Replit’s Vibe Coding: AI Platform Democratizes Software Building

Name: The $9B startup that wants to create a billion new developers
Uploaded: 2026-04-25T14:30:37+00:00
Duration: 39 min 11 s
Channel: Y Combinator
Description: Summary and key takeaways on Replit’s Vibe Coding: AI Platform Democratizes Software Building, covering Company Mission & Evolution Replit wants anyone who can

Y Combinator

Apr 25, 2026

•

39 min video

•

2 min read

YouTube video ID: kMYeTRqzAfc

Source: YouTube video by Y Combinator — Watch original video

PDF

Replit wants anyone who can read and write to deploy scalable, secure software. The company began by improving development environments and later abstracted code entirely through AI agents. “Vibe coding” lets users interact with the platform via natural language, comments, and drag‑and‑drop, while a hidden coding agent generates the underlying code.

User Base & Market

Replit serves “tech‑adjacent” creators such as product managers, designers, entrepreneurs, and domain experts. Physical therapists, for example, now build sophisticated health‑tech applications without hiring engineers. Enterprises adopt Replit to accelerate product development and replace rigid, siloed SaaS tools with custom internal solutions.

Product Strategy & AI Integration

Replit aligns its roadmap with AI capability step‑changes roughly every six months. Agent 4 adds parallel agents, asynchronous design canvases, and multi‑modal interaction across web, mobile, and other platforms. This shift creates a “post‑prompting” world where users supply high‑level goals and the system decomposes them into functional software.

Mechanisms

Vibe Coding: An AI coding agent works behind the scenes, turning natural‑language instructions, visual canvas actions, and high‑level goals into working code.
Parallel Agent Architecture: Multiple agents operate simultaneously, handling distinct tasks, resolving merge conflicts, and enabling asynchronous workflows.
Skill Integration (MCP): External APIs and “skills” such as Stripe or Salesforce become part of the agent’s context, allowing the AI to “download” capabilities on demand.
Internal Tooling Loop: Employees spot a bottleneck, build a custom solution on Replit, and deploy it, bypassing external SaaS procurement.

Business & Growth

Replit follows a product‑led growth model: individuals discover the tool for personal projects, then bring it into their workplaces. Sales teams act as evangelists, running hackathons that create internal champions and accelerate enterprise adoption. Y Combinator’s mentorship shaped the company’s fundraising, networking, and intense short‑term sprint culture.

The Future of Work

Replit envisions organizations composed of “builders” and “salespeople” who act as generalist entrepreneurs. Employees identify business problems, delegate solutions to AI agents, and launch internal tools that automate processes. This model blurs the line between product development and sales, redefining how companies create value.

Takeaways

Vibe coding lets anyone who can read and write create scalable, secure software using natural language, comments, and drag‑and‑drop, removing the need to write code manually.
The platform targets tech‑adjacent creators—product managers, designers, domain experts—enabling them to build personal apps, vertical SaaS, and internal enterprise tools without traditional engineering teams.
Agent 4 introduces parallel agents, asynchronous design canvases, and multi‑modal interaction, moving the product toward a “post‑prompting” model where users set high‑level goals instead of detailed prompts.
Replit grows through product‑led adoption that graduates users from personal projects to enterprise deployment, while sales teams run hackathons to create internal champions and accelerate adoption.
The future of work at Replit envisions “generalist entrepreneurs” who identify business problems and delegate solutions to AI agents, reshaping organizations into builder‑salesperson hybrids.

Frequently Asked Questions

What is vibe coding and how does it differ from traditional programming?

Vibe coding uses an AI coding agent that works behind the scenes, allowing users to describe functionality in natural language, add comments, or manipulate a visual canvas, while the agent translates those inputs into functional code, eliminating the need to write syntax manually.

How does Replit’s Agent 4 enable a post‑prompting development experience?

Agent 4 runs multiple parallel agents that handle separate tasks asynchronously, resolve merge conflicts, and integrate external skills such as Stripe or Salesforce, so users provide only high‑level goals; the system decomposes them, executes them across platforms, and delivers finished applications without granular prompts.

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.

Ergonomic Mechanical Keyboard For Programmers Recommended

Provides a tactile and comfortable typing experience for long sessions of 'vibe coding' and software development.

Amazon →

Ultrawide Computer Monitor For Multitasking

Offers the screen real estate needed to manage parallel agent workflows and visual design canvases simultaneously.

Amazon →

Book On Startup Growth And Strategy

Provides foundational knowledge on scaling companies and product-led growth, mirroring the themes discussed regarding YC and Replit's evolution.

Amazon →

Noise Cancelling Headphones For Deep Work

Helps maintain focus and flow state while building complex internal tools or managing AI-driven development projects.

Amazon →

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

Summarize another video

Full Transcript YouTube

Today I'm joined by Amjad Msad, the CEO
and co-founder of Replet. Replet is the
leading noode app builder uh for
consumers and enterprise. They just
recently raised their series D, a $400
million raise at a 9 billion valuation.
Amjad, thank you very much for joining
me and I'm excited to learn more about
Replet.
>> Thank you for having me.
>> Of course. Why don't you tell us a
little bit about what Replet is? Yeah.
So, um, you know, our ambition is that
anyone, no matter what level of skill
they have, anyone who can read and
write, basically, that's the skill that
that you need, can come in with an idea
and can leave with a an app that's
deployed, that's hosted, that's getting
traffic, that's can scale, and they
don't have to worry about any technical
aspect of the uh of building that thing.
Um, and it's been it's been a mission
we're on for for 10 years right now. So,
initially we solved the development
environment. You know, that was the
hardest thing is like setting up a
development environment. Then solve the
de deployment environment. Then what was
left is like the coding part was still
very intimidating for people. And
September 2024, Replet became the first
like what's called Vive coding product
where uh we abstracted away code
entirely. So there's like a coding agent
behind the scene, but you're just
interfacing with AI using natural
language. And more recently with agent
4, uh, also design interactions. So you
can write things, you can comment on
things, you can drag and drop things in
like a canvas. And we're thinking about
a lot of different modalities for how
people want to interact with agents
because I don't think it's always going
to be just text. I think at some point
it's going to be multimodal, maybe
video, maybe audio. But we want to
create a natural place where people can
express their ideas and those ideas can
turn almost magically into software that
is real software that it's not like a
toy software. Real software, secure
software, scalable software.
>> One of the things that's most
interesting to me about about your
company is that it really bridges the
the gap between a dev tool and and not I
I think it's the first dev tool I've
seen, but that's not marketed towards
engineers. like can you tell me more
about how you came to that uh decision
that you were going to do that and and
who is your primary user today?
>> Yeah, so you know I I started coding at
a very very young age but um I was
always interested in in the act of
creation. I was interested in
entrepreneurship. I built like my first
business when I was 13 14 and I always
thought that the developer tools were
getting in the way. It actually gotten
worse over time. So, you know, I
started, you know, coding on basic and
you just start the basic command line
interpreter and you can just like type a
little basic and that's good.
>> By the time I graduated from college,
like setting up, you know, a web app was
like a nightmare, right? Um, and so it
created this desire to just like build
tools that are that are more joyful,
more enriching, kind of focus on the act
of creation as opposed to the accidental
complexity of of of developer tools. And
so, uh, I started building a lot of
tools for myself. Uh, and eventually I,
you know, I built what would become the
kind of the first and browser IDE.
Initially it was like an open source
project. Later on I would also worked on
React and React Native. And I approached
every time I I built a developer tool
I've approached with the same thing. Can
you apply design sensibilities in the
same way that we you would work on a a
consumer app uh and that's been
successful in in many ways but
ultimately when we started replet the
goal was make programming accessible.
Later we updated our mission to create a
billion billion new developers.
As we progress in our in our mission and
in solving every part of the software
development life cycle, what we've
noticed is a lot of developers actually
like the pain.
>> They they like setting they the you know
setting up things and they like
configuring every aspect of it. It's
it's sort of like which is no knock
against that but like it's sort of like
a crafts person kind of liking to build
their own tools.
>> And what we've noticed that people that
are getting the most value out of a
product tend to be the more tech
adjacent ones. Maybe people product
managers have written code many years
ago but don't worry worry about the
development environment setup. Don't
want to worry about the deployment
setup. Um then designers, designers who
have ideas but are often blocked or
bottlenecked by engineers and they want
to be able to build their own ideas. Um
and later on as we layered on AI and the
product got better entrepreneurs and
when when I talked to these people it um
it reminded me of myself. Those are
people with with ideas, with passion,
with fire
>> in them,
>> but but but they're getting hamstrung by
by just like the need to learn to become
technical.
>> So at at some point in 2023,
um we just made it an explicit goal of
like we're not going after developers.
They're still developers. Like if you
walk around Replet today, it looks like
a dev tool companies. We're like
building for creators, right?
>> But they're not the traditional type of
developers. It's like a new generation
of developers that are coming up right
now because of AI. They're AI native
developers that are creating software
without having to worry about every
component in the system.
>> It's so interesting hearing you say this
because it it sort of reminds me I
learned to program on VB6
>> and and that's uh to me that's a real
you know this idea of the tools opening
up and making making it possible for
people that couldn't code. I couldn't
code. I learned to code in VB6 and and I
see sort of a very similar thing that
you're exploring with this.
>> Yeah, VB6 was better than setting up
React and Webpack
>> 100%. I remember the transition. Yes,
>> it's it's it got worse over time, which
is like now many things get worse in
life, but programming got worse and I
wanted
>> to bring it back,
>> make programming great again
essentially.
>> That's so interesting.
uh tell tell me about what people are
building today in in Replet.
>> Yeah, I mean there's a few different
categories. There's personal software,
there's enterprise software, and then
there's like entrepreneurs building
building products. One of the apps that
I really liked recently was talking to
uh a physical therapist uh and her
husband and she is a physical therapist
that's um built up so much know deep
knowledge in a subsection of physical
therapy that is like uh around fascia
and fascia release and she has all these
different methodologies and ways of
tracking uh the progress for her clients
and she wanted to build an app that is
very sophisticated. Mhm.
>> Uh like uh being able to take scans of
your body, being able to to scan your
range of motion um and to track that on
a 3D representation of your body. And
they spent hundreds of thousands of
dollars offshoring uh to developers
around the world.
>> And eventually they kind of were
frustrated by the entire process, took
matters into their own hands, built it
on Replet. And I when I saw that app, I
was like one of the best
>> like health tech apps I've ever seen.
And and they're not
>> Yeah. because the domain experts can now
actually build the product.
>> Yeah. People who are closest to the
problem can build the the the the
product they need. You know, I uh talked
to a founder the other day that's
building a SAS solution for
um for people who maintain pools like
and he grew up in a in a household where
their comp where their family business
is is a pool business. And so it's like
there's so much software to build in the
world. I also met a founder yesterday
who's building uh software for sports
clubs and he was showing me pictures of
the software they're using today which
is MS DOS based software.
>> Oh wow.
>> And so here in Silicon Valley we look
around us and we're like oh what else is
left to build? But but there's so many
walks of life. There's so many um things
that are kind of a blind spot for us.
And so suddenly when anyone can make
software a lot of parts of the economy
is just going to improve and that's like
I think a beautiful thing and a lot of
wealth creation is going to happen
productivity gain as well. Personal
software is also really cool like you
know talked to a lot of families that
built like you know healthcare software
like there's like a mom in Korea that
built software uh that helped manage a
very rare condition for her kid. Um a
lot of people build like personal like
uh you know healthcare software or like
tracking uh physical activity or
ingesting data from all their wearables
and and doing something with it.
>> A lot of software for families like a
mom built like a chore hero software
iPad that's like on the wall that shows
the kids how they're ranking on their
chores.
>> And then there's the enterprise use
case. And in enterprise it tends to be
two different use cases. One is uh
product development. So companies want
to move a lot faster on product
development. Everyone now is feeling the
pressure of of AI. We need to move
faster, faster, faster. And they're
realizing that it doesn't fall just on
their engineers. Their product people
can now build software. The designers
can build software. We hear from clients
uh like one of my favorite stories like
uh Whoop is a is a client of ours and
they told us like
>> the amount of ideas they can try has
grown by an order of magnitude
>> right like they used to you know get get
like a hundred ideas but are able to
only try five of them
>> right
>> but now they can try the 50 right
>> um so uh you know companies become a lot
more prolific they can release a lot new
features create new new business lines
new products And then there's like
internal tools and you know line of
business applications. Um there's a lot
of sales automations that that that is
happening at this company. Any any role
inside a company that's dealing with a
lot of data flow for example think of
revops that are the nexus of a lot of
different data flow from from their CRM
from their data people from gong from
and they want to be able to do things
with this data and typically they're
kind of bottlenecked by either
engineering resources or the different
SAS tools that they would want to buy.
The problem with SAS is you onboard a
new SAS tool that creates a new silo of
data that you can't really program.
>> So now they're taking matters into their
own hands and building things like you
know quote configurators and um and
being able to save the company hundreds
of thousands of dollars if not millions
of dollars on SAS tools and things like
that.
>> How how are you finding these users?
Like how do you market to them? How do
you get them to to try it out? And then
I I assume for enterprises it's got to
be a very complicated sales cycle like
you know getting getting into the
company like how does all that work?
Well, the cool thing about the the world
we're in right now and it's very similar
to developers. You know, I think the
insight that a lot of YC dev tool
companies has had like you know Stripe,
you know, page your duty. Um you uh you
uh developers were empowered to be able
to make decisions to bring software into
the company at some point like in the
last couple decades, right?
um that that shift is happening outside
of development right now where the
product manager, the designer, the
operations manager is empowered to bring
software. So the consumer use case, the
personal use case is often overlapping
with the work use case in that people
are playing with these tools on the
weekend. They're building their own
personal apps. They're like, "Oh,
>> look, the moment you understand that you
can solve a problem with code,
>> it changes your mind." like almost
there's like a neurological shift where
you start looking at the world
differently where you go around and
you're like oh I can solve this I can
fix this
>> I remember that when I learned to
program
>> yeah exactly and now that shift is
happening in a huge parts of the
population
>> and so it's you know PLG play I think is
still the gold standard just make a
product that's really good that people
want to recommend their to their friends
>> and make it easy to refer others you
know build a referral program do all of
that stuff
>> and then on the sales side, a lot of it
is very kind of sales assisted. So,
someone brings it to work, we talk to
them, they're the champion, we help
them. It's like, okay, what do you need
to convince your boss in order to bring
this to work? Let's work with them to do
a hackathon to bring create more
champions inside the company. Uh, let's
work with you. Let's teach you
leadership about AI. What? So, a lot of
what our sales people are doing is
evangelism, is education. So, it's it's
a different kind of sales motion, I
think, than the previous one. We still
have enterprise sales that's more top
down where a company comes to us, we're
like, "Hey, we're trying all the
different VIP coding tools. We heard
it's good for our business. Can you help
us kind of evaluate Replet?" And we go
into that and um and there we're often
winning because Replet has built um uh a
history of just like being uh super
trusted on security, on compliance, all
that basic enterprise stuff you still
you still have to do. Let's talk about
sort of what what the limits are today
of what you can do with a product like
Replet like what what kind of systems
can you build in replet and what do you
still need to have sort of traditional
uh software engineering approaches to
>> I can confidently say to any
entrepreneur out there that you can
build a SAS product a consumer product
um like an automation product on replic
comfortably
>> if you want to build like new cloud
platform or you want to build like a new
machine learning system. That's not
exactly what we're focused on today.
Some people still figure it out. Replet
is a versatile tool. We give you have a
virtual machine. We give you a general
purpose agent and if you have some
technical knowledge, you can bring that
and you can build sophisticated things.
>> But if you want to build things entirely
vibe coded without worrying about the
technical details, we have so many
examples on our website and what we talk
about uh in our marketing. uh we you
know that there's so many success
stories now that I can confidently tell
an entrepreneur with no technical
knowledge that they can build right
>> uh a piece of software so there's a lot
of consumer apps there are a lot of um
there are a lot of uh you know vertical
SAS products there are people that are
starting uh like replet native agencies
I was talking to
uh someone from Iceland yesterday and he
was telling me they're getting so much
business because they're 60 to 70%
cheaper
and more effective than traditional
agencies and they're all vioding on
replet. So they they get a client that
that client wants an application not
everyone knows that they can build the
application themselves. So they're going
in it's probably a period of time and
they're able to make a lot of money that
that way. you know, the internal tools,
automations, those tend to be kind of
fairly simple uh programs like you're
building an internal org chart, you're
building um a CPQ system like a quote
configurator where it's like pulling
from Hopspot or Salesforce. We have a
lot of MCP and integrations that allows
uh to do that. Um
>> how how do those integrations work? Like
have you gone through and built
integrations to all Okay.
>> Yes. Yes. We spent a lot of times
partnering with companies building those
integrations and now there's the what's
happening with the skills call
revolution where you know companies are
putting out skills and MCPs and we just
vet them and integrating into into the
product and so as you're talking to
replet and you're like I want to
integrate stripe and so we already built
a set of skills and sometimes code and
things like that and we'll search a
database we'll bring all that in and now
it's in context and know the agent is
Like it's sort of like Neo in the Matrix
where like you you download a new skill
and you're like, "Oh, I know how to fly
a helicopter now." So it's so cool to
see the agent is like, "Oh, let me
search this set of skills that I have
and suddenly it knows how to do that."
But we spend a lot of time just making
sure they're secure and they're safe.
>> That makes sense. Yeah. Having built a
dev tool community, like how is that
different? How do you build the sort of
like ground swell of support to pull
yourself into those organizations? Is it
different than how it how you would have
done it traditionally?
It's it's a little different in that you
need to show what's possible. I think
with with developers with more
traditional, you know, CS trained
developers, they know what's possible.
They know they'll read your docs.
They'll figure out what's possible.
They're on hacker news all the time and
they'll generally kind of are much more
resourceful in knowing what's possible.
With replet, there's a lot of education
that needs to be done. So we have like a
Devril uh team, but they're not the
traditional Devril. They're they're more
like educators.
>> Yeah.
>> They're like, you know, going and
spending a lot of times just teaching
people what's what's possible. Our
documentation, I think, is a lot simpler
than your typical dev tool
documentation. So you have to be really
good on content. you know, Stripe got,
you know, popular for this you being
really good on documentation
with if you're building a dev tools for
non-developers, you have to go above and
beyond on content. You have to produce a
lot of video content as well.
>> Um, also the agent itself needs to be
imbued with characteristics that is able
to talk to non-developers and tell them
what's possible and be a brainstorming
partner. Also this is why we also moved
from purely just prompting to also uh a
canvas and more visual interface so that
people can explore things uh more easily
and we have buttons where it says like
create different variations of this and
so we're adding a lot of different
shortcuts just to show people what's
possible on the enterprise side when we
go and we talk to leaders we're like
reserve any judgment don't pay for
replet we're going to come in bring the
group that's most excited about AI and
we're going to do a hackathon How do you
when you see someone how are you like
yes this is going to be the person who's
going to be so excited to use this? We
think a lot about that because you know
typically when you think about ICPs
you're thinking about oh that's someone
who's went to school for this degree or
someone who is a product manager or so
which is still true you can do that but
often times because we're able to sell
to sales departments marketing product
design
the set of traits that we think makes it
like the champion the person who's most
excited about replet who's going to
spread it inside the organization who's
going to build who's going to do the
educ ation evangelism. They tend to be
very entrepreneurial.
>> Yeah.
>> They're the kind of person that could
start their own company, but they're
able to kind of be very influential
internally. They're kind of like what YC
talks about like PGSA, like resourceful,
someone who's not going to get blocked,
who's going to figure out what other AI
tool I need to integrate, what can I
like go learn in order to figure it out.
So that entrepreneurial kind of founder
mindset I think is very important.
>> Got it. I mean while we're on the topic
of YC, you know, I'd love to hear about
your your experience here like how how
did YC influence you know what how
Replet started and what it became?
>> The main realization from YC is how much
you can get done in 3 months. So when we
when we came in and Sam stood uh uh when
he was still running it, Sam stood in
front of the batch and said for the next
three months tell your friend you're
going to be missing. You're not going to
be able to help them move. you're not
going to be able you'll come back into
their lives later, but for the next
three months, you need to be
hyperfocused on this company and you're
going to be able to achieve great things
if you really go intensely into it. When
we did YC, we had the we had a
whiteboard
>> and a countdown to to demo day.
>> Mhm.
>> Uh and a number of like a just a very
simple list of things we need to
achieve. Yep.
>> And every day we wake up, we erase that
number and we change it to the
countdown. Um and and when replet got
got into YC, it was still a ripple.
That's the name, right? It was still
just like a command line with like you
just type a little bit of code and run
it.
>> We exited YC, we already had like um web
development, we had like the initial
aspect of hosting, we had code
intellisense, we had like so many sort
of IDE features like got a lot more
feature complete just in three months.
And that that is an empower thing sort
of what we talked about with
programming.
uh just the fact that like oh you can be
so intensely focused and you can work so
hard and achieve so much in three months
and now every agent release there is
it's not three months it's more like
four weeks where we bring everyone
because some people are still kind of
remote or other offices bring everyone
to the office we provide breakfast lunch
and dinner coffee 24/7 and we're like
we're just going to hit this like very
very ambitious goal and that's like the
YC mentality. Another thing is the
compound growth, especially when you're
spinning up a new product or a new
initiative. This idea of like I don't
know where PG came up with the number
seven, but like 7% week overweek growth
>> is like a very good way to bootstrap a
new new product line. So, we're
constantly doing sort of we're
constantly going back to the YC basics
and doing them and all of that we
learned here.
>> Tell me about um how how did YC change
your life?
There are a few ways. One is before we
got into IC, we couldn't raise all that
much money. We raised like maybe
$500,000
and you know VCs did not want to meet
us. It was like the doors were not
really open for us. Uh after YC you
immediately first of all demo day you
got introduced to a lot of VCs. Um, and
then we were lucky because we got
rejected from YC four three or four
times and then we got invited to do YC
because Paul and Sam saw us on Hacker
News.
>> Yeah.
>> Uh, and so we got a direct relationship
with Paul and Sam and then at the end I
was like I want an intro to Mark and
Dreon
>> and you can still ask the partners here
like they they'll they'll try a lot to
to get you intros but I was like bold
enough to ask for that and so they got
me an intro to him. I went and had
breakfast at his house and pitched
Replet and A16Z ended up leading our
seed round.
>> Nice.
>> Uh and and like my network just expanded
tremendously after getting into YC and I
don't think we would have been as
successful without YC. Perhaps we would
have quit
>> if we couldn't fund raise or
>> so it gave it gave our our company life.
>> Yeah. Yeah. For us it was very similar.
We were complete outsiders and I think
having
>> having you know somebody sort of welcome
you in made made an enormous difference.
>> Yeah 100%.
>> So you guys just announced uh agent 4.
So what is that like the fourth major
revision of your um of your product?
>> Yeah. So we kind of an act of trying to
predict the future and based on what
we've seen at the time starting in kind
of 2024 we thought that broadly
the AI capabilities have massive step
changes uh twice a year.
>> Mhm. Uh so if you think about 2025
um there was the like vibe coding
revolution that happened in earlier 2025
and then there was the the you know the
the autonomous revolution that happened
in late 2025 with Opus 4.6 that led to
openclaw and and things like that. Uh
but I can tell you the same thing
happened in 2024. Mid 2024 cloud came
out and um uh Cloud Sonnet and for the
first time we could generate like a lot
of code as opposed to like GPT3.5 which
has that that laziness component. Um and
then late uh 2024 we started seeing
initial signs of the labs kind of um
doing long horizon reasoning. And so the
>> so that observation we're like okay we
want to align our road map to AI
capabilities.
>> Um and so every six months we release a
new uh a new agent version. Um and it's
an active predicting what's possible.
was also pushing the edge of what's
possible. So, for example, agent 3 was
the most autonomous agent on the market.
We want we knew that autonomy was coming
and we're like, okay, we want to be able
to run the agent for like two, three,
four hours. People should be able to put
put in a big prompt, go to launch, come
back and and see the software fully
fully built. And so, okay, what do we
need to do to update our platform? for
example, like we had we had to rewrite
everything on the back end to have these
long running containers in the
background doing work while the user is
not in front of the computer. And so we
did all of that work and although
autonomy didn't arrive until like maybe
like like true autonomy until like
uh November and and December, Replet
agent had demonstrated where the world
is headed by by September,
>> right?
>> With agent 4,
there were a few things. one parallel
agents. We thought that was finally
potentially possible to do parallel
agents. The the the thing that sucks
about autonomy is that you can put in a
big prompt and you were kind of sitting
back and just like watching it work like
what do you do next?
>> Okay, so that question, what do you do
next? You should be able to design,
should be able to kick off other type of
work. Um should be able to chat with
your agent and and plan for other
things. And so we wanted a more
asynchronous nature to the product. Uh
so we started building towards a multi-
sort of Asian architecture and power
Asian architecture. We had to solve
merge conflicts and so many things to
make that work.
>> Also as as the agent is building we want
to also unblock you from designing. So
we also had a kind of a more
asynchronous design agent and we're like
okay what what is the best interface for
that? So we designed this canvas.
>> Yeah. So now we have a built-in design
capability inside replet to um to be
able to explore the next page you're
going to build or the next feature we're
going to build while the agent is
building and then once you're ready
you'll kick that off into another thread
>> and it starts. So people are now sitting
in front of replet and just experiencing
the state of flow.
>> Yeah
>> because agents are slow that's fine but
they can work in the background
>> and finally uh teamwork. Once you solve
parallel agents, you've also solved
teamwork because every time someone
jumps into the session, you can start an
entirely new, you can fork entirely new
VM for them and they can work in
parallel to you.
>> And also because the orchestrator knows
how to subdivide tasks. You can also
prompt in the same chat window and we'll
figure out how to make all that work.
And because we had the canvas, it was
it's it's such a joy to just see other
cursors and people the the product just
becomes a lot a lot more live. So all
these components became uh Asian for oh
final one I forgot.
>> We wanted it so that when you make a
mobile app when you make a website when
you make uh a deck when you make a video
all of that should have the context of
your project. previously on replet and
now every other tool. It's basically you
need different tool for a website,
different tool for mobile app. So now
replet like you built a really cool web
app. It's like maybe I maybe I need an
app. You could just say make a mobile
app.
>> We'll like generate a mobile app. It
will lay it out on the canvas. When you
hit deploy, it deploys your web app to
the web. It deploys your app to test
flight or the uh or uh or your Android,
whatever it is. So, so now you can run
your entire company on Replet.
>> What What kind of skills do people need
to develop to to be able to make the
most of of products like yours? Like
what what is the you know what what do
people need to get practice at? Like in
your mind, is it is it people have to
get good at prompting or is it more that
the system is going to get better at
just understanding whatever it is you're
asking it to do?
>> I actually think we're headed to like a
postprompting world.
Um and you can you can sense that in the
sort of open claw and all the
derivatives and the way people are using
it um they're more giving it high level
goals like optimize my marketing funnel
you know things like that so prompting
will be there it's like a layers of
skills prompting will be there whenever
you want to do more interactive type
work but also you should be able to give
your agent highle commands so I think
perhaps agent five or maybe sooner you
should be able to tell replet every day
build me a SAS company and like try to
market it and see what works and make me
some revenue like you should be able to
give it like
>> that's pretty high level.
>> Yes,
>> I think we're almost there uh in order
to be able to do that. So what kind of
skills do you need? Understanding what
is possible is going to be important. So
playing with these tools a lot having
this like playful mindset of adopting
these tools, playing around with it,
being plugged in
>> is super important. At some point, I
think it was kind of a drawback or a
flaw to be someone who's like constantly
reading the news or being online,
>> but actually now it's very important to
know what's happening, what's coming
down the line.
>> Yeah.
>> Um I think um not giving up is
important. Like the thing that you try
to do today that they cannot do, try it
again in a month.
>> Yeah. And I tell some replet users if if
whatever you're trying to build replica
and build today, try it in a couple
weeks. It might be able to build it. Um,
and so that that mindset of like I'm
going to keep trying is is is important.
Idea generation is still going to be
important. Obviously, we're going to get
to a point where where AI is going to be
really good at helping you generate
ideas. Um, but being generative, just
constantly thinking about problems you
want to solve. Um, and just being
creative and like figuring out what what
what what does the world need right now?
And um, and being generative is
important because let's say you're a
small-time entrepreneur. Um, like you
know, Peter Levelvel's type of
entrepreneur or like this guy is very
famous on Twitter who built these
products. oftentimes the products go
through cycles like we build like a
product that is really good for that
moment and you can generate like a
couple million dollars but then that
product is no longer relevant. So you
need to be generative and creative in
order to continuously do that and you
can make millions of dollars doing that.
>> If you were starting Replet today,
what would you do differently?
>> I mean, I made a lot of mistakes along
the way. Uh, so if I was starting Replet
today, I'd probably hope not to make as
many mistakes. So, you know, culture is
very important. You know, at some point
we screw that up and we had to do we had
to do a reset and lay off and all of
that stuff. Um
I think being really honest with
yourself about like product market fit
is very important. It's very easy to
delude yourself.
>> Getting any kind of user is an amazing
achievement. Getting any money kind of
money from users is also an amazing
achievement and you should celebrate
every one of those moments. But true
product market fit is entire entirely
different. It's like an explosive thing.
>> Yes. And so being honest about that
because we've had periods we're like,
"Oh, it looks like that's successful.
Maybe it's working
>> and like you keep going down that path."
But in reality, you should have changed
directions a little earlier.
>> Yeah. When it works, it really works and
you know it.
>> In terms of AI development, AI
technology, what is something that you
were waiting for where you're like,
"This is going to unlock so much." You
know, I'd been waiting for computer use
models to to get better. They they're
slowly getting better. They're like one
of the things that is actually kind of
disappointing.
>> It's surprising to me that it's so hard
to build them because you'd think that
it'd be the easiest thing in the world
to get data for. And do you do you have
a sense of why that is?
>> Well, language is is a lot easier. uh
language also like much more can be
compressed down into like some kind of
highdimensional space a lot easier than
like like a video feed right
>> but still like if we can make progress
on self-driving surely we can make
progress on moving the mouse and
clicking on things so I think it's a
little bit of a of a mystery that being
said coding in coding agents turned to
be a workaround and hack Because a lot
of things that you could do in front of
the computer you could do with code
including you know scripting in Excel
sheet right the reason Excel sheet
agents got better is because coding got
better. The reason like you know some
commerce agents are starting to work is
because they can call APIs and they can
so coding turned out to be a bit of a
hack or workaround computer use agents
but there's still a lot of software in
the world that's like very legacy
software that computer use agents would
would be really good at also when you're
when you're creating a vibe coding
platform like like Replet you want to be
able to test the apps that people are
making and you want to be able to be
discerning as well like things that are
functionally right but also UX is not
very well you'll be able to give the
user feedback on that
>> but the models are not very good at that
so we spend a lot of time
>> taste question I take it
>> taste question so we spent a lot of time
prompting and augmenting
computer use models in order to to make
them good at testing because replet has
like a testing testing agent
um so that's one that I think will
unlock a lot of different capabilities
once we get to computer use agents.
Um
I think like
yeah I think it's become a buzzword but
like continual learning like the way
we're hacking around it right now is
we're writing to files
>> like the agent will learn some kind of
skill and like we'll write a skill MD
file or something like that but true
learning on the job is like not been
unlocked yet.
>> Yeah. Yeah. Uh, and so for us to be able
to like deploy an agent inside an
organization and for that agent just to
continues to get better for that org
itself is is is such a powerful thing.
But but that just still seems far away.
You know, I I I think, you know, one one
of the common themes I'm hearing here is
is like we're hey, we're going to have
like or we already have today with
Replet like agents to build software and
and you talked about how we're
increasing the abstraction until finally
you can just say, hey, I I want a
company that just makes money.
>> How how is this going to work? Like how
what does that leave left for people?
And and what does the company of the
future look like if it's all just agents
all the way down?
Um I think I think the company of the
future is made of builders
and sales people broadly and sales will
change and sales will be more like let's
help companies transform. Let's use the
technology that we're building here to
help companies transform. So when sales
people are more like I mentioned it
earlier like more like evangelists more
like educators
>> but I don't think that part is going to
go away because a lot of other companies
will want to talk to someone will want
to learn from someone right that's how a
lot of people learn um and they want you
they trust other humans and so the sales
probably part is like one of the more
defensible jobs
>> interesting
>> um and then builders uh there's always
more to automate. There's always I think
our our job will continuously like get
higher and higher level. I mean that's
already the case that with computers we
became a lot more higher levels like
computers were literally humans right
like there was like a you know bunch of
people like doing computations of
numbers
>> tables of numbers and then we're like we
you take that entire room and you put it
in a box that's computer and now there's
my job as like the operator of the
computer to use a computer for
productive use.
uh and there was a lot of still manual
work in order to make that computer do
interesting things uh software and now
we have an agent that's using that
computer to use so you know different
layers of abstractions I don't know when
you have a fully autonomous company
maybe at some point there's certain uh
styles of companies that could be fully
autonomous but I think you need business
generalists that understand customers,
understand what people need, understand
the economy, understand where the world
is headed, understand AI technology,
what where have some vision and you want
to give them so here's like the abstract
vision of a future company. It's like
almost everyone is a founder. They wake
up in the morning and they think how can
I make the company more successful? How
can I make the company make more
revenue? Mhm.
>> And then they go around the company
finding problems to solve and then
creating or deputizing agents in order
to go solve these these these problems.
We're already seeing that we actually
have a team like a vibe coding and
resident team at Replet
>> and they have a very vague mission
unlike like a product team or sales team
where it's very clear like go around the
company make it better. So they went to
the support team and was like, "Okay,
you know, you're using Zenex and all
these other tools. What are the main
problems?" And we're like, "Well,
we don't have a good way to prioritize
our uh support queue. Um, for example,
like there are customers that are paying
much more than other customers and there
are customers that, you know, have more
urgent tickets." And so they built like
a way to to visualize that visualize
that and and uh a way to like create
more priority cues. Um and they spent
some time with with the support team and
the seat score started going up
>> and then they go around they go around
their HR team. What are the problems?
Well, onboarding is a problem. There's
no place for people to to know all the
benefits and everything that we have.
Well, let's build an HR internal HR
platform. So they they do that and I
think that kind of role is more
is more where the future is headed and
more and more people inside the company
will be more generalist entrepreneurs
that are trying to make the business
successful.
>> Well, thank you very much for joining
me. A pleasure to have you here at YC. I
>> appreciate it. Thank you.

Help & FAQ

Beyond Bigger Models: Recursion As The Next Scaling Law In AI

Y Combinator

May 01, 2026

User Base & Market

Product Strategy & AI Integration

Mechanisms

Business & Growth

The Future of Work

Takeaways

Frequently Asked Questions

What is vibe coding and how does it differ from traditional programming?

How does Replit’s Agent 4 enable a post‑prompting development experience?

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

How does Replit’s Agent 4 enable a post‑prompting development experience?