0:00
Hello, I am Charlie, the co-founder of
0:02
Conductor, which is an app that lets you
0:04
orchestrate a bunch of coding agents on
0:06
your Mac. And we were YC Summer 24. Uh
0:10
and I'd love to show you my setup.
0:22
So, a recent thing that I can't live
0:24
without is this uh gooseneck microphone,
0:26
$20 on Amazon. We are all trying to talk
0:29
to our computers more. Um
0:31
one issue with having like an open floor
0:33
plan office is that can be pretty
0:35
distracting. So, one advantage of these
0:37
is you can like lean over and whisper
0:39
into Claude and be like, "Please uh
0:44
And it's a a little bit less disruptive.
0:46
I think we we all got these in an
0:47
attempt to encourage more talking to
0:49
computers. I spend most of my day in
0:51
Conductor. We're using Conductor to
0:53
build Conductor. One thing that I do is
0:55
I'm constantly kicking off new tasks.
0:57
So, I'm constantly um going command N.
1:00
Uh that was actually a sneak a sneak
1:02
peek of something we are working on,
1:04
which is Claude workspaces. But I'll
1:05
I'll I'll do command N and then I'll
1:07
speak into my computer and say, "Um can
1:10
you take a look at the latest linear
1:13
issue and give me uh a rough pass at how
1:16
you'd solve it." Stuff like that. And
1:18
then press enter. And then I can see
1:20
that it's running in the sidebar. And
1:21
while Claude is working, I'll go to
1:23
another chat. I'm very into keyboard
1:25
shortcuts. So, like try to make
1:26
everything have a keyboard shortcut. So,
1:27
in this case I'll do command shift Y. I
1:30
can see here that this workspace is
1:32
ready to merge. So, I'll take a look at
1:33
it, give Claude a quick review. Um in
1:36
this case it's a pretty small PR. And
1:38
so, it looks good for to me, but quite
1:40
often Claude won't get things exactly
1:42
right. And I'll give things uh a comment
1:46
like a GitHub-style comment, say, "Uh
1:48
this looks a little bit weird to me. Why
1:51
Press enter, get Claude running, and
1:53
then go back to a different workspace. A
1:56
big part of um how I use Conductor is
1:58
like experimentation. Like I'm always
2:00
kicking off workspaces to try different
2:02
ideas. Most of them don't make it in.
2:05
So, like you can see we have like four
2:06
PRs here that are in review, but like
2:09
there's a bunch of random ideas that
2:11
I've tried here that are in progress
2:13
that you know, may never see the light
2:15
of day. If I like it, then it might get
2:17
promoted to like an internal setting and
2:19
then an experimental setting. Okay, so
2:20
something I'm very excited about is on
2:22
the go. Uh I'm going to just speak into
2:24
my phone and it's say, "Let's add a new
2:26
feature where I can change the theme to
2:32
And then I'm going to click conduct.
2:35
And then my computer starts working on
2:37
it and I can uh conduct on the go.
2:40
>> Do you still write code today?
2:42
Yeah, no. Very occasionally I will like
2:44
edit Tailwind classes or like open up an
2:47
IDE to change like a .env file. We
2:51
actually added a mode that we call
2:54
caveman mode, which is uh you click this
2:57
and you can actually type with your
3:00
keyboard and like make changes in a
3:02
file. Once in a while you do need to
3:04
like make a change to a file by hand,
3:07
but it's like it's called caveman mode
3:09
for a reason. Most of the time if I want
3:10
small edits, I'll like highlight and
3:12
then uh tell the AI um about my comments
3:15
or I'll just like speak into my computer
3:17
and say, "That button looks a little too
3:19
wide. Like, can you uh can you make it
3:22
By the way, this thing is now ready to
3:25
wanted to show you I can now click
3:27
archive and it's gone from my side panel
3:30
and like merge into the codebase and uh
3:32
this one I can see that there are checks
3:34
running and once it's finished, I can
3:35
just like click merge and and get it in.
3:37
We recently added this thing called
3:39
status in the left. So, when something
3:42
like is kicked off, it's in progress and
3:44
then once there's a PR created, it's in
3:45
review and then once it's merged, it
3:47
goes into the done folder. We have this
3:49
new concept of a dashboard page where uh
3:53
from like one place you can see what all
3:55
your agents are working on and then like
3:57
take them to the next action. But, we're
4:00
still messing around with like what the
4:02
what the interface should look and feel
4:03
like. But, the ideal is like you should
4:05
feel like the CEO of a little company
4:07
and you can see all your agents working
4:09
for you and they'll bring you up like
4:10
digestible reports and then you can put
4:12
them in the right direction if they like
4:14
need some correction or just merging in
4:17
>> What are your other main applications,
4:20
main software that you use?
4:22
>> I use Telegram a decent amount to talk
4:23
to my open claw. That's that's been a
4:25
recent addition for me. I use spokenly
4:28
for text to speech. That's what comes
4:30
out when I press control space. It's
4:32
actually running a local model. It's
4:34
running Parakeet. I have a really beefed
4:36
up computer. So, it's like 128 GB of
4:39
RAM. Partly so I can like run local
4:40
models like Parakeet. But, as a side
4:43
note, I have just recently ordered the
4:45
MacBook Neo, like the the bottom of the
4:47
line lowest RAM, lowest memory. I I got
4:50
it basically to like force myself to
4:52
like use the lowest spec option.
4:53
>> Are there any tweaks that you do still
4:55
stand by that are the customizations
4:57
that actually do matter?
4:58
>> A couple of things like we put a lot of
5:01
time into our skills files and our like
5:03
Claude MD. If I like open it up like you
5:06
can see you can see like this is
5:08
probably a few hundred lines. There's
5:10
like some interesting things in here
5:12
like it's we say engineering practices.
5:14
We're a startup. You're probably used to
5:15
writing enterprise code, but that's not
5:17
how we do things around here. We have
5:19
like a lot of things like that that
5:20
we've like put into our Claude MD and
5:22
our skills files over time. What else do
5:24
I do? I always use fast mode. That's not
5:27
a default. If you're trying to token max
5:28
like you have to be in fast mode. I do
5:30
use a context 7 MCP. I think that's
5:32
pretty helpful to get documentation.
5:34
But, other than that like use most of
5:36
the most of the things out of the box.
5:38
One core thing is that we always run
5:41
Claude and dangerously accept all
5:42
permissions. Like that is not the
5:45
default and that is uh default way to
5:47
run Cloud and Conductor. I think
5:49
something that's really important to us
5:50
is having like clear boundaries between
5:53
uh well, we call them slot-free zones.
5:56
Um and having like parts of the codebase
5:57
or like parts of the documentation that
5:59
we like know is written by a human. It's
6:02
possible like that the AI the AI can
6:04
contribute to the the slot-free zones,
6:05
but like it has to be like every line
6:08
has to be read by a human. I think this
6:11
served us pretty well. Um because if if
6:14
you're not careful like the AI can like
6:17
get in a a vicious a vicious cycle where
6:19
like it sees bad code, and then it
6:20
writes more bad code as a result. And
6:22
the same thing can happen in like the
6:23
positive uh positive direction. We have
6:26
like some lines in our codebase that are
6:28
like do not touch if you are an AI. Like
6:31
this is for human eyes only.
6:32
>> What's the Conductor tech stack?
6:34
>> It's a Tauri app. So, it's using the
6:35
native uh Safari web renderer, and the
6:38
backend is technically Rust, but we
6:41
write almost everything in TypeScript.
6:43
Um so, it's like probably 90 95%
6:45
TypeScript on the like the desktop app.
6:47
The web app is Elixir. It's a it's a
6:50
Phoenix app. It's a very small app
6:53
because literally all you can do in it
6:54
right now is just like log in. But, I am
6:58
I'm a huge Elixir fan, and I am always
6:59
like pushing for more Elixir in our
7:01
codebase when we can. But, um
7:04
most of what we're doing is in
7:05
TypeScript. Another thing we talk about
7:07
is like don't let the AI be your
7:08
architect. Even the concept of like a
7:10
workspace here in the sidebar, which in
7:13
some ways is just like an abstraction
7:16
around a work tree, at least for right
7:17
now. Like that's actually going to
7:19
change soon. But, even that like concept
7:23
we as a human had to like think that
7:25
through. The other thing is like like
7:27
design and like interface decisions.
7:30
This concept of having like all your
7:31
chats here on the left, and then the
7:33
chat in the middle, and then the right
7:35
sidebar where you can like review code
7:37
changes or run your app. Like we put a
7:40
lot of thought into into those
7:42
decisions, and I think if you let the AI
7:43
make your like UI choices for you, you
7:46
can end up with something that like it
7:48
just doesn't feel like crafted. It's
7:49
like really important to us that it
7:51
feels crafted. Like even this decision,
7:53
so like we we thought for a long time
7:55
about how this open in button should
7:57
work, which is kind of funny because now
7:58
there's like so many apps have like have
8:00
the same pattern. The thing that we were
8:02
really thinking about is whether we
8:03
should show the icons in the top. I was
8:06
pretty against showing icons here at
8:10
first because it just feels like okay,
8:12
in the top bar of our app like we're
8:14
like advertising a different app, but
8:16
now I I really like it and it's like a
8:17
clear visual of like what's going to
8:19
happen when you click it. I think
8:21
something we would do a bit differently
8:22
is building the core of the app around
8:25
like human written APIs and like
8:27
contracts that the AI wouldn't
8:29
contribute to as much and then I think
8:31
that like it's important to have a big
8:33
chunks of your code base be like have
8:34
like free reign for the AI where you can
8:36
just like throw a ton of different ideas
8:38
at it and know that not going to affect
8:41
the core infrastructure. And I think
8:43
right now the the boundaries are a
8:44
little murky and like that's the thing
8:46
we're we're working on improving. I
8:48
think it's really important to us that
8:51
a little ahead of the frontier. Like
8:53
push push people's comfort zones a
8:56
little bit more than they'd expect. When
8:59
we first launched conductor,
9:01
most of the feedback we got was like
9:03
this is crazy. Like I barely can manage
9:05
like one cloud code or like one code X.
9:07
Like how am I going to manage like three
9:09
or like even five? We also purposely
9:12
made it so like you can't edit files
9:14
directly. Like we like made it so that
9:16
like anytime a workspace like has to be
9:19
a work tree has to then create a PR and
9:21
then you have to merge it. So we really
9:23
like enforce our workflow. I think like
9:25
what's what's exciting but also hard
9:27
about where we're at is like we have to
9:29
constantly adapt to where like the
9:31
models are going. So that's one reason
9:33
like we are putting so much work into
9:35
like cloud right now is right now like
9:37
you shut your laptop and the agents are
9:39
going to stop running. But like feels
9:41
like we're very quickly moving to a
9:42
world where the agents are going to run
9:44
for 10 times longer and they're going to
9:45
be 10 times smarter and they're going to
9:47
need to run in an environment that like
9:48
isn't constrained by like your Mac's
9:52
>> It seems like you're building conductor
9:53
in a very opinionated way. How do you
9:55
build the conviction behind your
9:57
>> Okay, that's a great that's a great
9:58
question cuz yeah like it especially for
10:00
like our audience, they want a lot of
10:02
like configuration. And I do think it is
10:04
important for the tool to like be
10:06
flexible and to like feel like yours.
10:08
But the way we build conviction is we
10:11
force ourselves to use it because
10:13
actually we don't even force it like
10:15
we we just use it every day. And so if
10:17
it doesn't feel right, like we like
10:19
quickly can can decide. But we we don't
10:22
we're not big on analytics or like
10:24
looking at like our AB testing or like
10:26
it's very much a like gut feel, this
10:29
feels right. Like when I click this, it
10:31
feels right that it opens in the center
10:33
and that way like I don't need a
10:34
separate composer and I can type
10:37
messages here and it all feels unified.
10:40
>> You sound like you default to cloud code
10:41
in a lot of places, but conductor
10:42
supports codex too. When do you reach
10:45
>> Oh yeah, I've recently actually been
10:46
using codex more. Codex is like the
10:48
workhorse. It will power through like a
10:50
specific problem or like uh it's not
10:53
afraid to do a ton of tool calls and
10:55
like debug something with me for a long
10:57
time. Cloud I'll reach for when I want a
10:59
little more like back and forth. I feel
11:01
like Opus is just like a little more
11:03
creative, like a little more
11:05
of a partner. And so I would say like
11:07
when I'm building out a new a new
11:09
feature, like I I probably would like
11:11
instinctively reach for Opus. And then
11:13
when I'm like, "Okay, now we just want
11:14
to get stuff done." like I'll go to
11:17
>> Why isn't just a terminal good enough?
11:21
uh we moved from terminal interfaces to
11:24
like GUI interfaces in the '80s. Like I
11:27
think humans are spatial visual
11:29
creatures and like having a a command
11:32
line interface just like feels like it's
11:34
feels very like restrictive and I think
11:36
it maybe works for the AI brains better
11:38
than the human brains, but I think just
11:39
like I want to know that okay, my chats
11:41
are over here and my like review panel
11:44
is here. I can talk to the AI in the
11:45
middle. I just think like yeah, bottom
11:47
line like humans are like visual visual
11:50
creatures. I also think like like out of
11:52
like a like zooming in a little bit like
11:54
there's a lot that you can't do in a
11:56
terminal um like that you can do with a
12:01
>> Let's talk about token maxing.
12:03
>> What's your high water mark on lines of
12:05
code in a day or spend in a month?
12:07
>> I think the highest spend was when we
12:09
were starting out conductor like in July
12:12
2025. I spent $22,000
12:16
on tokens that month. Granted that was
12:18
with uh like previous generation of
12:20
models. Um and the lines of code was
12:23
must have been like tens of thousands
12:25
that month. I'm very big on spending
12:28
like on token maxing, like using fast
12:30
mode, like think extra hard all like
12:32
high effort all the time, but we're not
12:34
big on lines of code. We we try and keep
12:36
the lines of code minimal actually.
12:38
There's a bunch of reasons for this, but
12:40
I think like you can quickly spiral your
12:42
code base can spiral out of control if
12:43
you're like not careful about the lines
12:45
of code added. But I I I think about it
12:47
very differently if I'm like starting up
12:49
an app versus like working in
12:51
established code base like conductor.
12:53
>> What's different about your workflows
12:55
today from say 6 months ago?
12:57
>> On a lot of like hard PRs, I would open
13:00
an IDE and make changes
13:03
by hand and I also use GitHub like the
13:05
web app a lot a lot less now because I
13:07
can just like review the code changes
13:09
here in conductor and like add comments
13:12
here if I need to. We do have like a lot
13:14
of PR checks that run um and uh so
13:18
that's why we recently added this like
13:20
uh this checks tab which lets us just
13:22
like add comments from GitHub like into
13:24
>> What's the most surprising thing you've
13:26
seen someone else do with Conductor?
13:28
>> One was like someone built like a mobile
13:30
version of Conductor by like hacking
13:32
together a bunch of our I don't actually
13:34
even really know how it works, but I
13:35
know it's like spoofing like IPC calls
13:38
to our desktop app, which is pretty
13:40
interesting. I think honestly
13:43
Gary has shown us a lot of what you can
13:45
do with Conductor. He is really putting
13:48
it to the test. I think I've learned
13:50
from him a bit about like how hard you
13:52
can go on skills. Like skills are very
13:54
much like a first-class thing in in G
13:56
stack and it's like it's there's some
13:58
like interesting ideas there, I think.
14:00
Like especially around like onboarding.
14:02
Then we've added uh actually a specific
14:04
mode for him called Gary mode, which by
14:06
default does not collapse any of the
14:08
tool calls. So, you can see all the tool
14:11
calls are default on collapse and uh you
14:14
can even actually see uh Gary's face
14:16
here if you're in Gary's mode.
14:17
>> What feels obvious to you and your team
14:19
that the rest of the world doesn't fully
14:22
>> Like I think there's like a lot of cool
14:23
stuff to explore with like collaboration
14:25
between humans and the AIs. Should you
14:27
be able to communicate with sub-agents.
14:29
Should you be able to have like
14:30
multiplayer chats where like multiple
14:32
people are working on the same thing
14:33
with the AIs. And then of course like
14:35
the a metaphor we'll we'll often talk
14:37
about is like feeling like the conductor
14:39
of an orchestra. You like uh wave the
14:41
baton and like the instruments are
14:44
playing in unison and then once in a
14:45
while you want to go to like the the
14:48
trumpet player and be like, "Okay,
14:49
you're out of tune." And then you want
14:50
to like zoom out to like the string
14:52
section and like uh you should play a
14:54
bit faster. But then most of the time
14:56
you're like conducting at the the
14:57
orchestra level. Code is almost like uh
15:00
sawdust now in that like it used to be
15:03
that code was the thing you were
15:05
building. It was like the structure. You
15:07
were putting time into like
15:09
in in into like crafting the code. And
15:11
now you're putting time into describing
15:14
what you want and how you want it to be
15:15
built and the code is almost just like
15:17
sawdust that comes out of that process.
15:20
And like that leads to like a lot of
15:21
like interesting conclusions. Like one
15:23
of them is like really what matters is
15:25
your prompts. And like when the next
15:26
generation of models come out, you can
15:28
just like rerun your prompts again and
15:30
then you'll get new code and the old
15:31
code didn't really matter. I think
15:33
that's one thing that like the world is
15:35
slowly waking up to. I think like the
15:37
submit a prompt like the prompt request
15:39
feature is sort of like an early
15:42
experiment with malleable software. I
15:44
the metaphor that I always think of when
15:46
I think of malleable software is like
15:48
video games and how like when you play
15:50
like Call of Duty, like the structure of
15:52
the game is the same for everyone and
15:54
like the skeleton is the same, but each
15:56
person can like I don't know like use
15:58
custom skins or like faster like reload
16:01
speeds or whatever. And like same way
16:02
you can like mod a video game, I want
16:05
you to be able to mod conductor and like
16:08
yeah, build in your own workflows a
16:09
little bit. It's important that like the
16:10
structure feels the same and like people
16:12
want software that's like been crafted
16:14
and been like really thought through,
16:16
but I also, you know, like video game
16:19
mods make make the game feel more like
16:21
your own and um I think that's going
16:23
[music] to happen with software as well.
