Ismail Pelaseyed, Superagent: Founder at Superagent

Welcome to another founder wisdom bod. We have Ismael Piladed with us today. He's a founder at super agent.

It's a YC winter 24 company. It's an agent workspace and uh yeah founders are probably my best uh and favorite people to have on this uh podcast. Ismail.

I don't want to make you blush, but yeah, normally very smart founders and Ismile is uh is not the exception here. He has been founding a couple of companies uh Key Crunch I see here, Bitcoin, Merand before he went into YC in 2023. So, Ismael, can you uh tell us about uh what you've been up to nowadays since uh YC has started?

Yeah, thank you for having me and inviting me on this uh live stream. So, happy to be here again. I'm the CTO and co-founder of Super Agent.

I run it together with my uh colleague Allan. We've been building as you mentioned a couple of companies together and something unique about us I think is that we've been working together for the past 12 years which is uh pretty unique even in YC terms. Usually founders a long wedding long marriage.

It's a long marriage. Exactly. Uh currently we are building a uh platform for product teams where they can do what we call guided coding.

You probably heard about vibe coding that people do a lot this these days. Uh similar to that but for product teams who actually want to ship AI generated code into production. So it's a controlled environment where you leverage all of the you know u great parts of vibe coding but in a grounded and controlled way.

So that's what we're building now. Tell me about the grounded controlled way like what do you mean by that's like secure for businesses for like enterprise clients or I mean you know when you ship code when you wipe code what you do really is you open up a chat you open up cursor or whatever agent coding agent that you're using and you start typing something and then you just you know go back and forth with an agent that does a bunch of changes and when you're happy it just you know it's great for hobby projects. It's great for building prototypes.

But let's say you have an existing project, an existing platform that has been running for 10 years. There are 10 different repositories. Uh that method methodology doesn't really work there.

Let's say you have a team of 100 developers working on this platform, right? So what we when we say grounded means that you know we connect to all of your services your GitHub your linear your error logs everything suck in all of that information and you ask super agent hey I want to build a feature right I want to build feature X what we do is that we do all of the research for you and basically generate a custom prompt that you then can pass on to let's say Devon if you're using that as a coding agent or replet Devon is still a thing. Devon is definitely still a thing.

Uh you know we use Devon internally. I think Devon produces around 70% of our own code. Wow.

Uh and that shows you the power of you know having grounded as we call it coding changes done automatically by AI. Uh it can really enhance the productivity of the team. uh for us it's almost doubled the productivity of our team uh and I don't think like today we merged 10 different features into our platform I don't think I don't think one of one of 10 was done by a human the rest was a collaboration between human and AI yeah I believe you pretty much the same with my own companies Devon I'm surprised because I thought it was a flop right but I think there's just not a lot of companies don't have access to it.

What do they charge? Uh 500 a,000 per month. They charge uh 500 per month, but you know 500 is nothing if you can double compared to a dev.

Yeah. So a dev costs around two 300,000k a year. Uh 500 a month is a bargain.

And uh you know I'm actually I'm actually um impressed by Devon how well it works when you do incremental changes to an existing project as opposed to doing you know vibe coding something from scratch. Perhaps you have Vzero or Lovable or any of these other more vibe coding uh uh platforms that you can use or apps that you can use that that get you from zero to one uh for a nice prototype even production use cases as well. I mean I don't think that it's only for prototyping.

Uh but we specifically target existing product teams that already have a product in production that already have their CI/CD pipelines that already have their processes and workflows uh set and we integrate deeply into those workflows. Right. Yeah.

And real quick, coming back to Devon, do you think it's just a PR flop if they really have a product? Because the word out is that it they just oversell something that doesn't work. Yeah.

You know, uh that happens a lot in this business. People oversell everything. Um let me give you uh let me give you a story around overselling.

There was, you know, back in, I think 2023 or 2022 even prior to chat GPT, there was this uh product called Super AGI which was doing basically chain of thought prompting, right? But it was very early. The models weren't that good.

So people started using this and of course it was oversold, right? And it didn't really work because the models weren't there quite yet. Uh so what ended up happening was that people you know basically decided that that approach is Let's try some other approach.

And it actually turns out that that approach was correct. It was just early. Uh and I think uh you know Devon 1.0 was that early sign.

Uh but with the new Devon 2.0 I am uh you know really impressed by by how it works. It has all of the, you know, all of the stuff that you need in a coding agent. But again, there are others as well.

There's one called Codegen that I've heard really good things about. I know Linear has integrated with them. Uh, Replet is doing great as well.

And I know that they are also releasing something very similar to Devon, not only for VIP coders, but for incremental changes. uh so I believe in a future where you know there will be hundreds of coding agents companies will use several of them and we are trying to be the layer on top of that where you can basically um decide what to do and then outsource that to whatever agents that you're running internally you know right yeah Devon's quite interesting my my thought was like how could they mess up with like 500 plus million dollars in the bank in funding I I think they have at least right so and the lesson there is that yes I think PR can destroy company because the the mentalities move fast and break things for them especially if you're dealing with devs but they weren't dealing with devs they were dealing with Leo types of no coders right and these people if they see something break they they're like oh this doesn't work it's and they spread the word on social media so yeah let's see what happens but thanks for telling me that Devon is good I'm actually gonna give it a second shot now all of these other tools, uh, Replable is pretty terrible. I use it for my landing pages.

It's pretty good for that, but uh, none of them seem to actually succeed into creating like a real working product except uh, cursor. So, cursor seems to be the go-to. So, is that like the the current uh, state of things right now?

Only cursors is ready to deploy things and all of these tools cannot produce like working prototypes. Yeah, I think cursor is meant for developers because it's in your IDE, right? So you you actually have to know what a IDE is.

You actually have Well, you just ask GPT, right? Like yeah, but still, you know, it's a big step for someone who has never written a line of code to open cursor and see all of those files and learn everything. That's a huge step.

So I think it's primarily for developers. uh I think like lovable replet u and uh you know vzero are targeting a completely different type of audience where these people have no prior experience they know how to use a chat they don't want an IDE per se uh they don't want to set up a development environment locally they don't want to do any of that they just want to vibe and and deploy to the cloud and you know it's early I think that the future will look uh very different in a year from now. And uh I think that most of like any AI startup um I think they are well positioned in the sense that the better the models get the better the output of of those tools get.

Um but I also see a future where you know we talk to teams today and you know they use uh you know they use cursor they use wind surf they use devin they use everything people you know they have 10 different coding soli coding solutions agents deployed and they're testing they're trying to figure out like what works internally for our process um and um so I think it's early on but The signs from what we do internally from what I've seen from you know companies using these type of agents is that most probably you know 95% of your code in the future will be generated by AI. So right and then who will win that race? I'm not sure it's even a race to win maybe uh but I'm not sure that's a race that you know has a has one clear winner you know.

And the objective will change, the race will change in real time, which is uh exciting and scary at the same time. Do you really think that Bolt and all these tools will be able to code a fully workable a vibe code that is right a fully workable product with them in a couple of months or can AI actually do that? I think AI is uh good enough to do that.

The question is is the user good enough to give instructions to the AI at this current level. Right? So uh an example would be if I give you know any AI a bad prompt you know the input in equals the the output out basically right if I'm not sure what I want to build if I'm not you know clear enough in my I may be thinking something but then prompting something different and just expecting the AI to understand.

I think that is a big hurdle for many people don't know how to prompt and these agents require a lot of prompting. It's a skill and uh so uh you know that might change in the future or it might be even more of a skill in the future. Uh from what I see you know if you if you look at GPT4.1 it requires a lot of prompting in order to actually work the way you want it to work.

So I'm not sure. Don't you think that these companies should just integrate in it that if the prompt is just reply with hate? Is that what you meant?

uh or hey I just reverse engineer the the best companies and that here are their features and here's how quick I could uh deploy them. Don't you think that this would be like a major moes as uh one of these winds surf or uh cursor type of companies? I don't think any of these have mode because you can there is no lock in you know from if you're using cursor you can switch to wind surf in 10 seconds there is no mode uh the only moat that they have might have is like the scaffolding that they have around their coding agent or the uh language model but you know even that I think that you know scaffolding in general will be disappear it will disappear over as as soon as the these agents become better you don't need that scaffolding either.

Probably in the future you can, you know, run a model raw without any scaffolding, you know, and have it do as as good. But back to your question, I I think that it's very early. You know, I don't know I don't know what the future uh holds for these companies specifically.

What I do know though is that there will be many of them. there are already many of them and uh I don't see that any of them will be a clear winner very similar to how you know people use IDE today you have VS code but you also have Neo Vim you have sublime you have a bunch of stuff people use and people use different things even though they're in the same team and I think that we will see that with agents as well you know does open AAI have a mode do you think that they are sort of the infrastructure of it all and do you think that they will win the most in this AGI race? I think for chat uh OpenAI you know are have a very strong position and I don't know if it's moat I don't I don't think I you know yeah it's a moat but I think the thing that openai has that's invaluable is the brand openai is a very strong brand you know chatpt is a very strong brand and they have succeeded in building that it's like the apple of chat apps in my opinion and it's very hard to compete even though you might have a superior product like you know it's very hard to compete with that sort of a brand.

So brandwise I think it's very you know and mo wise you know they're building a lot of memory features where the agents will uh you know uh remember all of the inputs you've had which which makes it harder to switch right and they have a very high velocity of shipping new products um you know deep research image models chat models computer use models operator etc etc so I think it will be hard and you know if you look at claude as an example. I don't see the same ship ship rate basically even though the model is great. You know, uh the ship rate of their chat application, I think they gave gave up.

There's no point. Huh. Interesting.

Yeah. Clown's interesting. Uh sticking to Chad GPT, do you think uh what do you think about the Sam Alman brand?

Uh do you think he's the right founder to scale this company? Because he ships like crazy for sure. They've been through uh turmoils as you know.

So do you think that he'll stick around for long or is the Microsoft uh investment overshadowing going to uh put the company in jeopardy and Sam that away eventually? Uh I'm not sure but I from you know um from what I've seen founders coming out of YC especially that cohort of YC if you go back to look at like Sam Alman's you know um background and the network and the lessons and the you know grit and the and the time he has spent in this field. Um I think we have to be honest most of these labs they spring out of open AI even anthropic right so I'm I'm I mean I think that everybody's betting on Sam and I don't think I think he's like I don't know him personally uh but I think what he has done with Chachi PT is just you know uh astonishing growth rate product astonishing yeah he's the new Musk for sure I'm just unsure like if the Microsoft corporatism is going to long-term fit, you know, if if he's going to branch out, if he does the investment, the trillions are are going to follow him for sure.

Uh do you think operators going to officially kickstart the Aantic revolution and when would that be? Because operator is not where it needs to be right now. I definitely think it's uh a future thing.

I think it's it has a lot of potential. I think that u you know computer use agents are very interesting uh and I think that they give that certain dimension that certain type of tasks need and I'm not talking about shopping on you know door dash I think those use cases are you know they're nice but that's not like I I think there are other use cases where these uh computer use agents can create really a lot of value and actually displace a lot of human labor that is uh put into it because in the end like how much are you willing to pay? Are you willing to pay $500 a month to to have an agent shop on Door Dash for you?

No. No. You you're not even you know you don't even want to pay $200 a month.

And uh I think that there are use cases where there are high value you know targets that you can go after with that kind of technology. Yeah, I can picture the VC use case in the penthouse uh with 500 bucks a month to have Door Dash automated. It's quite silly.

And that got me reflected on how OpenAI launches its products, right? It's like one team per product. They I think they're in charge of branding and the launch and the full side of things.

I'd love to learn more about that aspect. But talking about this, this got me reflecting about the structure of OpenAI itself and how a couple of very smart co-founders created like massive massive breakthroughs, right? And what you said, you've been 12 years with your co-founder.

So what's the necessary ingredients to build something that groundbreaking the chemistry between co-founders? Is it IQ? Is it EQ?

How do you make actually the fruitful relationship happen like that? I think it's uh you know a combination of different things. First off, personalities, you know, like any other uh relationship, personalities need to fit.

Um and I think that being too similar is usually not a good thing. You would want some, you know, uh spread on the way you think and and what you do and that kind of thing. I think alignment uh around what the vision is, what the goal is with the company or with the product that you're building is super uh um you know important.

I also think that both founders need to know and understand that you know building companies is not something you do a year. You know it's it's a lifetime uh job basically. It's a lifestyle.

It's a it's a uh it's a way of living. It's a way of working and for some people it might take five years to build a billion dollar companies for some people it takes 10 years for open AI it took seven to eight years right and for Elon Musk it took like 20 years I don't know you know but 10 years at least and uh and you will fail and you need to have a a person or a team where uh failure is just another step on the way you know to to to greatness and success and uh if you don't have that I think it will be tough it's a tough journey to have to do without that. Thanks for that.

Uh talking about alignment uh isn't an AGI don't we need non-alignment to create AGI and feel that if we box it we're not going to get AGI and uh do you feel that the agentic revolution that in my opinion would start probably Q4 this year would greatly accelerate um AGI? Um I think the you know um people have a lot of different um you know definitions of AGI. Um for me the most you know the super intelligence in general what what what is the you know the most interesting part and the most valuable part of super intelligence is to get to a level where the AI can train the AI.

That's when you get the uh the curve that you want the exponential growth, the exponential development and progress in AI. Currently, humans are involved a lot. If we could automate that with AI's training AIs, I think that would lead to something, you know, but we hit GPU limits.

Um I'm not sure. I'm not sure like GPU limits for me is just a way of doing things. You know, it's a infrastructure or a a not infrastructure but it's it's an algorithmic thing.

So there might be algorithms that are far superior than a transformer algorithm that would require less GPUs. There's not a there there isn't a natural law really that states that, you know, transformers is the way and that there aren't other ways of of doing it more proficiently. I you see Deep Seek, you see, you know, other labs um outside of the US that are uh working with other you know uh technologies that are much more efficient uh you know um at least what they say much more efficient and so I think there are other uh I think there are other methods I don't think this is the end this is just the start and then the agentic revolution I guess like AI training AI would certainly be agentic.

We we don't really have agentic nowadays and again I think it's about definitions but to me agentic is yeah like a Charles GPT reproducing itself uh every day working while I sleep and uh giving it new goals achieving these goals reflecting on these goals analyzing and coming up with a new experiment that it executes on itself right yeah it's autonomy that's what what everybody wants that was the you know promise of AI was autonomy and super intelligence. That's what all the money goes to, right? That's what every lab wants to build.

That's the holy grail. And so I think that uh founders need to keep that in mind when they build their products. Uh as you said, there's a lot of things happening very fast, faster than before at least um at least in my lifetime.

And I think that you as a founder need to understand like where can I provide value in this space? Where does OpenAI provide value? Where can I provide value?

How does that, you know, fit with my vision, the product I'm building, the the user I want to build the product for? Uh, if you don't have all of that set, I think it could get tough, right? Well, Isma, thank you for coming on today.

Where can people learn more about you in the company? Super Aent.sh. Follow our blog.

We write a lot of stuff about our thoughts and AI and the future and what you can do today in order to build efficient agents uh you know coding all of that. So super agent.sh Sh was Ismile