Teaching Agents to Code: Bleeding-Edge AI at Replit

Thanks hey everybody we're going to get started so uh thank you so much for coming out um and thanks for joining us uh on Tuesday at 816 started if you can hear me clap once if you can hear me clap twice over to you he's been doing this for a bit longer than I have uh so he's a bit better at it than I am but uh thanks so much for joining us on this Tuesday at a16 Tech week and uh I definitely want To get started by thanking first Andre and horn for putting on

Tech week and helping us amplify this event helping you all uh be aware of what we're doing here tonight um but then also a big shout out to Y combinator um for this amazing space uh I've never listened to a tech talk in an old like power generation company building with the giant turbine it's pretty cool um and also shout out to Maria at YC combinator helping us set everything up shout out To Beth at repet shout out to the entire uh support staff for helping this all come together but I'm Mt at repet I

work in developer Rel I kick it over to what you all are here for and that's to see Mel and Jared we're I'm going to give a Qui demo talk a bit about repet it'll be about 10 minutes long um just so you know what to expect and I'm going to talk about what repet is I'm going to talk about what the repet agent is and then I'm going to Talk about why you should care um and specifically I'll explain kind of how this differs from traditional ways of building but also AI tools one shot

AI tools uh that exist today um so this is repet just a very broad overview you know at a glance if you're non-technical that might be a bit confusing but if you are technical it might look a bit but like an IDE but it's a lot more than that and we'll talk about it today so if you're not technical the main takeaway Here is that repet is an all-in-one environment for building and deploying apps with code this is a code First Tool at the moment but we take care of the setup for you right so

you can focus on building um and we provide AI to simplify development the replate agent was released a couple weeks ago I'm going to show you guys a demo of that but then there are also inline AI tools for editing and modifying your code um and the most important thing I think is That we take care of deployment and environment configuration which means you can go from Local Host right you can go from running something to live it being deployed on the internet as fast as possible and we'll talk a bit about what that means

um but I think an important question is right so what this is a big fancy tool right there's all this code running it means that you can install any language in seconds so this GIF here is installing Python and node In its own isolated environment in this reple which you can think about if you're we'll talk about if you're more technical you think can think about it like a container but if you're not technical just think about it like it's its own place to write code it's its own separate computer so you don't have to

worry about configuring virtual environments or even figuring out what those are you can focus on writing code and Building Things um and that means You can install any language in seconds you don't have to worry about secrets we can configure databases for you you can collaborate in real time deploy those projects in two clicks and do it with the help of AI now if you are technical well actually show of hands how many people here are would consider themselves technical sick I'm not actually asking for this presentation I'm just going to tell jokes later and

you know I need to know if you're Technical or not um so if you are technical rep's a cloud-based IDE it's built on instant virtual environments or containers our head of engineering you know like your CTO is here he's probably going to correct something about what I'm saying but that's the gist of it essentially it's an auntu VM you can spin it up it has its own full file system and then we drop the repet editor on top of that you can use our workspace or you can connect via SSH with your f Um editor

there uh you can add seress postgress it's backed by Neon if you're familiar um or object storage in a single click and uh you can deploy a server horizontally autoscaling app Chron job or static site in minutes and that's sort of the magic of Reit and so this so what if you're technical right there's development deployment parity which means your development environment this repple that you're developing in a virtual container is exactly the same as Your deployment environment which means if it runs on repet it deploys on repit and I haven't found another cloud provider

that does that other Cloud providers are great right but I don't see the same functionality there it also means you can onboard teammates in minutes so if you're technical here right and maybe you have some semi-technical co-workers that are comfortable writing code they're comfortable using git but they're not Super comfortable configuring all this stuff on their own you could create an environment in repet and because when you Fork a reple you're cloning that entire environment you could just share that with your semi-technical teammates or onboard analysts or um you know others that want to get

started with code but don't really have the agency to do that themselves you can really uh facilitate that process and democratize um building applications across your Entire team uh like I said before you can deploy in two clicks and we now have repet teams uh that supports private uh deployments um backed by SSO um and again you can do that with the help of AI so um enough about that we're going to talk about repet agent so if you're non- technical repet agent is everything that's great about repet but it's with an AI system that

can configure environment edit multiple files and deploy your application um right this is Different from things like CLA or VZ at the moment uh and if you're Tech if you're technical I'm going to say agentic I promise I won't say agentic too many more times other people might say it it's in the title of this uh this talk um but it uses the latest llms and has access to all of the underlying files and tools in the repet workspace and again that's different because if you build something with uh a one-hot uh llm tool um

you don't have access to That you don't actually know what's going on and it's probably like on Rails some sort of very simple react application that's changing quickly but the repet agent has access to databases it has access to will have access to object storage and all of the um other great things in repet so I'm going to run through a quick demo here um this is not live I'm not that impressive to do a live uh AI demo but here we have uh a very simple app being spun up um and What we're going

to do is ask ask the agent to create a restaurant um voting app so everybody can go on your team can vote where they want to go for lunch and you could imagine this being uh more complex or an internal tool but immediate immediately the agent is jumping in and not just writing code right but configuring an entire environment so the first thing it did was create a postgress database now it's going through writing um python files Also wrote a static folder for HTML other assets for a simple um vanilla JS app um and now

it's installing packages and our engineering team recently added UV support which speeds that up I think 2 to 4X uh so the point I really want to drive home here is more than just creating an application it's creating an environment and then giving you that environment and enabling you to do really whatever you want with it whether that's continue to build share it with Others um or really anything you can think of uh this is also unedited I think I SP sped up like a couple steps but this is end to end exactly what you'd

see if you were building with the repet agent so once the AG done it's going to uh spin up this app um you can see it's making some changes to the database there um but at the end of the day we should have a working uh restaurant voting app in some in a couple seconds I talked a little too Fast so there you go and uh I can test this out first I'll say yes that does look good uh and we can try and add a restaurant awesome it's almost like I knew that was going

to work and so right this is different because the data will persist and it's backed by postgress this this data is being ridden to a database um and so we can have confidence that this is a production application with replit you have four different deployment types you can Configure those in just a couple clicks and then as I've been trying to explain the entire environment is going to be deployed uh instantly um using repet deployments um and there's a little call out for the private deployments which are available in repit teams which you can go

check out at rep.com teams we also have our head of sales here tonight I'm sure he loves to talk about it um or you can find me uh and so I did speed up the deployment step that usually takes Two or three minutes but taking this environment the entire environment packaging it up putting it live on the internet at its own URL and then just like that um it is available so this really end to end I would say right takes maybe five or 10 more minutes than I demoed here because this is a live

demo with the agent but I do want to illustrate that this really is what the Reet agent is and as Michaela likes to say I'm sure he'll talk about this is The worst it'll ever be right it's only getting better from here as the underlying Technologies improve um so that's all I have for you that's an intro to repet and the repet agent again I'm Matt at repet workan developer relations if you'd like I'll be around tonight come find me say hi if you can't you can find me on X or LinkedIn above also we

are hiring um so come find to me talk to me about that uh and since we're mostly technical here um how do two git Commits resolve an argument some people are giggling I don't know if that means they get it they hash it out all right that's it for me I'll let I'll let you guys get to the real talk so one of the things I'd love to start with is like what's happened so for for folks who don't realize um you guys just launched this was almost exactly three weeks ago now right so three

weeks ago and this is the first I Mean there are a lot of people who are working on like coding AI agents of various flavors but this this is the first one that can actually build a complete web app and deploy it by itself with no human intervention required that's ever been launched to the public this is completely new yes yes um there's been there's been a lot of uh sort of like um semi agentic things uh and they're great things like there's like vzo there's um you know cursor Composer feature uh but but they

they kind of uh are a lot narrower and when we were designing uh rapet agent uh you know we we took this like do you know this toothbrush test um the idea is that um you you want to be able to design something that people want to use every day and so anytime you have an idea you should be able to just go to repid Ag and and for it to do it so if um if that's the case that it needs it needs to be able to talk to the cloud it needs To be

able to store things in the database it needs to be able to deploy to the cloud and it needs to be really full stack and so that's been the the driving thing for us and the other cool thing is that even though repet has like millions and millions of users you didn't like do some super limited Alpha release where you need to be one of the chosen people to use this you just like put this up and let anyone who's willing to pay just like pay and start using it Right yeah do you want to

talk about that yeah like like day one you're just like go to town folks I might not be here if the launch going the right way no I'm just kidding um yeah I mean we did it the way because ultimately it's better to ship and get feedback from the users and we learned so much in the past few weeks that it wouldn't have made much sense for us to do like a very uh small release only to a few folks we Definitely did a lot of dog fooding internally and we learn a lot by doing

that but eventually you just got to ship and see what people think about it um when you launch something at rep means that you get a lot of love and a lot of backlash because the community is huge and we didn't make every single use are happy but we also made a lot of people full in love with a product they never seen in their life so I I don't regret yall do this big launch okay so three Weeks ago you took this like totally new thing that the world had never seen before and just

like threw open the doors and let your millions of users just like bang on it like what's happened what have you learned yeah I mean the most the most exciting thing is that people who have never coded before people who uh don't don't have the inclination or talent for coding can now make things and experience the feeling of of coders so Programmers here in the audience do you remember like the first time you made something it feels great right there are a lot of people that don't experience that and I think that the most exciting

thing was when people like you know uh there's a very famous designer that tweeted something a couple days ago Andrew Wilkinson who runs the tiny firm who owns a bunch of design ass sets like uh dribble and things like that uh he says I've been a designer all my life And uh at every stage of my life I had to go to developers to ask them to prototype even simple things and for the first time in my life I am able to to build build these things uh by myself and I think that's that's really

profound and the next tweet was like I need more qua can you give me more and the Tweet after that is like make the design better I'm a designer yeah so there's a lot for us to to work on but but but I would say for me at Least the the most inspiring thing is that is that it's empowering do you have some specific use cases specific user stories that have really been like oh wow like I would have never expected that to happen yeah um my my favorite one is uh uh there is uh

a a small company of real estate agents in Ohio that are using the product to build like back office automations and processes these are people that are not really connected to the Silicon Valley bubble Right so this is sort of escape the Silicon Valley bubble and now empowering normal everyday people to run their businesses more efficiently be because of rapet agent and and that that's one that that I'm excited about do you have one I changed my evening gits instead of Doom scrolling on X I'm actually watching videos of Rapid agent on YouTube and my

my favorite one has been I think the is like a retired guy who has built an up to collect all his Memories like all his pictures and put them on a map with pins and kind of rebuild the history of his own family and you could see at the end when it actually worked it was almost like about to you know start to weep and cry I know that was very emotional like figuring out that to empower people to build something that they wanted to do for literally decades I think it's this individual been trying

15 years 15 years exactly it's my my was my highlight so We we started thinking about this idea of personal software so obviously there's like business automation software which is which is going to be really great but the idea of personal software is that you want to make software for your family for your kids for yourself another another one I saw was like basically Nobody Does that because it's way too expensive to make software so like nobody writes software for themselves except you know some some Like hardcore programmers do but like regular people don't write

their own software and I I feel like we're at the at the precipice of this so this guy wanted to teach his kids about color mixing how you can mix RGB to create any any any kind of color he asked the agent to do that they created the app really quickly and his kids just had a blast with it um and so I think personal software is probably going to be as big as personal Computing uh you know 1984 was like when the Mac was launched and like personal computer became mainstream maybe 2024 is like

when personal software becomes mainstream and if that sounds crazy to you remember the personal computer sounded crazy to most people at the time it was invented to the idea that people would have a computer in their homes like why would anybody need a computer in their homes was what most people thought yeah another mind-blowing thing Is like a large percentage of users use it on their phone like use ret anyone here has a ring on their phone we have a few people here look nice that's awesome do you guys write code on your phone yeah

wow that kind of blows my mind did you imagine this when we were going through YC oh you you too oh there we go repet out over there okay oh you deploy from your phone oh That makes sense you can like check up on the on the on the deployment St you don't have to be like tied to your computer the whole time W I'm not going to tell your employer go um okay let's let's talk a little bit about the history so like a year ago repet had already been around for how many years

oh almost almost a decade yeah yeah almost a decade it had a very large very successful business Incidentally raise your hand if you've used repet oh my God it's like every hand thank you do you do you is it is it still weird to walk into a room and like everyone is one of your users it's still surprising that it works okay app works that's great yeah so so a year ago you had this very large like really successful business and you decided to make a really big bet on basically a totally new product Direction

on all this LM stuff you didn't have to do that you could just kept operating repet like what made you decide to place this you know to to allocate such a large percentage of your company's scarce resources towards this new thing yeah so so um uh like like I think from um basically the first time I saw a large language model the idea of being able to do actions and not just uh complete text Was something that that um that really captured my attention uh early 22 I think I wrote this huge threat about agents

uh which is now like more than three years old about how agents are going to really revolutionize software it's gonna it's G to make it a lot easier for people to to to write program 20222 this pre-chat GPT uh yeah no one was talking about agents back then no one was talking yeah someone can blow it up but um uh someone else was Thinking about agents which is which is Michel and actually uh we met Michel when we were uh starting to talk to to Google um and do you want to tell them about the

demo that you you showed us yeah so I I was in the team build palm and I think before we even release it a sham J demo where on repet I was connecting to services on gcp and generating all the code on the Fly and I was taking slides from one of your I think from raelon and passing them Basically through a model that did OCR and giving that back to an API and it was all automated on RIT so that was the sort of the first glimpse of oh you can do these agent orchestration

things with the with the cloud um and then we tried something like that internally but but but like it wasn't there the context window was especially very very small around 22 early early 23 um and then uh mcky uh joined and and wrote the manifesto if you want to talk about that Yeah mine is much younger only 15 months old but if you go through that you will read basically all the high level details of the agent that we just shipped few weeks ago and I think it was like a slow role in the sense

that you know for many months we build infrastructure and every now and then we try to see if you know we're almost there to really go Full Steam on the agent the reason why we didn't do it is models were simply not there yet and I Think you know you see yourself with all the other companies that are building similar products not only on coding um I think there's been like a step function improvement with the latest open anthropic models and the moment we started to see that our demos were working well then more and

more people in the company R behind it and the rest is which model was the big unlock my favorite lock has been CLA 3.5 suet in terms of coding capabilities Because we're building a coding agent but I think the the latest Frontier models in 2024 are really what unlock agents as a whole and a lot the fact that the community is building in public the llm research community of course is all entrenched into big laabs but most of the agent Builders are here and they WR about it and and can you talk a bit about

what it took to build this like this thing that we just built how many people worked on it how long did it take Them how was the actual process to build it like yeah so so April uh April this year maybe March uh uh you know when one of our uh uh people on on Michael's team on the AI team Jen is Jen here y there you go wrote the Prototype and remember showing it to me and I I literally saw this like baby agent on his on his Local Host um and and Micha and

I decided that the time was right that right now based on what we saw that we can really invest in it and we tend To not dabble when we when we really like something so we took I would say like half the engineering team yeah yeah roughly like half the engineering team maybe a third of the company into into a room uh I think in April and we sat down and I said okay this is the agent task force and that was half the company yeah so for folks who haven't run a company the scale

of replic it's very unusual to for a company that has a product that's working really well and it's millions And millions of users are growing really fast to take half the company that was working on the current product and H have them instead work on something that's like a prototype on like Jen's Local Host yes and a lot of people disagreed uh internally uh some people left because of that but it's like uh yeah you have to do it once you see the future like I think you have to you have to really bet on

that and and uh and we Did and um it was it was really uh it was really awesome it was really chaotic there were some tears there was no blood luckily not at all so smell War rooms though yes but but it was really intense it was um it was almost like YC again uh and at ret we like to run these like mini YC programs almost we like what we learned from YC is like you can go really really hard for two to three months at a time at least you know uh no one

can keep it Year long and so we went went really hard for two to three months Michaela every day would like would like uh decrement the counter to the deadline and the deadline was pushed maybe two or three times uh but uh and I would say the last three weeks were the most inspiring three weeks of my career because um we changed the architecture fundamentally the AI team came up and you can talk about it in a second came came up with with a new architecture Which changed the ux fundamentally essentially the last three weeks

were as intense as the previous three months and that that allowed us to to launch on that day how did you change the architecture we wanted to build something even more ambitious I would say we set a deadline a month before the deadline we felt okay this is kind of working but maybe it's not as reliable as we wanted can we do more can we do something Bolder and rather than working On small iterative Improvement uh we went on the White and said hey can we do something much crazier and the the key Insight is

the longer our Legend trajectory runs as I was mentioning before the more probability you have that it's going to go the wrong way so a trick there rather than waiting for models to become better is to do as much as you can with the fewest possible amount of steps and we basically rehold the entire thing and if you see right Now we are very fast creating the initial prototype except when the Wi-Fi is not working well we need to make it more resilient to to Wi-Fi issues that's not an AI problem though just to make

sure yeah and it was um yeah but it was worth I think everyone in the agent task force eventually was was behind it but you know that that's that's intensity that Jad was talking about if you show up on a Mond and say we spend the weekend rethinking the entire Architecture you ruffle some feathers but it's worth it okay so speaking of the architecture um I'll bet a bunch of people here are working on AI agents yeah uh raise your hand if you're like working on an AI agent or thinking about working in the AI

agent space okay quite a few people all right so for those folks can you talk about like the Tactical architecture like how does the thing work under the hood and what what sort of like advice could you drop for People who want to build their own AI agents and have them work as well as as this one does I okay so the architecture is multi- Agent the reason why we did that is because you don't want to put in scope as many tools as you created and a lot of different instructions so one of the

key recipes to build a more reliable agent is to have the most narrow scope you can think of so we have a we have a team that looks very much like a small YC company we have a manager we have Software editors we have a verifier and then we have a constellation of different tools which in all honesty we didn't have to build during this project rapit has been around for several years so most of the tools were there the challenge is to make the tools talk the right language that you you know allows you

to orchestrate them reliably with an llm so there is this field I think that is starting now in the research called ACI so agent computer interface and we We have to come up with the right way to build the interface between tools and the agent and yeah I can go I if anyone has questions I can go like a bit more in detail about uh the challenges definitely if you work with code you realize that creating files is relatively easy with with a Frontier Model editing files is much harder if you want want to do

that reliably if you want to make sure that the model is not lazy you can basically destroy your Application with just one llm call and that's the reason why we allow you to revert and we take a lot of checkpoints it's hard to detect that so we rather have the user in the loop rescue the situation we tried a bit of oh yeah the question is did I'm going to paraphrase did you put more effort in F tuning a model to make the agent better or is it rather the ACI problem like taking cloud and

using it correctly And also like all the HCI work on the product per se I would say it's more the latter we started by a relatively simple architecture and we tried to F tune models to make it more reliable and then we realized that relying more on Sonet and on its EDI capabilities and then building the right interface to the tools really gave us a big lift and from there we definitely have work going on right now where we are still fun tuning models or using like some capabilities That are still not available at large

uh but there is so much you can do with a Frontier Model and the right way to prompt it and to use it shall we open it up for questions guys yeah all right let's do it what's your workflow like QA like offline kind measuring great question um so the question is how do you measure Quality quality yeah all right so the way we do evals is first of all Vibe checks are extremely important when you're dog fooding the Best thing you can do is you kind of force everyone inside the company to use it

all the time luckily after the product started to work people wanted to use it because it was fun so we got a lot of internal feedback um it's easier to Evol agents on a step per step basis so like we have a relatively good way right now to understand if from one step to another the agent is following the behavior we expected or not it's much harder to do that end to end so we have Some ways of testing end to end building one of these relatively simple applications and we are building something more advanced

now where we're going to basically imitate what a user does when they're in front of a web application um you see part of that logic already in the proc product when they take when we take a screenshot and we interpret it and then we make decisions based on that but we're going to go all the way to having another Verifier agent that literally clicks and sends input to the web application and we're basically going to study those traces we do so the checkpointing is already there like I'm J was pting during the demo and we

also have a reflection Loop in the agent which means every few steps we try to assess if we made progress or not if we didn't we roll back automatically uh so that we do our best not to get you stuck in our local Minimum there's an important other quality test that amjad you were telling me about the the other day which is uh omj out as a service where was it every Friday afternoon yeah every Friday afternoon you would sit down with the whole team who was working on this and you would live test it

with a prompt and every week you personally would find quality issues and then the team would go and fix them yeah yeah I I I believe Actually am I wrong did we coin Vibe testing and in the AL Community I think we did uh Lucas from vit Andes quotes you as the yeah yeah so so like even going back to to two or three years ago um there was this amja test which is like a Vibe test that I I I I've developed quite an affinity to large language models maybe I think like them a

little bit but but I I was able to to kind of aren't we all large language models ourselves maybe maybe we uh so Every Friday we call the agent Salon we get everyone half the company we sit down and uh I I I run a test and and those are this that that hour is either the most miserable hour of my life or the best hour of my life right because the Run could go really badly and everyone is like that I'm like we're gonna fail we're GNA all get fired uh including me uh but

but when it works it's super fun like one time we built like this collaborative uh game of uh Conways Game of Life and everyone was like playing with it during and we're like oh my God we're gonna make it and that was a lot of fun I also the horror stories like the game days before the launch you want to we had game days which means we use alpan hour of the entire company everyone was eating the agent with multiple uh sessions and I think we had two of them the two days before launch and

one day before lunch the two days Before lunch was an absolute disaster uh so many things failed that we never expected and we introduce a regression the night before something like that so and we had a DM later in the night and and AMJ asked me do you think we're still launching and I think I told Jeff I'm 80% confident that we're going to be there and we were all locked in the war room that night you know to make sure that the bug will be fixed which actually happened at 3:00 a.m. if I Rec

Correctly um and then the the game day right before launch was was Flawless and then we said okay this is going to happen that's how it goes okay so we have this comically large microphone for questions s amazing you the the the secret to the comically large microphone is you have to toss it to the next person oh sweet it's designed to be tossed love it um I'm lucky to known am Jud for many years be before he was a mega celebrity Uh early on you mentioned that for hiring when candidate would be talking to

some of the bigger companies you wish you would just skip talking to them I'm curious and that advice is by the way super true I'm curious for like especially the stage from like three people to 10 which a lot of people have advice for like 10 25 plus but the the very special moment of like three people to 10 do you have any hiring advice uh for Founders out there yeah so so so uh Uh you know what what we used to do is like we'll ask them who are you interviewing with and if they're

interviewing with with fangs literally hang up the phone they're not going to work at a startup especially at the time when when the offers were atrous glad you didn't do it with me um and and so so um you know three to to 10 people what would we do every interview we'll do a phone screen but then uh we'll bring them to office and We'll sit down and we'll we'll hack with them for for a day uh sometime sometimes we'll even pay them for that uh if we're in a good mood but we'll we'll we'll

sit down um and you can tell so much uh about about about that person just from from spending you know six to 10 to to 12 hours but by the way if the last 12 hours that's also that's also a good sign because we were we were there grinding for for 12 hours and um and uh you could tell so much about about their Code their their work ethic their culture fit and so I you know for ear super early stage just bring people into the office and just spend a ton of time with them

and super quick I remember you did your toes into crypto at some point and I'm curious do you have any like thoughts right I know it's very untrendy now but do you we're still super bullish on on bitcoin by the way that that thread that I wrote about with with agents it talked about coordination Between AIS and humans using using Bitcoin and we're still thinking uh that uh that's part probably part of the plan in the future is like I think we're still want to bring in some some more uh market dynamics into the agent

all right can you toss it to the person right next to across has to be across the room okay all right hi um I'm I'm building a gen AI agent like uh code AI agent for you know let's you build apps from doodles from drawings uh can I Stream the output of my agent to repet kind of like in a similar way like in real time like I I'm thinking of using like repet as like IPI mm to deploy like I think would be cool to to like do like a real time like streaming from

my agent you know to rep environment come to us I think alha the company is around you yeah I'll show show you my with us yeah yeah definitely yeah but by the way we're we're super open to to kind of Working and supporting other other sort of Agents running on repet there's a bunch of them AER others uh so we're we're pro that I I I would love to so I think really cool toss it to someone heyj um question for you uh so this is boiler plate you know you start a project but uh

are you in a position or is it on the road map to uh take the existing code and say okay here's what I have and I want to expand this Particular section and build more features there yes so we we have a feature coming uh pretty soon um that allows you to do that in any code base and then starting the agent on existing code bases we have like some research effort here I don't want to promise it but I think it's going to be there in a month so sounds like a promise yeah good

toss good toss um hey y'all I'm Omar uh thanks for doing this Uh talk it's very uh informative I'm curious about like uh personally uh times that you failed uh could you kind of let us a little bit uh into your minds as to you know times you had to push back deadlines uh times that things didn't work they weren't So Glorious uh what did you learn um and you know what do you take with you now um especially now that you're on this uptrend and you know things seem to be um you know going

Beautifully yeah I mean I think we mentioned a few of them like quite literally that two days before the launch uh you know I messaged them I said this doesn't work like you know it's like literally and and and so um I I I think so I I think you know there's a sense in which in which you got to keep going I think that's the story behind repet I mean repet even even before before being a company was Sort of a side project and I had this feeling it's like you show up every day

and and you do the thing and over long Horizon you're going to win right and so I I think that's really a deep belief that I have is that you know there's the uh there's the ups and downs on The Daily by the way in a startup uh there's the ups and downs of like sometimes you're growing sometimes you're plateauing uh other times you're you're Hopefully you know uh not uh declining but it could happen uh but at the end of the day if you zoom out uh over over decades are are you um are

you headed the the right direction right and I think what that's what really matters and you judge yourself based on are you sort of writing the same the right waves are you making the right predictions about the future of technology and then are you positioning yourself to to make use of that and I think if you Continuously do that eventually you'll win yeah for folks who don't know Reet has a fascinating backstory this might be a good time to just tell people like about like the first seven years of Breet like for for years repet

did not seem like he was going to be successful wasn't even trying to be successful wasn't even trying to be a company and then it it turned into one like the the world pulled the company out of you it's a Very unusual story yeah so you know goes back to uh 2008 when I was when I was in college and um I I I simply wanted a a way to write code uh without having to spend a ton of time installing crap right and by the way that's literally what we still solv with with the

agent right you know the agent solves a ton of crap um and so um you know I didn't have a laptop and so I would go to the computer lab in in school and continuously do this laborious task of Installing debugging packages and all of that so I thought surely you know someone would have written like an online compiler that could like you could write code in Well turns out no one had done that by that point which was super surprising to me I was like how hard could it be so I started working on

that uh had had some initial success by just like running JavaScript into the browser writing a few languages that that could run run JavaScript uh But hit hit kind of a hit hit a wall when you know I wanted to run python like how do you even run python in the browser writing a python interpreter in JavaScript just felt like a crazy thing although I tried to do it um and in in 2011 I kind of had another breakthrough which was being able to compile the C uh implementation of python to to JavaScript and this

is what I was saying about kind of are you on the right trend of of technology and so the browsers Were getting more and more powerful so you could do something like that and we were the first to do that um and and released what became repet on Hacker News and went super viral uh by just like being this thing that could run a bunch of programmed language in the browser and people have never seen that before and so it was really really surprising to people um but even after that that like it wasn't like

the day after it went viral on Hacker News you Quit your job and like Turned rep into a startup I tried I was back in Jordan there was no VC to speak speak of I didn't have my own money to fund a company and so I couldn't do that um and then luckily a YC company uh uh reached out to me and they were using my software that that's code academy uh Zach and Ryan they came out to Jordan recruited me because I kept saying no uh they gave me founding engineer title uh I came

out to to New York and at that Point kind of repet was abandoned I was working on code academy really focused on it um and code academy by the way was a huge success millions of people learned how to code sold for half a billion dollars uh recently um and uh but but along the way kind of repet died died a little bit and uh I went to work at Facebook was lucky to work on react and react native and a bunch of like JavaScript tools uh was kind of the early team working on those

and it Wasn't until 2016 that um that decided to work on repet again and and the idea that was five years after The Hacker News launch yeah so you'd been doing other stuff in life and repet had just quietly been growing the whole time on its own with nobody really working on it nobody really working on it I might fix a bug or two on the weekend but people loved it like the idea of just like going to the browser and writing code and and Running it was something that people I was like you know

there was Alternatives as well but people kept coming and it was buggy and it was like really Weighing on me and I was like ah I guess I guess I'll fix it and then and then you know Docker came out it was possible to do like you know server side stuff and I I I started fixing the stuff and it started growing really fast then around 2015 16 and it was growing so fast I was spending so much money on it Just to keep it up uh and at that point like like Jared says it

kind of exploded itself on the world and it kind of forced us to create a company around it because otherwise we wouldn't be able to sustain it all right who's next toss it to the back hey thanks so much for sharing um I'm really interested in the future of personal software I've been devouring the mother of all demos and thinking about how that was almost 60 years ago And we're still like playing in that Plateau sandbox Bo of what engelbart was talking about and I've been thinking about the Computing environment and if you build it

in a proper way then really anyone can do anything in there and you were talking about the register of of this tool of like coders get it and then non-coders are kind of like I don't know how to do this but then you've built this like access point so could you just share a little bit about Your vision of the future for personal software yeah um so people were not familiar um the the mother of old demos was it 68 I think 67 67 uh in Silicon Valley I think SRI SRI Douglas engelbart uh does

this demo that basically set the stage for the next you know uh 50 60 years of computing he basically demoed Google Docs Skype and all the things notion all the things that we like today in one demo that was called the mother of all demo you can go Watch on YouTube and just like and it was unveiling The Mouse and the corded key set it it was phenomenal yeah and and and I think so the the idea of like you know personal software has has been around for a long time is that you know you

are no longer the sort of consumer of software typically the relationship right now we have to today is that there are developers and then there are consumers of software the vision of early Computing P Pioneers was that We're all creators of software like our machines our environments have like multiple uh sort of uh software suite of tools that allows us to build software as as we go and so you know the the the mission for repet uh you know is to is to make it so that there's a billion people creating software and so hopefully

more than that in the future but but we thought like a billion is like a is like a good good number to start with and and so um and so the idea Is that we're just going to build a lot of these tools um and we've been building it for the long time which is what Michel and and the team was able to orchestrate and so we want to build all the tools necessary All The Primitives and then use Ai and agents to be able to orchestrate them in a way that human creators would want

anything that I love you mention that demo because literally two weeks ago we had an internal session of youas which we call like the mother World demos at rid like we wanted to figure out what can we build in the next like five to 10 years how is personal software going to look like uh one of the examples that really resonated I think with all of us is the fact that Excel seems to be like the common denominator for personal software and like allowing you to use your quantitative skills it's amazing it has more like

a billion uh users today if I recall correctly but at the same time It's possibly like the worst possible like generic interface that that you can give to people and we we want want to break that barrier like we want to make sure that everyone can build their own Excel customized to their own needs it seems to be there like right now rapid agent is amazing at building dashboards and building like basic tools that allow you to manipulate data you can only imagine what's going to happen like next year so I think all the software

that Everyone is used to that has like literally like a more than a billion users today will probably be reinvented in the next few years by being completely personalized and that's why we're building the agent all right I think we got time for two more questions choose someone to toss it to hello so I'm curious so with the release of open ai's new voice model I would like to know how soon will I be Able to talk to my phone and make it make an app you can already do it you can dictate The Prompt

definitely and tell me will tell me it doesn't respond back with voice yet uh but you can definitely dictate The Prompt how soon until it responds resp with voice um one month three weeks does anyone want to work at rapet come talk to us later okay we have a lot of cool ideas but it's a small company but we'll get there you a big Toss there you go all right um in terms of product like do we envision this being as like multi-agents like if imagine if I want to build a more complex system could

have like a small team of different agents working on a different flow and then putting them all together at the same time like yeah imagine like a mini Erp or something and then you'd be surprised some people are writing mini prds in the prom today like we have a Slack Channel where we see every single problem that people are submitting because you know we are curious to see what people are building uh some are very short some others are writing markdown instructions you know with different headings and going very much in detail unfortunately we can't

really follow every single instruction because it's overwhelming uh but people are already getting there that being said we have all the architecture in place to do That what is awesome about the rapid infrastructure is that every rle is mapped to an isolated container you can do whatever you want in there if it crashes nothing happens so we can have like a lot of Agents spawn in parallel each one of them doing a different task or uh think of what is you know you probably heard of MCTS and all these Technologies where not only you generate

but you also generate in parallel trajectories and see which one of them Is best you can think of a future version of the agent where it becomes cheaper to run where rather than doing one run we're going to do five or 10 and then we're going to chry pick the best one so the moment that comput is there llms are getting more powerful and cheaper sky is the limit at that point I think it would be really cool to um give you the user the ability to select how much compute M you want to spend

on a on a task and so if if the task really Matters to you or you're really rich and you want to spend a ton of a compute on it you should be totally able to just say like hey like you know spin up a hundred agents to work on this and maybe we seeed them differently and then and then we select the best the best one for you all right we've gotten so good at tossing we can do one more want to toss to someone thank you that was easy So currently reation supports Mostly

right do you have any plans to support other languages soon or better of other platforms like iOS apps I promise not to give a deadline on this one otherwise people are rapit are going to shoot me in the back um we even have demos internally working well with JavaScript right now I think we try to optimize for building something that is reliable as reliable as possible for as many people and we're going to give we're going to allow you to work with Another stock as soon as we feel confident with that it's definitely in the

road map excode is terrible like just build stuff for iOS and now you're a new customer what what language are you requesting Swift swift is a bit harder because then we would have also to run the entire iOS in place let's we're going to stick to web apps and scripts for a while it's it's better to do that on all right in The back can you talk to the and the white shirt almost there thanks um I'm curious are you using uh the agent to make rapet better internally ret as a product not yet RIT

as a company yes we have people that built personal apps or that are building dashboards like to Tru internal metrics and it has been very exciting to see noncore Engineers building applications around and being more productive it kind Of you know confirms that our thesis is correct that we are are empowering everyone to create software awesome thanks I have a very related question uh so in terms so in the demo I saw that you did not have a code review interface right away like the GitHub co-pilot workspace project so uh when you imagine uh as

you said you're going to launch something to start working on an existing project when you imagine rep using replate on existing project uh do You imagine the experience to be similar to this or do you imagine the experience to be more uh similar to how existing Engineers work with each other asly you will be the people telling us exactly what you like in my opinion it will look similar because we are using it as a backand so every time the agent does a substantial step forward we commit it and then if you don't like it

you can always revert but nothing prevents us to stop and Generate a PR and then ask if you want to merge it or not like the the complexity here is not to decide when we check point it's rather doing the right thing that's where we're focusing sorry so uh the interface itself of uh verifying whether what has been built is correct or not I'm talking about that like do you think that will be more on the chat side and uh you know manual testing side or uh do you think that a review interface similar to

what you Have on say get yeah no I think it will be more on the chat side that's what we love about the agent the fact that you speak in natural language and you can make progress just by expressing your requirements and your needs and we're going to try to stick to that Paradigm as for as long as possible rather we're going to try to make it even simpler than chatting if it can show UI affordances even better because it makes the product even easier to use Uh there there are a lot of tools for

The Other Extreme of the spect for professional developers they're really happy with that you know we're building for new Developers all right I got a I got a final question um I'm Jad you've been thinking about agents for a long time since 2022 before like anyone was talking about this what do you think is the future of Agents especially for folks who are like interested in this area like what what What can you predict about the future that might help people like know what to build yeah uh I think like one way to try to

figure out like where where the world is headed is like see like where everyone is optimizing and by everyone I mean where the big AI labs and the big efforts uh are going and certainly you can see that now uh there's a lot lot of effort going into um into reasoning so like models like open AI 01 and there are a lot of efforts going into having Models better interact with the world around them right and I think that is uh that means that 2025 probably the year where agents uh are going to work really

really well are gonna start automating large amounts of jobs it's going to start start having a real effect on the economy you know there there's there's all these questions whether llms are going to have effect on the economy and um and and I think agents is the way because you know The thing in the economy that that you know we spent a lot of money on is is labor so making people more efficient um is is the way where where AI will have a real uh demonal impact that maybe on GDP even in 2026 so

I think 2025 agents will get a lot easier to build um and and the way we design our agent I'll give you kind of a free tip is that uh build something just on the edge of what's possible because then when the uh when all the AI researchers are really Working for you to make your app better and so so in a month or two your app's going to get better month after that your app is going to get better and you're writing you're writing sort of a wave I love that it's crazy future um

do you guys have final announcements for folks before we wrap up um do we have anything use replit agent it's freaking awesome we are looking for people coming To help us build the agent I I think we'll publish tomorrow a hiring challenge is like a nice AI uh project to build it's one of the problems that we face with the agent and we cracked we want to see which one of you can crack it as well and if you do a great job we're going to reach out so that's one announcement that I can make

I'm also going to put uh ret team on on the spot can you raise your hands if if you work at repet first of all thank you everyone Who works at repet oh there's a a whole bunch of folks upstairs are are these all the folks who actually built the thing that we just saw whoa so like a whole chunk of the team is actually here Round of Applause for everyone and and please please find them and and Grill them on on repa agent and and anything anything you want amazing all right there's going to

be food over there please feel free to hang out and talk to the whole rep team they built an Incredible product thank you thank you