front wheel was had had
a leak.
Yeah, no like I started to trust it a
lot more. I feel like it's way more
automated now. I can because you use a
dynamic workflow, you set it to auto
mode, you go for I went for breakfast
and I see I I check my GitHub and I see
comments happening like as I was eating.
The auto mode just keeps going. Um I
have a I have a few really really nice
stories about auto mode like situations
where I was very impressed by by it it's
>> Yeah, yeah, I'm using auto mode. I I I
had like a an alias for super cloud
dangerously skip permissions but I never
use it anymore. On auto auto does the
trick.
So yeah, it's
So what does it do that impressed you?
It gives you confidence. It's all like
it tells you like I stopped this.
>> Exactly like
I mean, the thing is it does it blocks
certain actions and then the the LLM
tries to work around them and then auto
mode like blocks them or the LLM the the
auto mode gives it a good reason and
then the LLM ask me but sometimes I it's
been trying and it's like it tries three
four times to circumvent auto mode and
then it's it stops it says like yeah,
I'm I'm not able to do this and it went
really far. It went like way off the
mark like what I what we went off and
off and I was like no, no, no, stop.
Your first action was correct." And the
correct thing it needs to do is it needs
to pause, and I need to say it
explicitly. For example,
um I was doing work on a development AWS
account, and I asked or I said, "And
while you're at it, check if staging is
also aligned or would also face this
issue." And so, while I was doing all
the work, a little bit, you know, going
down the context, it finally came to the
point, "Let me go check staging." And
Auto Mode says, "You can't. It's blocked
because this is like um exploration past
your scope. You're not You're You're
working on dev. You're not supposed to
go into other accounts." And and basic
Yeah, and it stopped At that time, it
stopped really nicely, told me, "I've
been blocked because I'm going into
staging." And it only allows it if I
explicitly say, "Now go to staging and
check." And then Auto Mode will say,
"Okay, you're allowed to go." So, so
that's uh one of the stories.
Um
>> Mhm.
>> There was another one.
>> But but but but it sounds like you're
totally Claude red-pilled right now. I
mean, it Do you ever use Codex and the
the other
>> I'm actually setting up a Hermes agent.
I mean, my son is on summer holiday, so
I'm asking him to set up Hermes for his
exploration, and we're using our chat
GPT Codex
uh what for this, yes. Cuz otherwise,
API cost on Anthropic would be insane.
And in terms of like, do I try other
tools? I'm still not hitting my budget
with with Claude code, so I have no
incentive. Like, if I do, and I still
have budget, like I used to have budget
on Gemini CLI, but apparently that's
been completely locked down, and the
cost has gone up. And Codex is probably
the only other one that I would use, and
I'm trying to get use out of it with
Hermes right now.
>> One of the last things I sent you or was
trying to tell you about was I tried
anti-gravity, too, and it was
>> Yes.
>> not a great experience.
>> Well, your your Your issue was that it
that it didn't give you any any warning
that it was going to run out of tokens,
and it just stopped in the middle of the
work and left your your work a little
bit unfinished, which was frustrating.
Um but you should have committed the
working versions in between.
>> [laughter]
>> Well.
>> another cool version uh story with Auto
mode I just remembered. I was working
and I created a public repo and I was
working locally on on code for quite a
while and I told it to push the remote
to the remote and uh like do the do a
bunch of work cuz usually it's like Auto
mode doesn't see the original request
included my approval to go push. Um or
it was just being extra careful. Anyway,
I told it to go push and and Auto mode
blocked it uh saying that this is a
public repo. You must be aware of the
risk.
>> Okay.
>> whatever you going to post is going to
be public and um you need to approve
that. Like are you aware and approve it?
Only then. Yeah. Yeah, that's like
>> Okay, getting Okay, let's talk about uh
Git commits here.
As you as you were saying, how do you
instruct the you you you you said for
example in my anti-gravity two uh
session, I should have been committing
as I went. So, how do you think I should
have done that? I should have just
just uh at at at the phases dropped into
a shell
and uh committed the code myself. What
what do you What What was What was I
supposed to do there?
>> I don't like to commit myself because
then I have to come up with the commit
message and I find the model is doing it
way better. So, intermediately I I when
something works, I ask it let's commit.
And and it depends on your prompt, but I
find Opus is very very um you know,
maybe anti-gravity doesn't do that, but
Opus is like, "Hey, we finished a a
decent amount of work. Let's commit."
Like it's pushing me. It's like, "Let's
commit. It's working. Let's commit." And
I'm like, "Let's do one more." I mean,
it's like we have all the other stuff to
do, but we're at a decent point where we
should like be able to commit. And it
says, "Do you want me to commit and then
continue or do you want to just continue
and then do a bigger commit?"
>> So, what what Claude is doing that by
default for you then?
>> I guess it's a little bit related to the
prompt and the type of work. With Spec
Ledger now, even I have like branches
and they're like 20 commits and I didn't
never ask it to commit. It's on auto
mode. It has like things to do and it
just go maybe one time when it's like
generates the spec, I say like push and
commit for team alignment alignment. I
always want to like I want to build into
the model the concept that we have to
get alignment from other people even
though it's just me.
>> increasingly that I need it it sounds
increasingly that I I have to use Claude
and nothing else because Claude does get
commits right and get and Claude does
commissioning right and Claude does
>> It's just I I
I I have I have a budget on it that I
paid for it so I'm using it and I'm I'm
happy with it. Um, I'm excited to try
out Pilot Dev and to rebuild it. I'm
also using open code with Git GitHub
Copilot's seats for like just for demos
just to show to people that hey, this is
how you can use this and that but like
my main driver is mainly Claude Code,
yeah. Cuz I have the budget for it,
yeah.
>> We should rename the podcast the Claude
AI infrastructure.
>> Because this is always always we're
always looking at the different
trade-offs and like a lot of people are
complaining uh that Opus 4.7 was really
missing the mark and Codex give you so
much more for
you know, to cost value.
>> So tell me about 4.8. What's your
experience?
Between 4.8 and 4.7.
I guess you just prefer the larger
model.
>> Yeah, I mean I don't really notice. Um,
I'm not going I don't have a benchmark.
I don't do the same task to see how it
does better or anything like that. I
just want things to get done and it's
getting things done and I feel 4.8 plus
auto mode plus um the new workflow
feature with the budget on max is
getting done things done for me and and
I'm very happy like I was able to turn
through quite a lot of things. I'm still
very
careful because I think the biggest
issue
is that you and this was an interesting
conversation I had with my friend
yesterday is that
you still need to have a way to
define what needs to be validated
against, right? So, either you are very
explicit in your plan or you're you're
specifying your user stories and then
you can get a high confidence that you
get what you want. Then you can have
multiple adversarial reviews.
>> was That was my next question. So,
you're you've adapted your spec hit
workflow for
leveraging Claude, right? You're You're
using that
workflow to get your work done.
Right? Or not?
>> the moment I have not up streamed it
into the binary. It's It's a It's a
custom prompt in my current project. I
do want to take those updated templates
and put them They're actually They're
the same. They're They're called
verify-workflow and implement-workflow
and they're They have adjustments saying
that we do not need the task to be
broken down because we will break down
the tasks. So, so there's a difference
between the normal task and implement
and now there's a just one
implement-workflow that does both. So,
they're separate. I could keep them
separate and make an a little annotation
that this only works with the Claude
>> remember your your spec hit has a has a
UI. Do you actually use the UI to like
track
the implementation?
>> tried so many. So, the problem of
collaborative reviews or even just
getting like a nice markdown web view of
the of the specification, the user
stories, the ability to highlight a
section, leave a comment on that
section. Uh that's obviously a hot
topic. There's a hundreds of tools that
do this. You shared with me Trit.
>> Yeah, well, why didn't it work for you?
It works for me. I'm I'm able to look at
it differently.
Put comments in and then it goes back
into the agent and
>> Okay, but that's not what I use it for,
right? I I use it for single artifact. I
don't know if you tried that. But I
I tried
Crit has two modes. Shows how much you
use it. Crit has has a
a diff mode with multiple files or has
an a single artifact mode.
>> Okay. I haven't used that one. So, yeah.
>> Yes, it doesn't work. And also the
folder in like 2 weeks the the source
code folder became littered with like I
don't know, 20 go source code files in
the in the root. The web app is like I
don't I don't it's it's it's completely
sloppy fine in just 2 weeks. I I was
really surprised looking at the code
after 2 weeks and I was I had no
confidence. If you cannot as an
architect tell the agents to build the
code in like a well like organized way,
then I'm already have a lot less
confidence. Like, you know how people
share their project on Reddit and they
go like I built this with with AI or
like look at my project and then the
first comment is like I looked at the
source code. It's It's absolute slop.
And I think
I mean, it's it's a stupid criticism
because I think if it works, it works.
But still,
you need to be able to instruct the
models to do it properly.
>> Okay, single single artifact didn't
work. Did you Did you try the diff mode
at least?
>> No, no, no. But that's not my target. My
My target is to review the user stories,
highlight This is very very like This is
the hardest part of the job right now
for me, which is like
do these user stories make make sense?
Like for example, I'm building a tool
that does a a verify, which is a content
sha verification on on skills to verify
that the skills haven't been maliciously
manipulated. And and I then also have a
doctor version, which is doing that
verify but also a full review of the
setup if if the like according certain
rules.
Um it's not an AI like rule verifier.
It's just like, hey, um
is the origin set
um correctly? Are the skills like
correct? So it must it must very very
validate some things. And somehow the
Opus thinks that this is a full offline
mode. But at the same sentence it says,
but we are doing a remote check like to
get remote. And I'm like, yeah, but
that's that doesn't make sense. And then
when I highlight that one thing saying
like, hey, you said it's offline and
then here you say we're validating the
get remote, then that's not offline. And
then it goes through the whole document
and says, actually you're right. There's
quite a few areas where there's this
tension and there's some
inconsistencies. So so so going through
this like 500 lines and trying to
identify this is is is damn hard. Um,
and and then once I'm feel confident in
that because
>> five 500 lines of user stories you mean,
right?
>> Yeah, 500 lines markdown.
>> That's what you focus on.
>> functional requirements, acceptance
criteria. Yeah, I'm trying to identify
if they make sense.
And then I also need to make a
verification that no
term confusion because it will it will
have the terminology in the user story
will be one word and then in the
implementation plan originally in the it
will will be another word. And then the
agent doing the implementation would be
confused
because of the meaning behind the word
and then it would do things that are
unexpected. So I found that those are a
lot
>> I'm boarding a train now
and so probably won't work.
Yeah, probably won't work.
Things are starting to move.
>> Uh
>> Cuz you're in like this underdeveloped
country, right? I guess the internet
doesn't work in trains.
>> It doesn't really work.
>> It works in Vietnam, though.
>> It does?
>> Not in UK as far as I know.
>> Okay, anyway,
let's just call it a day right now.
And uh
and I'll I'll catch up with you tomorrow
or something, okay?
>> Yeah, so I think
yeah, have a nice uh trip.
>> Thanks.
I'll give you updates on WhatsApp. Bye.
>> Yeah, bye-bye.