Conductor Podcast: AI Orchestration and Human‑In‑The‑Loop

Name: How Conductor CEO Charlie Holtz Sets Up His Team Of AI Agents
Uploaded: 2026-06-04T14:00:22+00:00
Duration: 16 min 35 s
Channel: Y Combinator
Description: Summary and key takeaways on Conductor Podcast: AI Orchestration and Human‑In‑The‑Loop, covering Product Philosophy Conductor treats coding agents like a

Y Combinator

Jun 04, 2026

•

16 min video

•

2 min read

YouTube video ID: fQmlML9Lay4

Source: YouTube video by Y Combinator — Watch original video

PDF

Conductor treats coding agents like a musical ensemble, letting CEOs of projects orchestrate AI‑generated work through prompts and digestible reports. Code has become “sawdust,” a byproduct rather than the primary focus, while the platform’s goal is to give users the feeling of running a company that directs agents. Human‑led design remains essential; AI should not dictate UI choices because the result can feel uncrafted. The software is envisioned as “malleable,” allowing users to mod workflows much like video‑game skins, keeping the core structure stable while customizing the experience.

Workflow & Tooling

Agents operate inside workspaces where each task produces a pull request (PR). The PR sits in a “in‑review” stage until a human reviews, comments, or merges it. If AI‑generated changes are insufficient, users switch to “caveman mode,” manually editing files before the agent resumes. Certain parts of the codebase are marked as “slot‑free zones,” mandating human oversight to prevent AI from developing harmful patterns. The app tracks progress through a state machine—“in‑progress,” “in‑review,” and “done”—and presents a dashboard view of all active agent work.

Technical Implementation

Conductor is built as a Tauri desktop app with a Rust backend and a TypeScript frontend, while authentication and web services run on Phoenix/Elixir. Engineering practices and constraints are encoded in “skills files” and a cloud.md document that guide the agents. The team runs local models such as Parakeet on high‑end hardware equipped with 128 GB RAM, using a “fast mode” to maximize token efficiency. Permissions are set to “dangerously accept all,” granting agents broad autonomy within the defined workflow.

Future Vision

The roadmap emphasizes software that is both malleable and decoupled from infrastructure. Users will be able to mod their development environment—changing prompt speeds, UI skins, or agent collaboration patterns—without altering the underlying codebase. Multi‑agent collaboration is expected to expand, allowing several specialized AIs to work together while the human remains the final arbiter of quality and direction.

Takeaways

Conductor treats coding agents like a musical ensemble, letting CEOs of projects orchestrate AI‑generated work through prompts and digestible reports.
The platform enforces a human‑in‑the‑loop workflow by requiring every AI‑produced change to appear as a pull request that a person must review, comment on, or edit in “caveman mode.”
Its UI philosophy favors visual, spatial interfaces over terminal‑style interactions, arguing that AI‑chosen UI elements often feel uncrafted.
Built with Tauri (Rust backend, TypeScript frontend) and a Phoenix/Elixir web service, Conductor runs local models like Parakeet on high‑end hardware (128 GB RAM) and uses “fast mode” to keep token usage efficient.
Looking ahead, Conductor envisions software as a malleable, mod‑friendly layer where users can customize workflows like game skins while the underlying code and infrastructure stay decoupled.

Frequently Asked Questions

What is “caveman mode” in Conductor’s workflow?

Caveman mode allows users to bypass AI‑generated changes and edit files manually. When an agent’s suggestion is insufficient, the developer opens the affected files, makes direct edits, and then saves the changes, after which the agent can continue its work. This fallback keeps human control when AI output falls short.

How does Conductor ensure AI cannot commit directly to the main branch?

Conductor enforces a human‑AI boundary by routing every AI task through isolated workspaces that generate pull requests. The system’s state machine moves each task from “in‑progress” to “in‑review” and finally “done,” but the PR must be reviewed and merged by a human before any code reaches the main branch, preventing autonomous commits.

Who is Y Combinator on YouTube?

Y Combinator is a YouTube channel that publishes videos on a range of topics. Browse more summaries from this channel below.

Does this page include the full transcript of the video?

Yes, the full transcript for this video is available on this page. Click 'Show transcript' in the sidebar to read it.

Helpful resources related to this video

If you want to practice or explore the concepts discussed in the video, these commonly used tools may help.

High Performance Desktop Computer 128gb Ram Recommended

Provides the necessary memory capacity to run local AI models and manage complex agent orchestration workflows efficiently.

Amazon →

Adjustable Gooseneck Microphone For Voice Commands

Enables hands-free voice-to-AI interaction, which is a core method for initiating tasks within the Conductor workflow.

Amazon →

Ergonomic Mechanical Keyboard For Software Developers

Supports the 'caveman mode' manual coding and human-in-the-loop review processes by providing a tactile, comfortable interface for long sessions.

Amazon →

Ultrawide Monitor For Software Development Workflow

Provides the screen real estate needed to manage multiple agent workspaces, PR reviews, and status dashboards simultaneously.

Amazon →

Links may be affiliate links. We only include resources that are genuinely relevant to the topic.

Summarize another video

Full Transcript YouTube

Hello, I am Charlie, the co-founder of
Conductor, which is an app that lets you
orchestrate a bunch of coding agents on
your Mac. And we were YC Summer 24. Uh,
and I'd love to show you my setup.
So, a recent thing that I can't live
without is this uh gooseeneck
microphone. $20 on Amazon. we are all
trying to talk to our computers more.
Um, one issue with having like an open
floor plan office is that can be pretty
distracting. So, one advantage of these
is you can like lean over and whisper
into Claude and be like, "Please uh
merge PR 3475
and it's a little bit less disruptive. I
we we all got these in an attempt to
encourage more talking to computers." I
spend most of my day in conductor. We're
using conductor to build conductor. One
thing that I do is I'm constantly
kicking off new tasks. So I'm constantly
um going commandn. Uh that was actually
a sneak a sneak peek of something we are
working on which is cloud workspaces.
But I'll I'll do command n and then I'll
speak into my computer and say um can
you take a look at the latest linear
issue and give me a rough pass at how
you'd solve it? Stuff like that. And
then press enter. And then I can see
that it's running in the sidebar. And
while cloud is working I'll go to
another chat. I'm very into keyboard
shortcuts. to like try and make
everything have a keyboard shortcut. So
in this case I'll do command shift Y. I
can see here that this workspace is
ready to merge. So I'll take a look at
it. Give Claude a quick review. Um in
this case it's a pretty small PR and so
it looks good to me. But quite often
Claude won't get things exactly right
and I'll give things a a comment like a
GitHub style comment say uh this looks a
little bit weird to me. Why do we need
this? Press enter. get cloud running and
then go back to a different workspace. A
big part of um how I use conductor is
like experimentation. Like I'm always
kicking off workspaces to try different
ideas. Most of them don't make it in. So
like you can see we have like four PRs
here that are in review, but like
there's a bunch of random ideas that
I've tried here that are in progress
that you know may never see the light of
day. If I like it, then it might get
promoted to like an internal setting and
then an experimental setting. Okay.
Okay. So, something I'm very excited
about is on the go. Uh, I'm going to
just speak into my phone and say, let's
add a new feature where I can change the
theme to hacker mode.
And then I'm going to click conduct. And
then my computer starts working on it
and I can uh conduct on the go.
>> You still write code today?
>> No. Yeah. No. Very occasionally I will
like edit Tailwind classes or like open
up an IDE to change like a ENV file. We
actually added a mode that we call
caveman mode which is uh you click this
and you can actually type with your
keyboard and like make changes in a
file. Once in a while you do need to
like make a change to a file by hand,
but it's like it's called caveman mode
for a reason. Most of the time if I want
small edits, I'll like highlight and
then tell the AI um about my comments or
I'll just like speak into my computer
and say that button looks a little too
wide like can you uh can you make it
smaller? By the way, this thing is now
ready to merge. Um so I just wanted to
show you I can now click archive and
it's gone from my side panel and like
merge into the codebase and uh this one
I can see that there are checks running
and once it's finished I can just like
click merge and and get it in. We
recently added this thing called status
in the left. So when something like is
kicked off, it's in progress and then
once there's a PR created, it's in
review and then once it's merged, it
goes into the done folder. We have this
new concept of a dashboard page where
from like one place you can see what all
your agents are working on and then like
take them to the next action. But we're
still messing around with like what the
what the interface should look and feel
like. But the ideal is like you should
feel like the CEO of a little company
and you can see all your agents working
for you and they'll bring you up like
digestible reports and then you can
point them in the right direction if
they like need some correction or just
merge it in if it looks good.
>> What are your other main applications,
main software that you use?
>> I use Telegram a decent amount to talk
to my open claw. That's that's been a
recent addition for me. I use spokenly
for text to speech. That's what comes up
when I press uh control space. It's
actually running a local model. It's
running parakeet. Um, I have a really
beefed up computer, so it's like 128
gigabytes of RAM. Partly so I can like
run local models like Parakeet. But as a
side note, I have just recently ordered
the MacBook Neo, like the the bottom of
the line, lowest RAM, lowest memory. I I
I got it basically to like force myself
to like use the lowest spec option.
>> Are there any tweaks that you do still
stand by that are the customizations
that actually do matter?
>> Uh, a couple things. Like we put a lot
of time into our skills files and our
like cloud MD. If I like open it up like
you can you can see like this is uh
probably a few hundred lines. There's
like some interesting things in here
like it's we say engineering practices.
We're a startup. You're probably used to
writing enterprise code but that's not
how we do things around here. And we
have like a lot of things like that that
we've like put into our cloud MD and our
skills files over time. What else do I
do? I always use fast mode. That's not a
default. if you're trying to token max
like you have to be in fast mode. I do
use the context 7 MCP. I think that's
pretty helpful to get documentation but
other than that like use most of the
most of the things out of the box. One
core thing is that we always run claude
and dangerously accept all permissions.
Um like that is not the default and that
is uh the default way to run uh cloud in
conductor. I think something that's
really important to us is having like
clear boundaries between uh well we call
them slot free zones um and having like
parts of the code base or like parts of
the documentation that we like know is
written by a human. It's possible like
that the a the AI can contribute to the
the slot free zones but like it has to
be like every line has to be read by a
human. I think this actually like served
us pretty well. Um because if if you're
not careful like the AI can like get in
a vicious a vicious cycle where like it
sees bad code and then it writes more
bad code as a result. And the same thing
can happen in like the positive uh
positive direction. We have like some
lines in our codebase that are like do
not touch if you are an AI like this is
for human eyes only.
>> What's the connector text stack?
>> It's a towery app. So it's using the
native uh Safari web renderer and the
back end is technically Rust but we
write almost everything in TypeScript.
Um so it's like probably 90 95%
TypeScript on the like the desktop app.
The web app is Elixir. It's a it's a
Phoenix app. It's a very small app
because literally all you can do in it
right now is just like log in. But I am
I'm a huge Elixir fan and I am always
like pushing for more Elixir in our
codebase when we can. But um most of
what we're doing is in Typescript.
Another thing we talk about is like
don't let the AI be your architect. Even
the concept of like a workspace here in
the sidebar which in some ways is just
like an abstraction around a work tree
at least for right now. Like that's
actually going to change soon. But even
that like concept of a workspace like we
as a human had to like think that
through. The other thing is like like
design and like interface decisions.
this concept of having like all your
chats here on the left and then the chat
in the middle and then the right sidebar
where you can like review code changes
or run your app like we put a lot of
thought into into those decisions. And I
think if you let the AI make your like
UI choices for you, you can end up with
something that like it just doesn't feel
like crafted. It's like really important
to us that it it feels crafted like even
this decision. So like we we thought for
a long time about how this open in
button should work which is kind of
funny because now there's like so many
apps have like have the same pattern.
The thing that we were really thinking
about is whether we should show the
icons in the top. I was pretty against
showing icons here at first because it
just feels like okay in the top bar of
our app like we're like advertising a
different app but now I I really like it
and it's like a clear visual of like
what's going to happen um when you click
it. I think something we would do a bit
differently is building the core of the
app um around like human written APIs
and like contracts that the AI wouldn't
contribute to as much. And then I think
that like it's important to have a big
chunks of your codebase be like have
like free reign for the AI where you can
just like throw a ton of different ideas
at it and know that it's like not going
to affect the core infrastructure. And I
think right now the the boundaries are a
little murky and like um that's
something we're we're working on
improving. I think it's really important
to us that we stay like a little ahead
of the frontier, like push push people's
comfort zones a little bit more than u
they'd expect. When we first launched
Conductor, uh most of the feedback we
got was like this is crazy. Like I
barely can manage like one cloud code or
like one codeex. Like how am I going to
manage like three or like even five? We
also purposely made it so like you can't
edit files directly. like we like made
it so that like anytime a workspace like
has to be a work tree and it has to then
create a PR and then you have to merge
it. So we really like enforced our
workflow. I think like what's what's
exciting but also hard about where we're
at is like we have to constantly adapt
to where like the models are going. So
that's one reason like we are putting so
much work into like cloud right now is
right now like you shut your laptop and
the agents are going to stop running.
But like feels like we're very quickly
moving to a world where the agents are
going to run for 10 times longer and
they're going to be 10 times smarter and
they're going to need to run in an
environment that like isn't constrained
by like your max like CPU.
>> It seems like you're building conductor
in a very opinionated way. How do you
build a conviction behind your
decisions? That that's a great that's a
great question because yeah like it
especially for like our audience they
want a lot of like configuration and I
do think it is important for the tool to
like be flexible and to like feel like
yours but the way we build conviction is
we force ourselves to use it because
actually we don't even force like we we
just use it every day and so if it
doesn't feel right like we like quickly
can can decide but we we we're not big
on analytics or like looking at like our
AB testing or like It's very much a like
gut feel. This feels right. Like when I
click this, it feels right that it opens
in the center. And that way I don't need
a separate composer and I can type
messages here and it all feels unified.
>> You sound like you default to cloud code
in a lot of places, but conductor
supports codeex too. When do you reach
for codeex?
>> Okay. I've recently actually been using
codeex more. Codex is like the
workhorse. it will power through like a
specific problem or like uh it's not
afraid to do a ton of tool calls and
like debug something with me for a long
time. Cloud I'll reach for when I want a
little more like back and forth. I feel
like Opus is just like a little more
creative, like a little more uh of a
partner. And so I would say like when
I'm building out a new a new feature
like I I probably would like
instinctively reach for Opus. And then
when I'm like okay now we just want to
get stuff done like I I'll go to Codeex.
Why isn't just a terminal good enough?
>> There's a reason uh we moved from
terminal interfaces to like gooey
interfaces in the 80s. Like I think
humans are spatial visual creatures and
like having a a command line interface
just like feels like it's feels very
like restrictive and I think it maybe
works for the AI brains better than the
human brains. But I think just like I
want to know that okay my chats are over
here and my like review panel is here. I
can talk to the AI in the middle. I just
think like yeah bottom line like humans
are like visual visual creatures. I also
think like like out of like a like
zooming in a little bit like there's a
lot that you can't do in a terminal um
like that you can do with a uh user
interface.
>> Let's talk about token maxing.
>> Yeah.
>> What's your high water mark on lines of
code in a day or spend in a month?
>> I think the highest spend was when we
were starting out conductor like in July
2025. I spent $22,000
on tokens that month. Granted, that was
with like previous generation of models.
Um, and the lines of code was must have
been like tens of thousands that month.
I'm very big on spending like on token
maxing like using fast mode like think
extra hard all like high effort all the
time, but we're not big on lines of
code. We uh we try and keep the lines of
code minimal actually. There's a bunch
of reasons for this, but I think like
you can quickly spiral. Your codebase
can spiral out of control if you're like
not careful about the lines of code
added. But I I I think about it very
differently if I'm like starting up an
app versus like working established
codebase like Conductor.
>> What's different about your workflows
today from say 6 months ago?
>> On a lot of like hard PRs, I would open
an IDE and make changes by hand. And I
also use GitHub like the web app a lot a
lot less now because I can just like
review the code changes here in
conductor and like add comments here if
I need to. We do have like a lot of PR
checks that run. Um and uh so that's why
we recently added this like uh this
checks tab which lets us just like add
comments from GitHub like into
Conductor.
>> What's the most surprising thing you've
seen someone else do with Conductor? One
was like someone built like a mobile
version of Conductor by like hacking
together a bunch of our I don't actually
even really know how it works, but I
know it's like spoofing like IPC calls
to our desktop app, which is pretty
interesting. I think honestly
Gary has shown us a lot of what you can
do with conductor. He is really putting
it to the test. I think I've learned
from him a bit about like how hard you
can go on skills. like skills are very
much like a first class thing in in
GStack and it's like it's there's some
like interesting ideas there I think
like especially around like onboarding
and we've added actually a specific mode
for him called Gary mode which by
default does not collapse any of the
tool calls so you can see all the tool
calls are default on collapse and you
can even actually see uh Gary's face
here if you're in Gary's mode
>> what feels obvious to you and your team
that the rest of the world doesn't fully
understand yet
>> like I think there's like a lot of cool
stuff to explore with like collaboration
between humans and the AIs. Should you
be able to communicate with sub agents?
Should you be able to have like
multiplayer chats where like multiple
people are working on the same thing
with the AIS? And then of course like
the a metaphor we we'll often talk about
is like feeling like the conductor of an
orchestra. You like uh wave the baton
and like the instruments are playing in
unison and then once in a while you want
to go to like the the trumpet player and
be like, "Okay, you're out of tune." And
then you want to like zoom out to like
the string section and like uh you
should play a bit faster, but then most
of the time you're like conducting at
the the orchestra level. Code is almost
like uh sawdust now in that like it used
to be that code was the thing you were
building. It was like the structure. You
were putting time into like in in into
like crafting the code and now you're
putting time into describing what you
want and how you want it to be built.
And the code is almost just like sawdust
that comes out of that process. And like
that leads to like a lot of like
interesting conclusions. Like one of
them is like really what matters is your
prompts. And like when the next
generation of models come out, you can
just like rerun your prompts again and
then you'll get new code and the old
code didn't really matter. I think
that's one thing that like the world is
slowly waking up to. I think like the
submit a prompt like the prompt request
feature is sort of like an early
experiment with malleable software. I
the the metaphor that I always think of
when I think of malleable software is
like video games and how like when you
play like Call of Duty like the
structure of the game is the same for
everyone and like the skeleton is the
same but each person can like I don't
know like use custom skins or like
faster like reload speeds or whatever
and like the same way you can like mod a
video game. I want you to be able to mod
Conductor and like yeah build in your
own workflows a little bit. It's
important that like the structure feels
the same and like people want software
that's like been crafted and been like
really thought through. But I also, you
know, like video game mods make make the
game feel more like your own. And um I
think that's going to happen with
software as well.

Help & FAQ

Emergent: How Six Months of Tinkering Led To A $100M ARR Company

Y Combinator

Jun 06, 2026

Workflow & Tooling

Technical Implementation

Future Vision

Takeaways

Frequently Asked Questions

What is “caveman mode” in Conductor’s workflow?

How does Conductor ensure AI cannot commit directly to the main branch?

Who is Y Combinator on YouTube?

Does this page include the full transcript of the video?

Helpful resources related to this video

Share This Summary

Embed This Summary