May 27, 2025 • 53 mins
Unlock the secrets to world-class developer experience for AI agents! Discover how to build powerful, reliable AI agents with Ahmad Awais, founder and CEO of Lang Base, a serverless platform revolutionizing AI development. We dive deep into overcoming complex tooling and steep learning curves, exploring AI primitives, and the groundbreaking AI coding agent, Chai, which builds other AI agents for you. Learn about agentic RAG, memory agents, scaling agent workflows, and the future of AI agent creation. Ahmad shares his journey from open-source contributor to AI innovator, and how Lang Base and Chai are tearing down barriers to innovation in the AI space. If you're building, designing, or dreaming of the next generation of AI, this episode is essential viewing.
Lang Base: https://langbase.com/
Chai: https://chai.new/
Sign up for A.I. coaching for professionals at: https://www.anetic.co
Get FREE AI tools
pip install tool-use-ai
Connect with us
https://x.com/ToolUseAI
https://x.com/MikeBirdTech
https://x.com/MrAhmadAwais
00:00:00 - Intro
00:00:46 - Ahmad Awais' Journey into AI & Open Source
00:09:03 - Deep Dive: Lang Base Chai - AI Agent Builder Explained
00:33:45 - Most Compelling AI Agents Built with Lang Base & Chai
00:36:32 - Principles for Great Developer Experience (DevEx) in AI Tooling
00:45:49 - Tools & Advice for Advanced AI Agent Development (ft. MCP)
00:51:04 - Bonus: Ahmad's Open Source Code on Mars with NASA
Subscribe for more insights on AI tools, productivity, and AI agents.
Tool Use is a weekly conversation with AI experts brought to you by Anetic.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
(00:00):
We demand brilliance for our AI agents, but what about the
experience for their creators? The truth is, building powerful,
reliable AI agents can be a messof complex tooling, steep
learning curves, and frustratingroadblocks.
Friction is the enemy of innovation.
So how do we tear down these barriers?
What does a world class developer experience for AI
agents look like? Today, we're exploring the
crucial yet often overlooked world of the developer
(00:20):
experience for AI agents, Essential knowledge for anyone
building, designing, or dreamingof the next generation of AI.
Welcome to episode 41 of Tool Use, the weekly conversation
about AI tools and strategies toempower forward thinking minds
brought to you by Anetic. I'm Mike Byrd, and this week
we're joined by Amit Awais, the founder and CEO of Langbase, A
serverless platform for buildingAI agents.
Ahmad, welcome to Tulius. Thank you for having me here
(00:41):
Mike. Super excited to talk about
agents. Oh, it's always a blast.
Would you mind giving the audience a little background how
you got into AI and what got youhere today?
Yeah, of course. I think I've been around the
block for like, I don't know, 25years now, building one thing
after another. And I remember back in the day,
I started contributing with the word background, you know, where
(01:02):
we were building Web 2 point O, which is now known as HTML5,
right? And that led me to a career in
open source. Like I ended up contributing a
lot to the core software of WordPress, then a lot of
JavaScript libraries, then the core of Node JS.
There was a time when I was leading the community
committee's outreach program forNode JS, right?
(01:24):
Got deeply involved with Open API Enterprise Governance Board
and got on a Google developers Advisory Board.
I'm practically very, very technical.
I've built a lot of software. I have like I think created
about, I must say, more than 150open source packages and they
get like 40-50 million downloadsevery year.
(01:47):
And I have like one of the top 10 VS Code themes called shades
of purple. I love the purple color.
And so, so, so, so is the name of the theme, right?
And over the years I have done alot of open source advocacy.
I remember about I think 11 years ago, I got Facebook to
open source React and graph QR, right?
(02:08):
That was when my career absolutely changed.
Like I saw first hand how React became the other half of the
Internet is how I call it. Like half of the Internet is
WordPress and half of the Internet is React, right?
So I gently tend to joke about it like there's no website you
cannot visit without my small little my, you know, piece of
tiny contribution that I might have made, right?
(02:28):
I take pride in that, right? And I remember like, because I
became this developer advocate of sorts, I ended up doing a lot
of talks, right? People wanted to understand,
especially tool use, right? Understand how I'm using these
tools, what I care about as a developer.
Why this? Why that to a point where when I
(02:50):
switched from Sublime to VS Codein 2017, when VS Code was like a
new editor, right? Everybody was like, you know,
how are you doing this? Like what is this thing?
What is that, right? To a point where I ended up
building a course called VS Codedot pro.
I think I've helped probably more than 56,000 developers
switch to VS Code over the years, right?
So and about Lang base, I'm now building language, right?
(03:14):
So Lang base is serverless AI platform.
So anything you want to do with agents, from building an agent
to deploying it and scaling it, that is what we do.
And we are contrary to, you know, the normal wisdom that
people talk about like, well, everybody's building an AI
(03:34):
framework. I don't think you can build an
AI framework in this space. I think the space is moving too
fast and things are breaking tooquickly.
I think what we need are primitives, good primitives,
like for example AWS has S3A primitive where you can upload
your images and download them. You don't really care how they
(03:56):
are doing the scaling and charting behind the scene.
You just get a simple API and that API just works.
It's similar to how open AI, your developers experience with
open AI is like you get an API for their models and it just
works, right? You don't really get to see or
have to worry about how they arescaling their models, right?
So at LAN base we are building AI primitives, A primitive for
(04:19):
agent workflows, A primitive foryou know, building an agent
called Pipe, a primitive for agents with memory like human,
human like long term memory which have vector stores in
them, a primitive for threads. So you can easily store
coordinated database problem problem, you know solve the
(04:40):
problem of coordinated database with threads like you know, all
the tool calls and tool results should be coordinated in in the
same order and as fast as possible.
Then parse, search, anchor and Bradley, you name it, right?
We have a bunch of primitives, right?
And how all of this happened wasin 2020, I was travelling a lot.
(05:00):
I had like, I don't know, 43 conferences.
That was going to be the year where I was.
And I, I absolutely hate travelling.
You know, I'm not adventurous atall, you know, like programming
was my hobby. I'm practically boring.
This is all I do, you know? But in 2020 I decided, OK, let's
(05:20):
change things, right? And well, you know what
happened? COVID happened, right?
All my, you know, plans to travel got deprecated.
I remember one of the conferences I was going to share
stage with Obama as one of the keynote speakers.
I was super excited about it. And then I got to do it on hop
in online platform and it was sobad he didn't even show up
(05:43):
right. Anyway, I was I, I found myself
building this thing called Corona CLI and Corona CLI was
what perplexity is, but only forCOVID data and in 2020, right?
So you could ask a dumb questionlike if I go to Italy, will I
die? Or if I go to, I don't know,
like pick up TV from Costco on 4th street.
(06:06):
But what what happens, right? And it went absolutely while I
was just building it because I was travelling a lot and it went
viral when COVID became a globalpandemic, right?
And I meanwhile, like, no puns and like 10 billion API calls in
a month or so, right? And that got me an introduction
to Sam Altman and Greg Brockman.And this is 2020.
(06:28):
Like, open AI is not a big deal in 2020.
Nobody knows about open AI. You know, I didn't know about
much about open AI. They gave me access to GPT 3.
They had just finished training of GPT 3.
And I think June 2020. And I was talking to them in
July 2020, right? And I remember talking to Greg
Ruckman and telling him like, I want to suggest the next line of
(06:50):
code, right? And this is, this is like a year
before GitHub Gopalit was launched.
GitHub Gopalit was launched I think in October 2021, right?
So I was already excited about like this use case of like
suggesting something new. And I did a couple of workshops
for an open EI and a bunch of things.
Basically, I built a couple of software that eventually became
(07:12):
land based, right? A playground of sorts and easy
to build. An easy infrastructure to build
your agents without having any machine learning knowledge.
Right? If you have machine learning
knowledge, that's good. I think a lot of startups are
building for different machine learning engineers.
I just saw a big gap. Like what about people who wear
shorts, who don't go to peer reviewed conferences and who
(07:36):
will never pick up a research paper and read it?
Who would never understand this complex problem how do we fix
that for and how do we enable everybody to build agents like
how do we make it how do we makea way that is the best way to
build an agent right every I think now everybody everybody's
after that right I ended up pointing a term called pipe pipe
(07:58):
was the first standard for multiagent protocol of source prompt
instruction personalization engine.
Now we and this was like in 2022, late 2022 in late 2023 is
in is when it kind of got some heat around it, right?
Some community being building around it.
Now we have so many things now we have A to a now we have MCP
(08:22):
and I I also read about something called Agent C,
something like that, right? Like everybody has a standard
now, right? So not having a standard is more
cool now. So but yeah, that's me.
We have just launched Chai Computer Human AI.
That is a computer coding agent that has access to a lot of
(08:44):
computers and it lets you vibe code any AI agent.
So it's an agent that can build agents for you.
And these agents are not a formal.
They survive the test of the time.
You can throw in like terabytes of data, they can have memory,
all of that fun stuff, right? So yeah, excited to chat about
that. Yeah, actually, I wouldn't mind
just dying to try right away. I saw the announcement.
(09:06):
It's caught fire on Twitter. Could you just walk me through
what type of person it's targeted for, what the process
is? If I go on and give it a use
case and it spits on some code, is it hosted there?
Could you just walk me through the whole process there?
Just a little bit of a history, right?
So Landbase is a serverless coding platform, right?
For AI agents, we started with pipes and pipes again were
(09:29):
prompts, instruction, personalization engine.
Think of it as a new compute primitive.
Think of it like AWS Lambda but for agents.
If you had a building block thatyou would build on top of and it
would become an agent, that was pipe, right?
And what it allowed people to dowas, you know, connect like 600
(09:51):
different LLMS or a unified API which looks similar to Open AI
and connect your data like for RAG or whatever and build a
workflow, an agent that becomes an agent for you, right?
So personalization layer was thedata layer behind the scenes,
right? And then we built a studio
(10:12):
around it, right? So I think I built, I started
building this thing. This was not a, you know, a
startup back then. I started building this in 2020,
right? So this was like the only
playground of sorts for Open EI back in the day, right?
And how it worked and it still works that way is like, you
(10:32):
know, you can version this. This is you should feel very
familiar vibe of this with GitHub or something like you can
fork it. You can get the code for this or
what not, right? And in here you can build an
agent either by looking by usingthe studio or by using code,
right? Both of these things work,
(10:53):
right? The next thing we did was we
created memory. So memory agents are really near
and dead to my hardware. You upload your data and these
agents have the ability to figure out everything that is
required with their data. It will parse your PDFs, it will
chunk it the right way. You have control on each of
these steps, but you don't need to know all of it if you are
(11:17):
just using it right. And then you can just ask it
questions and let me like give you a demo of what that looks
like. I actually happened to have an
explainer here how they can findif that generally really helps.
And I'm explaining. There you go.
So right now I've created an agent.
I'm gonna prove that this is quite urgentic.
(11:40):
And what this one is doing is I'm asking it to ignore it's pre
training so it doesn't hallucinate, right?
And I've picked Chad GPD 4 old model again.
You have access to like 600 plusmodels here, all on the same
API. I'm asking so many questions.
The most fun question is how tall is the founder of Lang
Base, right? So as a human, it's very easy to
(12:01):
understand that I'm asking for the height of the founder of
Lang Base, right? We have attention by default,
built in LLM's are actually built on top of this idea of
self retention. They can do this kind of thing,
they can figure out the combination and relationship
between different things, right?But if you just run it in one
go, it will definitely going to hallucinate.
(12:23):
So this is not my name and it's definitely not sure how tall I
am, right? So let's do this thing.
Let's give it access to all of our product documentation and
some of my health data, right? So what I've done is I've added,
I've connected 2 memory agents to this pipe agent.
These are agents. They both have data in them.
(12:46):
They have vector stores in them,right?
This is how a memory agent works, right?
In the Lang based memory agent, we have our all of our product
docs. So they know I'm the founder.
And in my health docs it, it probably stays how tall I am
right now if I run this, you will see that it will
automatically reason over these two memory agents and figure out
(13:07):
answers to all of these questions, right?
This is called agentic rag. This is not simple rag.
It's retrieval, but it is agentic and I will tell you why.
So it knows I'm the founder and I'm 185cm tall, right?
So here's what's happening when I ask how tall is the founder?
This should not work with our language agent, memory agent,
(13:27):
because it does not have my height data in it.
It knows who is the founder though, right?
So again, on a similarity search, if I asked this memory
agent MLS, how tall is the founder?
This one doesn't know I'm a founder, so it should also fail
or hallucinate. But overall, as a human, we have
attention. We can kind of connect the
(13:48):
relationship between 2:00. So this is what our agentic
reasoning architecture does behind the scene.
It basically rewrites this question using the context to
how tall is MMS, and then it basically very easily finds the
answer here, right? There's a lot going on behind
the scene, but for you, you're simply sending the user messages
to an API, right? API is just doing a lot for you,
(14:11):
right? And you have control over every
single step here. Like how this reasoning happens,
you control all of it, right? And I can change this to, I
don't know, let's try with something like in Gemini.
Yeah, let's go with Gemini too. For some reason, I don't have
Gemini API keys, but I was just running it.
I'll probably add it later on. I was probably running it here.
(14:34):
I can show you right here. This is GPT 4.
There's not a single demo that won't break live right, right.
So I was just running it here somewhere, I think.
Where are you? There you go right.
So this was a preview run on Gemini 2.0 Flash and this was a
(14:58):
pre run on chat GB forum, right?The completion was again 185cm
tall and again, the completion was 185cm tall, right?
So I had just changed the model on this one, but look at the
difference. If you run it a million times,
this is like $22,000 versus 400 bucks for the same answers,
(15:19):
right? So this happens without you
having to do anything. You can just switch a model from
one thing to another and it justworks, right?
And we talked quite deeply aboutthis on state of AI agents
research that we did. We understood about 200 billion
tokens here and we figured out alot.
We created distillations for each provider for almost all of
(15:41):
their models to figure out how to help you.
It's like Mike, when when you became a founder, there are like
10 different questions people mostly ask you and your brain
kind of keeps a warm memory of that.
Like, this is what I do. This is the name of the company
this, right? This is what our agentic
architecture can do for reasoning, right?
(16:02):
OK, now let's take a step back from this.
And we also try to help people alot in figuring out how to build
all of this themselves. Like we created tools and
infrastructures that would actually help.
We created an explore tab. So you can see like I think
about 20,000 agents here that people are publishing that are
(16:26):
open source. You can see resumes you put
here, whatever, right? You can explore these.
Like here's the scene maker. You can probably go in there and
figure out, you know, what they are trying to do with this or
what not. You can fork it or what not,
right? And then we also published our
agentic architectures piece. These are 8 different agent
architectures that cover most ofthe agents that you want to
(16:49):
build. Like this is my favorite
orchestrator worker, where when based on an input, this agent
can spin up like 200 different agents to solve that problem on
the go, right? Like you don't need a framework
for this. You can see this is like 90
lines of code. There's no framework, plain
TypeScript or JavaScript or Python.
It's just basic API calls, right?
(17:10):
There's nothing you would need to run, you know, learn here,
except for our API versus, again, any API looks like this,
right? And by doing this, we have done
a lot, right? So we've created so many
examples how to use memory threads, this and that, but
there was still something missing.
It was still very technical. And our goal was like, be the
best way to build an agent, right?
(17:32):
How do you make it so simple without you actually losing
control? I don't want to make it like a
drag and drop green Weaver. You are dragging and dropping
and connecting like something like that.
I have nothing against it. It's just as a programmer, once
you have felt the control that code gives you, you can you can
never move away from it. You know the the dynamic nature
(17:57):
of code and how it helps, right?So internally, for last 4 1/2
years, I have been building a coding agent.
And that was just for me, right?You know, when Greg gave me
access, I started building an agent.
I don't use cursor or any other coding agents because my coding
agent is really good. It's like really trained on my
ideas over the course of last five years.
(18:21):
And that agent started doing something weird last year.
It started solving a road map without any supervision and I
was like, holy shit, what is happening?
What we figured out was because we created so many things in
such a good infrastructure behind the scene, our agent was
able to, our coding agent was able to use our infrastructure
(18:44):
to create more agents. You cannot do that with ChatGPT.
Like, you know, like you can go ask ChatGPT to build you a
perplexity. It will not do that right?
But when we asked our agent, theinternal agent, to build a
perplexity, it started building a very many version of
perplexity. And we're like blown away.
Like, holy shit, this is amazing.
And we started using it for all silly things like, you know, in
(19:05):
a POC, when you were telling us,you know, build this and this
type of agent, we would show youlike, oh, we already have
something like this. And people would say, wow, we
never thought that it could become a product, right?
And that is what we launched as Chai.
So Chai, we put human between a computer and AI.
You've created an auto scalable community space infrastructure
(19:26):
where we can scale to thousands of machines to solve a problem.
Right now Chai is not a general purpose agent.
It is built for just one purpose, to build agents for
you, right? No other agent right now can do
that. I am very well connected.
I have not seen any agent havingagentic capabilities to a point
(19:46):
where it can build agents for you.
And those agents are also agenting, right?
It's like a same loop, right? So this is what we have
launched. You know, I'm a huge Chai
drinker, right? So it's just I was looking for a
word that had AI in it in some funds.
A Chai, you know, prompts a friendship, right?
What it can do is you can just ask it to build an agent.
(20:08):
Like, for example, I don't know,make me an e-mail agent.
Oh, there you go. Yeah, log me out, right.
Yeah, it's a bit of product and we are figuring out a lot here,
but let's let's go with something like.
What we were thinking, I'm gonnaask it to build me an agent and
(20:28):
imagine that I'm somebody who has never built an agent, who
may or may not have programming knowledge.
This thing Chai is going to bothcode that agent and build me the
UI and the API for that agent right so that was the discovery
We were like oh, what if like I don't like I'm nothing against
(20:48):
lovable or visa or anything, butI don't need to build like 100
websites. I know how to build a website
yes with lovable or bold or all of that.
It's fun to build a website, butto a point where, oh, what do I
need a website for anymore? Like maybe after a couple of but
dude, I need hundreds of agents for my stuff, right?
And as somebody who's building acompany, I still never got
(21:11):
enough time to build a lot of agents.
You know, like you have to code it.
There's something to be said about no codes it it tends to be
fast, but like, but then you lose control.
You build something fun and thenyou know, I can't do this.
I want to do a four while loop, right?
But like, how do I pull that up,right?
So let's let's do that, right? So I've asked her to, and by the
way, totally expect something tofail live demo, right?
(21:34):
So I'm just asking you to build me an e-mail agent.
So Chai does only one thing right now.
It builds you the agents, not the connections, nothing else,
right? Right now the focus is just to
build the agents, right? And what I've asked it to do is
build me an agent that will readmy e-mail and will perform
several tasks in a sequence. It will analyze and summarize
e-mail and it will figure out ifthis is an e-mail that is not
(21:58):
spam and if it needs a response and maybe then write the
response as well, right. So now it has started e-mail,
you know, telling me what it is doing e-mail processing agent.
This is what I'm doing. The workflow it is building will
simultaneously analyze. This is important.
I these are two separate tasks. I want them to happen in better.
(22:22):
And then this should be a block gate where, you know, we need to
determine if our response is needed.
Then we use that all of the context to build that agent.
Now, this is really easy to build.
We have a bunch of guides on this as well.
Like I think we call it a composable AI e-mail agent
architecture. Like when you get a spam e-mail,
basically, you know, they basically figure out that this
(22:44):
is a spam e-mail. It doesn't do anything.
If you get a valid e-mail, like you know your API is broken, the
CM maker will create a response quickly and start writing the
response. Like, we've talked about this a
lot in our docs, right? But what Chai is doing here is
different. Chai has gone ahead and built
something for me. I'm gonna go and deploy it.
So while this is deploying, let's review what the code looks
(23:06):
like. So in the code, it has created
an agent. This part is analyzing the
sentiment. I can explain it and prove it,
whatever you want, right? So I can like just say, you
know, this part is not doing what I want to do this, right?
That part is there. So this is the step one, and
we've made sure that our workflow is extremely simple.
There's a step, there's an ID, and then there's a run function
(23:27):
where you just add whatever you want, and it's probably giving a
response. It's already created the UI for
it, by the way. It's just deployed it, right?
But let's go back to the code, analyse sentiment and it has
created a schema. This is more advanced than I
thought it would like, you know,sentiment or whatnot.
Like as a programmer, I would not probably do this in one go.
I would be like, you know, I'll get to that point someday.
(23:50):
And then it is feeding that too.It has decided to use GPD 4.1
Mini. So we have trained a lot of
things on different models and we are continuing to do that.
It can basically understand whatthe task is at hand and what
would be the right model for that.
And you can optimize that. You can ask, you know, make the
responses much better. So use a premium model or maybe
use something that I could use for free like from Gemini or
(24:11):
what not. And it has also has the ability
to create prompts. You can obviously improve these
prompts, right? But this was the first agent,
this is the second one. Summarize the e-mail.
This is the step where it is determining if the response is
new year. It has also created a schema for
that. If the response is needed, why
(24:33):
is it needed? The reason, the priority, right?
It's automatically doing this. You can completely change this.
You can just say, you know, add more labels here over now and
then I think it is just simply generating a response.
If the response is needed, thereyou go, right?
The Step 3. And by the way, you don't need
to look at this this way. We also give you a flow which
(24:55):
you can just see and this is what it's doing and it's gonna
process the e-mail sentiment, somebody determine if it is
needed, if it is needed this notskip log it and be done right
and you can regenerate it or what not soon.
We'll also give you the traces. You know, you can see how this
is running and if there's a memory required, the memory is
also created for this use case. Memory is not required.
(25:17):
So well, here's what the agent UI looks like.
Let's give it something like, I don't know what was the e-mail
here. I can just probably write
something. Your billing API is broken.
This is not good, right? So now it is processing all of
that for flow, right? It is figuring out if we need a
(25:39):
response or what not. And this is just a silly UI,
right? But I think what we have sort of
ended up inventing is nobody wasbuilding UI for agents,
especially not dynamic US. Like what would a e-mail agents
UI look like, right? So here's the summary, Here's
the sentiment analysis. Sentiment is negative.
You know, urgency is high. And it has also gone ahead, told
(26:03):
me the reason why it should be responding.
And it has also drafted a response, right?
And this is pretty powerful, right?
I just created this thing, right?
And I don't have to use this UI,I can just use it based on this
API, right? The Node Python call, I can just
send in my e-mail from whatever to and it will do exactly this,
(26:26):
right? So it is an actual agentic
workflow where multiple agents are doing multiple things,
right? And for a lot of agents, what we
have figured out is a simple UI is all you need, right?
So I don't know, you won a dollar 100 million lottery.
Sometimes it doesn't do the spamcheck really well and I have to
tell it like you're letting the spam pass through.
(26:50):
Let's see. OK, Summary emergency is higher.
It's pretty positive. Yeah.
And there you go, not respondingto this one, right.
I guess it's probably a spam, right?
Yeah, right. So sometime it doesn't do it.
It starts responding. And then I have to like come and
tell it like, you know, I don't want you to do this, right?
(27:11):
And then there's so many that I've created.
I think you have a question. Yeah, sorry, just to jump in.
So this super impressive click of a button or basic prompt can
generate all of this, I assume with the deploy button, what
not. This can all be hosted on Lang
base. So I can just kind of go in,
create a little army of agents and just by giving my prompt.
This has all the infrastructure needs, the ability to expose the
APIs, and then I can just kind of operate as I would anything
(27:33):
else that's just cloud host and serverless.
Yeah, that's that's the point, right.
So you don't have to worry abouthow to scale this.
You can send it like 100 millionemails a day and it will just
process it very easily. We are generally doing a hundred
150 million agent runs a month at this stage, right, So all of
this is deployed. Whatever we needed to work out
(27:54):
for this deployment happened behind the scene and also this
UI is also deployed. I know it's beta.
You don't see a URL right now, but we'll the first thing we are
launching this week is you should be able to share this
agent with anyone, right? Like you know what so source
graph actually uses Lang base todo this.
You know, AI bought in their docs, right?
(28:18):
So somebody in our docs team where they had an idea was, you
know, it's because it's always me saying something like this.
I don't know, like, let's figureout some feature pages.
Our feature pages are not great.So sorry everybody in our docs
team, right? So versioning, OK, versioning is
good, right? So fork, let's say I don't like
this page. And what I'm gonna do is I'm
(28:39):
gonna copy this link and I'm gonna go to Slack and I'm gonna
say maybe this page should have more content, right?
So what our docs team has done now is let's go over here
somewhere, right? So what they have done is the
moment I post something in Slackthat, well, this URL, this page
(28:59):
is not great. It should have like more content
here or what not. They think that this is a
prompt. They throw this to Chai
automatically. They have built a small dog
spot. Chai actually builds all the
agents needed to fix this because Sai has computers in it.
(29:21):
It is connected to GitHub, it can download data from GitHub.
It can figure out where the filefor this link is.
Our docs are on GitHub, right? And then it runs those agents
and that feedback to fix that problem.
And then it finally creates a PR.
Now this workflow is not new. You get a PR and you have an
(29:41):
improved doc. This workflow is not new.
You could do this with anything.What's new is the next time I do
this, they will not send, they will not tag Chai.
They will have a very defined particular workflow that will
work the way they want, right? It's like, you know when you do
one thing with Cursor, it's, there's no guarantee that the
same thing will happen next time.
(30:02):
If you add an API endpoint and you're really happy with it,
Cursor has not learned anything.You just have prompted a lot.
And when you're building something with Chai, you're
actually building real life agents that are there that you
can continue to improve. Like for example, I've built a
very mini version of API Route Creator.
We have a very different way of creating API routes internally.
(30:23):
I couldn't teach that to cursor.Maybe I could add cursor rules
or whatnot, which are just prompts.
Agents are agentic, right? So now I have a API Route
Creator agent through Chai that works locally.
By the way, it's not deployed anywhere.
You can ask, try to create something, you can take this
code and just use it locally, right?
So that is still true. You can use it with Olama.
(30:45):
We integrate with Olama, right? Maybe you don't like this UI or
you know you have some update inthe UI.
You can go here, ask some changes for UI, right?
One more thing that I really like is like I've been building
so many like I've not built thatmany agents myself, but right
nowadays I've been like, I don'tknow, like here's here's a
perplexity one, right? So a little bit perplexity.
(31:06):
I think I did one shot where I asked I make a deep researcher
like perplexity. And it basically tried and I
deployed it, but then when I ranit, I got this error, some
schema or something. I didn't even read it.
It's wipe coding anyway, right? So I ask it like, fix this
error, right? And in the second shot, I think
(31:28):
it got it right. So this is what we have now.
This is our deep researcher. I don't know, like let's what
did Open AI launch last month, which was April?
I don't know if they launched anything.
I don't know what it's gonna do.It's using EXA behind the scene
(31:49):
excise, you know, Internet scraping or search tool of
sorts. There you go in April 2024.
Oh, there you go 2024. Like it's obviously it's not
publicity, but it's working, right?
And like you can basically ask it like, I don't know what has,
(32:10):
I don't know if this will work on X lately, right.
So this is, I don't know if Twitter allows you to scrap
data, but I'm pretty sure their API is pretty down but exists
pretty good. It should probably bring
something up. It's waiting watch, it's taking
it's free time. There you go.
(32:30):
OK, some of my tweets by script.It's bringing up old data.
I think what's happening here isEXA does not have access to the
live X data and it's just when Twitter was open and they
allowed it to scrap. That is what they're bringing
up. But it's pretty good
architecture like you can see how it is happening, how and
what it is doing and people havebeen building so many fun stuff
(32:51):
like for example, a simple thingis as simple as like I don't
know, chat with PDF, right. So it should probably just build
a memory agent now where I should be able to upload APDF or
any file and talk to it, right? So this common problem, right?
While it is doing that, I can also probably show you a little
(33:13):
bit. I'm probably gonna zoom in a
little bit. So we don't see their names.
So there are usernames here. And by the way, in five to six
days, 35,000 agents have been built with Chai.
Wow. Let the sink in right?
People are building amazing stuff like this.
Is today's May 6th you look at the time stamps here it's almost
(33:35):
every other minute there's like an agent being built by somebody
right so PDF chat, Firefly summarizer, e-mail responder,
mail sorter, some of the. Most compelling agents that
you've seen developed, whether with Chai or otherwise?
I don't know, it's really up to you.
Anything. People are building so many
things. Like I saw somebody from
(33:57):
Change.org create a partition agent which scraps your
Change.org partition and helps you both write the content and
talk to people in comments. I'm like, holy shit, right here
you go Forex trader, like all sorts of fun stuff.
My biggest really, which I was that people are were gonna go
and use it like ChatGPT and thatwould mess it up, right?
(34:20):
Like, for example, I know if yougo to ChatGPT and you say make
me our deep researcher like perplexity, it's probably and
let's go over there. I don't know what it's going to
do. It's definitely not going to
build me an agent, right? So I.
(34:41):
Thought the mindset. I don't know what it is doing
right. So I thought people will start
using Chai as Chai GPD or cloud or something and that would
totally not be cool. But but unbelievably, like 95%
of people are like, you know, look at this like NAV centric
and this is just somebody going,you know, nutrition planner
(35:02):
mind. These people are building all
sorts of fun stuff. What they are doing is like task
refining. What it takes to build an agent
is automated now, right you I have, I've been asking people to
not use frameworks because they slow you down.
Use primitives. Now it doesn't even matter
because something else is writing code for you.
You have the complete control. This is not DSL as you can see.
(35:26):
Let me deploy it basically went ahead like this is not DSL.
This is not some weird weird configuration file that is super
vendor locked into LAN base. It is an actual code written for
you for the purpose of your agent that you are building,
right? It's, you know, like how would
you build a generate answer agent?
Like it has also created a really good, you know, prompt
(35:49):
for that purpose as well. So I've been having so much fun
just reading through what peopleare doing.
NER data scout agent. I know what that is.
Race engineer fit plan or whatnot.
Lightning code expert. A lot of people are actually
building coding agents. A lot of people have ideas of
never does this or love people never does this does that right?
(36:10):
They don't feel like they have the control over improving the
agent they are using. With Chai, you are actually
building the agent that does what you want and you have the
control of improving that agent over time.
As your project grows, your requirement requirements grow
over. So the people are building all
sorts of fun stuff here. Like I just like there are 54
pages of these. These are just workflows, right?
(36:31):
Yeah, actually a. Quick question on that, Yeah,
when someone makes their own coding agent, because there's so
many options out there and they want to make it best for them,
what would you say are some gooddeveloper experience principles
that people should follow or just be aware of so that the
tooling that they build is goingto be beneficial to them rather
than just something that's flashy?
I don't know if do you know about Tailwind.
(36:51):
Slightly familiar, but do you want to tell the audience?
Yeah, so Tailwind is this framework that was created by
two amazing designers and developers Adam and Steve in
for. For someone as technical as me
who always wrote CSS from scratch, I had like this weird
think about frameworks. Like I can't use CSS frameworks.
I'm a Bootstrap contributor, butI don't use Bootstrap.
(37:15):
It's just very, very weird when talking about it that way,
right? And for the longest of the times
I was like, I don't like Tailwind.
And then it hit me one day that Tailwind is not a framework,
it's actually an API, a primitive of sorts, which allows
you a defined way of building CSS.
And they have really pretty and really awesome, amazing docs
(37:36):
that anybody can read because everybody who's building, I was
writing CSS from scratch of their own ways.
I had my own way. But you on my team, if I hired
10 more people, they didn't knowmy way and took me time to teach
them. So Tailwind built really good
documentation, really good primitive that anybody could
learn. And I could just hire somebody
who knew Tailwind a little bit. They would create anything
(37:58):
because it's not a framework, it's just an API and you can do
anything with API, right? And there was this curve I had
where I was like I went from, you know, like I think like, I
don't know, there was a like a no to tailwind and oh, this,
this is pretty cool, right? Like my emotion emotion went
like this, right? This is the same thing that has
(38:18):
happened with vibe coding, right?
So vibe coding can both be helpful or very fleshy and
wrong, right? So what I see people do is what
they do is which is by the way, what I'm doing here right?
So chat with PDO. This is the wrong way to wipe
code by the way, it's fun like for example, it has gone ahead
created this PDF document memory.
(38:40):
I'm gonna for the sake of this chat, I'm gonna upload something
here. Let's upload this has some data
on my this is the home page of my personal website, right?
And I'm probably going to throw in another text file as well,
which has which is the API keys portion of our docs, right?
(39:02):
Just to ask these questions and it's probably going to go ahead
and process this in a minute while we chat, right?
There you go. These are ready.
Let's go to the UI and let's askme what is the name of Emmett's
company and how do we get API keys from that?
(39:26):
Fingers crossed. I hope this works right.
So I just added two documents, one that knows who am I.
So it should probably figure outthat the new other companies
land is there you go and you basically go to the settings
menu and there's this and there you go.
I just asked chat with PDF and this is what I got.
I can use it through this UI. So you can use this as an API or
(39:50):
UI. It really is up to you, right?
So what a lot of people do is they get this weird wrong idea
about vibe coding. So I think vibe coding.
So I'll take a step back. I think Vibe coding is going to
replace no code and no code. Right, like we always had like
no code, low code website builders.
(40:10):
We had green Weaver, I don't know, like 1990 or something.
You could drag and drop and build a website.
But even then lovable and bold exist and they give you code.
You don't need to use that code.But if you want to or if you
have someone who can help when you need it can help, right?
Like if you throw, you know, your no code to tool to
(40:33):
developer and they were like, Oh, I can't help with that.
They'll be like, you know, I'm going to build you something.
I'll charge you like Python bugsfor that or not.
Or you can just ask like, this code is doing this.
Can you just do this small implementation that I'm not able
to figure out? That's super powerful, right?
So for the longest of the time Iwas like, ah, white coding is
weird to a point where I figuredout, oh, white coding is
actually replacing no code and no code with actual code that
(40:57):
you still have to not put in a lot of effort to build, right?
That is super powerful, right? So I got over that hill where
I'm like, OK, vibe coding is great, right?
And this is coming from someone who's deeply technical, right?
But what people get wrong is, let's say you want to get to, I
don't know, 10, right? And what people try to do is
(41:19):
they try to do 5 + 5 or 6 + 4 equal to 10 with prompt
engineering or vibe coding. What actually generally works is
1 + 2 + 3 plus, I don't know whatever the rest is.
To get to that result right, Youhave to take it step by step.
It's similar to let's say you hire an employee or let's say
(41:41):
you're looking for a Co founder.You don't just throw all of the
context at them. You don't just say, you know, be
my Co founder and they'll be your Co founder or be my
employee and do all of the things that I expect.
You have to slowly build that relationship, slowly build the
understanding of the feedback loop of loop of you know where
(42:02):
now build this now do this now do that right?
Like a human in the loop of sorts, right?
That generally works really well.
So what I often tell people is start with something simple,
very simple. So you know, it works.
You get familiar with. So Agent DSS, the agent, agent
TSXS, the UI, which I can host anywhere.
(42:23):
It's a simple React code, right?It's conveniently hosted for
you. But if you wanted to take this
code and do whatever the fun thing you want to do with it,
yeah, be my guest, right? Then you slowly build this thing
up into something. I don't want to name names
because I don't have permissionsfrom them, but there's this
company that in the last five days, I think they have created
hundreds of agents, which I in there this the CEO is doing
(42:48):
this. They're pushing the limits with
China. And I'm like, I'm mostly looking
at their chairs. I'm like, oh, oh, it can do
that. I didn't know, right?
Like I built this thing. But they're like, you know, let
people upload their logos and from their logos.
It's a swag company, by the way,like their very common name.
Everybody uses them to build their T-shirts or stuff like
(43:10):
that. When you upload a logo, they are
able to figure out through theirinventory using an agent that
they are continuously building, which I that what can they
deliver? What colors by when and what is
like possible. This is like an extremely
complicated job. And I think that he's like, you
(43:30):
know, probably 3050 messages in one thread.
So now we are thinking about like, oh, this is too much.
We've just figured out like you cannot have longer as long chats
and shared GPT or Claude. But we we create a child for
these agents not to be formal. They are always going to be
there and you'll always try to improve them, but you cannot
(43:51):
have 10,000 messages here. So now we have this user
experience issue where we're trying to figure out, OK, maybe
we create some sort of done withthe last 10 versions, stash them
away. Let's summarize the context and
be here. We're calling it something weird
internally where we are very excited about that feature,
right? So anybody can use it.
(44:13):
It is just how you use it that matters.
You could be less technical and you could have much better
identic ideas than I do, which is what is happening.
I, I go look at the space and I'm like, oh, style guide,
etcetera. People are building things that
I didn't know we could build, you know, with this, OK, I think
I am on page one, which we just saw, but I should probably be
(44:37):
zooming in. People are building all sorts of
fun stuff here that I never likefought from Spring agent right.
So I don't know Swift UI could build a French wedding maestro,
right? Like anybody can build anything,
right? It's, it's the fact that it gets
(44:58):
you to, in most of the cases, 9095% done.
And then you have still the complete control over chatting
and fixing its behavior, improving it or starting fresh
or bringing in an actual expert,like, you know, somebody who's a
developer, who's a developer on your team.
Let's say you're a PM and you'rebuilding something.
(45:19):
It's fun, it's working. You have bookmarked it in Slack.
And now there's a bug that you just cannot figure out, right?
Maybe you used to code like 10 years ago and now you're not
brushed up on your React skill. You just tease your front end
guy a little bit or girl a little bit and they can just
say, oh, just fix this and be done right?
(45:40):
That they cannot do with a no core tool, right, right.
So Yep, what do you think? I think it's incredible.
I did have one other question for the For the last remaining
5%, if someone were to develop an agent gets pushed away there,
pull it into local Idi. Remember you mentioned that you
don't use cursor your VS Code because you have your own coding
(46:03):
agent, but what advice or tools have you seen in the AI age for
development that you think are useful?
Does MCP help with your dev experience, or are there any
other tools that you think people should start exploring as
they're trying to build to get these last few percentage points
over the line? Anything really that helps, you
know, back in the day. So I'm here today because I
(46:23):
started writing content and I started blogging when I was very
young, I think late 90s and all of my careers because I started
writing one day. I remember watching Cristiano
Ronaldo 2001 FIFA, I think Euro World Cup and I'm like, this is
awesome. I should write about it, right?
So you should just use tools, right?
(46:45):
No matter what, you don't know where what one tool would lead.
You're taught to. I love reading source code.
Like you can often find me like at 1:00 AM in the morning
reading the core code of Expressjust because, right?
I love to just read source code as in read as in just look at
it. I think my brain picks up on
weird things that later on and I'm like, I don't know how I
(47:08):
know this, but I know this, right?
So MCPS like this, right? MCP, every tool has pros and
cons, just having the knowledge,I think.
So here's my advice about that. When I was young, when I went to
college, I thought if I knew a lot of programming language I
would languages I would do really well.
(47:28):
So I started like, I don't know 26 different programming
languages. And for the longest of time that
turned out to be the wrong choice.
I ended up just getting really good at just a couple of
languages like JavaScript, TypeScript, Node, C, Python,
right? But now I really enjoy the fact
that I know more than 20 programming languages.
(47:49):
I know there is a solution in the lexler that can fix this
concurrency issue struggling with I know how to think about
it and if I give the right and wrong, these agents can do fun
stuff, right? So my advice from over the years
has now changed. I'm like read everything.
Potentially you'll have this is agents are just a new type of
(48:11):
code. They are going to be very
helpful and tools are going to be your augmentation.
I think humans have this capability.
So legs and arms are our tools. My father is a doctor.
I asked him how does our brain learn things?
And he's like, so legs and your hands are tools, but your brain
has the capability to create newtools.
(48:32):
Like you can learn to drive a car using these four tools,
right? And that capability is what they
call a process that he has this entire diagram of C1C2C3
cognition levels where you get effects, you understand effects
on C2, then you basically understand what to do or what
not to do, like reasoning. And 4th 1 is planning.
(48:53):
Like even a dog can reason like if it should, you know, sit on
your bed or not, right? But they cannot plan like
humans, right? I think we are going to have
planning tools. LLMS are going to get better.
And you, the human is going to be the best planner there is.
You will be augmented in the middle, right?
Like computer, human AI, you're the planner here.
(49:15):
You need to know all those toolsto understand the knowledge.
You may or may not use MCP, right?
But you may like something from MCP that you can ask an agent to
build for you, right? Or you can just use one MCP to
do that. One thing.
MCPS are cool. One thing I don't like about
MCPS are do you know what is a CLI like?
(49:38):
You know you have a command linetool and it does exactly what
you want. I think MCPS are poor man's
Clis. People who don't know how to
build a CLI can use MCP to potentially automate something.
The only problem I have with MCPis that it just consumes a lot
of tokens, right? Just to do a simple thing.
(50:01):
Like I was talking to Claude theother day about my Stripe
account. He did one analysis and he was
like, you're done for the day. Come back 5:00 AM.
I'm like, what? Right.
So I just only asked, like, you know, how many pro subscriptions
of Chai happened in April, in May, and it just consumed my
entire day's worth of tokens. That was unexpected, right?
(50:21):
I could have just built a CLI, one shot at it and done that,
right? So each tool, knowledge of those
tools will actually help you understand if that is something
you need or not need. And someday that knowledge will
actually really be helpful. Now you have all heard and learn
something because it will be helpful someday.
Now it can actually be helpful. Any knowledge of any technology
(50:44):
can actually be helpful because you don't actually need to learn
everything. You don't need to learn React
view and everything. But if you just know React, you
can kind of build something withview because these agents and
agentic systems and LLMS are getting really, really good at
it, right? So yeah, MCPS, well, definitely
try them out. So as someone who's super
(51:04):
interested in space and loves Team Open Source, could you tell
me a story about how you got your code running on another
planet? So that's a fun story.
When I was young, I wanted to bean astronaut and an astronaut
who was absolutely afraid of heights.
So I kind of settled for being AI, thought I'll be a scientist,
(51:28):
right? I'll invent things that would go
in space. You know how kids are, right?
And a lot of people say that, you know, two of the hardest
thing in computer science are naming things and cache and
validation. And I really like naming things
because I wanted to be a scientist, right?
So like, I wouldn't rinse stuff,right?
I think all of that led to small, little, pretty small
(51:52):
little contribution in the NASA Infinity Mars helicopter
mission. In fun fact, it was actually AUI
code that they are running in the Ingenuity helicopter
mission. Like, you know, who knew being a
web developer would actually, you know, you can, you can throw
your code to Mars, right? So, so yeah, and that happened.
(52:12):
I was like super excited when they told me like, you know,
this is the project where this code is going to be.
And then I later on and got a little bit more into it, you
know, got to see the project a little bit more from, you know,
like you know what that actuallycould have been even in an
alternate universe being espirant and also got down to
(52:32):
side down with Satya Nadella andtalk about it like, you know,
like what it felt like. Like I think about 10,000 or so
developers whose open source board actually made the Internet
E helicopter mission possible. And for me, open source was how
it happened, right? It's not like, you know, I got a
job at NASA and did that, but you know, I eventually got to do
(52:55):
the same thing. So it was super fun, you know,
So we're really excited that that happened.
I know my mother and father werelike, oh, so you do do important
stuff. I'm like, Oh yeah, yeah, yeah.
So, yeah. Fun story.
Love it. Yep, encourage ever to export
open source. You never know what happens.
Exactly. Ahmad, this was a blast.
(53:17):
Thank you so much for sharing your perspective.
Chai looks awesome. Courage everyone to try it
before I let you go. Is there anything you want the
audience to know? No, that was awesome.
Thanks a lot for having me read the source code.
Use your code for good. Absolutely.
All right. Have a great day.
We demand brilliance for our AI agents, but what about the
experience for their creators? The truth is, building powerful,
reliable AI agents can be a messof complex tooling, steep
learning curves, and frustratingroadblocks.
Friction is the enemy of innovation.
So how do we tear down these barriers?
What does a world class developer experience for AI
agents look like? Today, we're exploring the
crucial yet often overlooked world of the developer
(00:20):
experience for AI agents, Essential knowledge for anyone
building, designing, or dreamingof the next generation of AI.
Welcome to episode 41 of Tool Use, the weekly conversation
about AI tools and strategies toempower forward thinking minds
brought to you by Anetic. I'm Mike Byrd, and this week
we're joined by Amit Awais, the founder and CEO of Langbase, A
serverless platform for buildingAI agents.
Ahmad, welcome to Tulius. Thank you for having me here
(00:41):
Mike. Super excited to talk about
agents. Oh, it's always a blast.
Would you mind giving the audience a little background how
you got into AI and what got youhere today?
Yeah, of course. I think I've been around the
block for like, I don't know, 25years now, building one thing
after another. And I remember back in the day,
I started contributing with the word background, you know, where
(01:02):
we were building Web 2 point O, which is now known as HTML5,
right? And that led me to a career in
open source. Like I ended up contributing a
lot to the core software of WordPress, then a lot of
JavaScript libraries, then the core of Node JS.
There was a time when I was leading the community
committee's outreach program forNode JS, right?
(01:24):
Got deeply involved with Open API Enterprise Governance Board
and got on a Google developers Advisory Board.
I'm practically very, very technical.
I've built a lot of software. I have like I think created
about, I must say, more than 150open source packages and they
get like 40-50 million downloadsevery year.
(01:47):
And I have like one of the top 10 VS Code themes called shades
of purple. I love the purple color.
And so, so, so, so is the name of the theme, right?
And over the years I have done alot of open source advocacy.
I remember about I think 11 years ago, I got Facebook to
open source React and graph QR, right?
(02:08):
That was when my career absolutely changed.
Like I saw first hand how React became the other half of the
Internet is how I call it. Like half of the Internet is
WordPress and half of the Internet is React, right?
So I gently tend to joke about it like there's no website you
cannot visit without my small little my, you know, piece of
tiny contribution that I might have made, right?
(02:28):
I take pride in that, right? And I remember like, because I
became this developer advocate of sorts, I ended up doing a lot
of talks, right? People wanted to understand,
especially tool use, right? Understand how I'm using these
tools, what I care about as a developer.
Why this? Why that to a point where when I
(02:50):
switched from Sublime to VS Codein 2017, when VS Code was like a
new editor, right? Everybody was like, you know,
how are you doing this? Like what is this thing?
What is that, right? To a point where I ended up
building a course called VS Codedot pro.
I think I've helped probably more than 56,000 developers
switch to VS Code over the years, right?
So and about Lang base, I'm now building language, right?
(03:14):
So Lang base is serverless AI platform.
So anything you want to do with agents, from building an agent
to deploying it and scaling it, that is what we do.
And we are contrary to, you know, the normal wisdom that
people talk about like, well, everybody's building an AI
(03:34):
framework. I don't think you can build an
AI framework in this space. I think the space is moving too
fast and things are breaking tooquickly.
I think what we need are primitives, good primitives,
like for example AWS has S3A primitive where you can upload
your images and download them. You don't really care how they
(03:56):
are doing the scaling and charting behind the scene.
You just get a simple API and that API just works.
It's similar to how open AI, your developers experience with
open AI is like you get an API for their models and it just
works, right? You don't really get to see or
have to worry about how they arescaling their models, right?
So at LAN base we are building AI primitives, A primitive for
(04:19):
agent workflows, A primitive foryou know, building an agent
called Pipe, a primitive for agents with memory like human,
human like long term memory which have vector stores in
them, a primitive for threads. So you can easily store
coordinated database problem problem, you know solve the
(04:40):
problem of coordinated database with threads like you know, all
the tool calls and tool results should be coordinated in in the
same order and as fast as possible.
Then parse, search, anchor and Bradley, you name it, right?
We have a bunch of primitives, right?
And how all of this happened wasin 2020, I was travelling a lot.
(05:00):
I had like, I don't know, 43 conferences.
That was going to be the year where I was.
And I, I absolutely hate travelling.
You know, I'm not adventurous atall, you know, like programming
was my hobby. I'm practically boring.
This is all I do, you know? But in 2020 I decided, OK, let's
(05:20):
change things, right? And well, you know what
happened? COVID happened, right?
All my, you know, plans to travel got deprecated.
I remember one of the conferences I was going to share
stage with Obama as one of the keynote speakers.
I was super excited about it. And then I got to do it on hop
in online platform and it was sobad he didn't even show up
(05:43):
right. Anyway, I was I, I found myself
building this thing called Corona CLI and Corona CLI was
what perplexity is, but only forCOVID data and in 2020, right?
So you could ask a dumb questionlike if I go to Italy, will I
die? Or if I go to, I don't know,
like pick up TV from Costco on 4th street.
(06:06):
But what what happens, right? And it went absolutely while I
was just building it because I was travelling a lot and it went
viral when COVID became a globalpandemic, right?
And I meanwhile, like, no puns and like 10 billion API calls in
a month or so, right? And that got me an introduction
to Sam Altman and Greg Brockman.And this is 2020.
(06:28):
Like, open AI is not a big deal in 2020.
Nobody knows about open AI. You know, I didn't know about
much about open AI. They gave me access to GPT 3.
They had just finished training of GPT 3.
And I think June 2020. And I was talking to them in
July 2020, right? And I remember talking to Greg
Ruckman and telling him like, I want to suggest the next line of
(06:50):
code, right? And this is, this is like a year
before GitHub Gopalit was launched.
GitHub Gopalit was launched I think in October 2021, right?
So I was already excited about like this use case of like
suggesting something new. And I did a couple of workshops
for an open EI and a bunch of things.
Basically, I built a couple of software that eventually became
(07:12):
land based, right? A playground of sorts and easy
to build. An easy infrastructure to build
your agents without having any machine learning knowledge.
Right? If you have machine learning
knowledge, that's good. I think a lot of startups are
building for different machine learning engineers.
I just saw a big gap. Like what about people who wear
shorts, who don't go to peer reviewed conferences and who
(07:36):
will never pick up a research paper and read it?
Who would never understand this complex problem how do we fix
that for and how do we enable everybody to build agents like
how do we make it how do we makea way that is the best way to
build an agent right every I think now everybody everybody's
after that right I ended up pointing a term called pipe pipe
(07:58):
was the first standard for multiagent protocol of source prompt
instruction personalization engine.
Now we and this was like in 2022, late 2022 in late 2023 is
in is when it kind of got some heat around it, right?
Some community being building around it.
Now we have so many things now we have A to a now we have MCP
(08:22):
and I I also read about something called Agent C,
something like that, right? Like everybody has a standard
now, right? So not having a standard is more
cool now. So but yeah, that's me.
We have just launched Chai Computer Human AI.
That is a computer coding agent that has access to a lot of
(08:44):
computers and it lets you vibe code any AI agent.
So it's an agent that can build agents for you.
And these agents are not a formal.
They survive the test of the time.
You can throw in like terabytes of data, they can have memory,
all of that fun stuff, right? So yeah, excited to chat about
that. Yeah, actually, I wouldn't mind
just dying to try right away. I saw the announcement.
(09:06):
It's caught fire on Twitter. Could you just walk me through
what type of person it's targeted for, what the process
is? If I go on and give it a use
case and it spits on some code, is it hosted there?
Could you just walk me through the whole process there?
Just a little bit of a history, right?
So Landbase is a serverless coding platform, right?
For AI agents, we started with pipes and pipes again were
(09:29):
prompts, instruction, personalization engine.
Think of it as a new compute primitive.
Think of it like AWS Lambda but for agents.
If you had a building block thatyou would build on top of and it
would become an agent, that was pipe, right?
And what it allowed people to dowas, you know, connect like 600
(09:51):
different LLMS or a unified API which looks similar to Open AI
and connect your data like for RAG or whatever and build a
workflow, an agent that becomes an agent for you, right?
So personalization layer was thedata layer behind the scenes,
right? And then we built a studio
(10:12):
around it, right? So I think I built, I started
building this thing. This was not a, you know, a
startup back then. I started building this in 2020,
right? So this was like the only
playground of sorts for Open EI back in the day, right?
And how it worked and it still works that way is like, you
(10:32):
know, you can version this. This is you should feel very
familiar vibe of this with GitHub or something like you can
fork it. You can get the code for this or
what not, right? And in here you can build an
agent either by looking by usingthe studio or by using code,
right? Both of these things work,
(10:53):
right? The next thing we did was we
created memory. So memory agents are really near
and dead to my hardware. You upload your data and these
agents have the ability to figure out everything that is
required with their data. It will parse your PDFs, it will
chunk it the right way. You have control on each of
these steps, but you don't need to know all of it if you are
(11:17):
just using it right. And then you can just ask it
questions and let me like give you a demo of what that looks
like. I actually happened to have an
explainer here how they can findif that generally really helps.
And I'm explaining. There you go.
So right now I've created an agent.
I'm gonna prove that this is quite urgentic.
(11:40):
And what this one is doing is I'm asking it to ignore it's pre
training so it doesn't hallucinate, right?
And I've picked Chad GPD 4 old model again.
You have access to like 600 plusmodels here, all on the same
API. I'm asking so many questions.
The most fun question is how tall is the founder of Lang
Base, right? So as a human, it's very easy to
(12:01):
understand that I'm asking for the height of the founder of
Lang Base, right? We have attention by default,
built in LLM's are actually built on top of this idea of
self retention. They can do this kind of thing,
they can figure out the combination and relationship
between different things, right?But if you just run it in one
go, it will definitely going to hallucinate.
(12:23):
So this is not my name and it's definitely not sure how tall I
am, right? So let's do this thing.
Let's give it access to all of our product documentation and
some of my health data, right? So what I've done is I've added,
I've connected 2 memory agents to this pipe agent.
These are agents. They both have data in them.
(12:46):
They have vector stores in them,right?
This is how a memory agent works, right?
In the Lang based memory agent, we have our all of our product
docs. So they know I'm the founder.
And in my health docs it, it probably stays how tall I am
right now if I run this, you will see that it will
automatically reason over these two memory agents and figure out
(13:07):
answers to all of these questions, right?
This is called agentic rag. This is not simple rag.
It's retrieval, but it is agentic and I will tell you why.
So it knows I'm the founder and I'm 185cm tall, right?
So here's what's happening when I ask how tall is the founder?
This should not work with our language agent, memory agent,
(13:27):
because it does not have my height data in it.
It knows who is the founder though, right?
So again, on a similarity search, if I asked this memory
agent MLS, how tall is the founder?
This one doesn't know I'm a founder, so it should also fail
or hallucinate. But overall, as a human, we have
attention. We can kind of connect the
(13:48):
relationship between 2:00. So this is what our agentic
reasoning architecture does behind the scene.
It basically rewrites this question using the context to
how tall is MMS, and then it basically very easily finds the
answer here, right? There's a lot going on behind
the scene, but for you, you're simply sending the user messages
to an API, right? API is just doing a lot for you,
(14:11):
right? And you have control over every
single step here. Like how this reasoning happens,
you control all of it, right? And I can change this to, I
don't know, let's try with something like in Gemini.
Yeah, let's go with Gemini too. For some reason, I don't have
Gemini API keys, but I was just running it.
I'll probably add it later on. I was probably running it here.
(14:34):
I can show you right here. This is GPT 4.
There's not a single demo that won't break live right, right.
So I was just running it here somewhere, I think.
Where are you? There you go right.
So this was a preview run on Gemini 2.0 Flash and this was a
(14:58):
pre run on chat GB forum, right?The completion was again 185cm
tall and again, the completion was 185cm tall, right?
So I had just changed the model on this one, but look at the
difference. If you run it a million times,
this is like $22,000 versus 400 bucks for the same answers,
(15:19):
right? So this happens without you
having to do anything. You can just switch a model from
one thing to another and it justworks, right?
And we talked quite deeply aboutthis on state of AI agents
research that we did. We understood about 200 billion
tokens here and we figured out alot.
We created distillations for each provider for almost all of
(15:41):
their models to figure out how to help you.
It's like Mike, when when you became a founder, there are like
10 different questions people mostly ask you and your brain
kind of keeps a warm memory of that.
Like, this is what I do. This is the name of the company
this, right? This is what our agentic
architecture can do for reasoning, right?
(16:02):
OK, now let's take a step back from this.
And we also try to help people alot in figuring out how to build
all of this themselves. Like we created tools and
infrastructures that would actually help.
We created an explore tab. So you can see like I think
about 20,000 agents here that people are publishing that are
(16:26):
open source. You can see resumes you put
here, whatever, right? You can explore these.
Like here's the scene maker. You can probably go in there and
figure out, you know, what they are trying to do with this or
what not. You can fork it or what not,
right? And then we also published our
agentic architectures piece. These are 8 different agent
architectures that cover most ofthe agents that you want to
(16:49):
build. Like this is my favorite
orchestrator worker, where when based on an input, this agent
can spin up like 200 different agents to solve that problem on
the go, right? Like you don't need a framework
for this. You can see this is like 90
lines of code. There's no framework, plain
TypeScript or JavaScript or Python.
It's just basic API calls, right?
(17:10):
There's nothing you would need to run, you know, learn here,
except for our API versus, again, any API looks like this,
right? And by doing this, we have done
a lot, right? So we've created so many
examples how to use memory threads, this and that, but
there was still something missing.
It was still very technical. And our goal was like, be the
best way to build an agent, right?
(17:32):
How do you make it so simple without you actually losing
control? I don't want to make it like a
drag and drop green Weaver. You are dragging and dropping
and connecting like something like that.
I have nothing against it. It's just as a programmer, once
you have felt the control that code gives you, you can you can
never move away from it. You know the the dynamic nature
(17:57):
of code and how it helps, right?So internally, for last 4 1/2
years, I have been building a coding agent.
And that was just for me, right?You know, when Greg gave me
access, I started building an agent.
I don't use cursor or any other coding agents because my coding
agent is really good. It's like really trained on my
ideas over the course of last five years.
(18:21):
And that agent started doing something weird last year.
It started solving a road map without any supervision and I
was like, holy shit, what is happening?
What we figured out was because we created so many things in
such a good infrastructure behind the scene, our agent was
able to, our coding agent was able to use our infrastructure
(18:44):
to create more agents. You cannot do that with ChatGPT.
Like, you know, like you can go ask ChatGPT to build you a
perplexity. It will not do that right?
But when we asked our agent, theinternal agent, to build a
perplexity, it started building a very many version of
perplexity. And we're like blown away.
Like, holy shit, this is amazing.
And we started using it for all silly things like, you know, in
(19:05):
a POC, when you were telling us,you know, build this and this
type of agent, we would show youlike, oh, we already have
something like this. And people would say, wow, we
never thought that it could become a product, right?
And that is what we launched as Chai.
So Chai, we put human between a computer and AI.
You've created an auto scalable community space infrastructure
(19:26):
where we can scale to thousands of machines to solve a problem.
Right now Chai is not a general purpose agent.
It is built for just one purpose, to build agents for
you, right? No other agent right now can do
that. I am very well connected.
I have not seen any agent havingagentic capabilities to a point
(19:46):
where it can build agents for you.
And those agents are also agenting, right?
It's like a same loop, right? So this is what we have
launched. You know, I'm a huge Chai
drinker, right? So it's just I was looking for a
word that had AI in it in some funds.
A Chai, you know, prompts a friendship, right?
What it can do is you can just ask it to build an agent.
(20:08):
Like, for example, I don't know,make me an e-mail agent.
Oh, there you go. Yeah, log me out, right.
Yeah, it's a bit of product and we are figuring out a lot here,
but let's let's go with something like.
What we were thinking, I'm gonnaask it to build me an agent and
(20:28):
imagine that I'm somebody who has never built an agent, who
may or may not have programming knowledge.
This thing Chai is going to bothcode that agent and build me the
UI and the API for that agent right so that was the discovery
We were like oh, what if like I don't like I'm nothing against
(20:48):
lovable or visa or anything, butI don't need to build like 100
websites. I know how to build a website
yes with lovable or bold or all of that.
It's fun to build a website, butto a point where, oh, what do I
need a website for anymore? Like maybe after a couple of but
dude, I need hundreds of agents for my stuff, right?
And as somebody who's building acompany, I still never got
(21:11):
enough time to build a lot of agents.
You know, like you have to code it.
There's something to be said about no codes it it tends to be
fast, but like, but then you lose control.
You build something fun and thenyou know, I can't do this.
I want to do a four while loop, right?
But like, how do I pull that up,right?
So let's let's do that, right? So I've asked her to, and by the
way, totally expect something tofail live demo, right?
(21:34):
So I'm just asking you to build me an e-mail agent.
So Chai does only one thing right now.
It builds you the agents, not the connections, nothing else,
right? Right now the focus is just to
build the agents, right? And what I've asked it to do is
build me an agent that will readmy e-mail and will perform
several tasks in a sequence. It will analyze and summarize
e-mail and it will figure out ifthis is an e-mail that is not
(21:58):
spam and if it needs a response and maybe then write the
response as well, right. So now it has started e-mail,
you know, telling me what it is doing e-mail processing agent.
This is what I'm doing. The workflow it is building will
simultaneously analyze. This is important.
I these are two separate tasks. I want them to happen in better.
(22:22):
And then this should be a block gate where, you know, we need to
determine if our response is needed.
Then we use that all of the context to build that agent.
Now, this is really easy to build.
We have a bunch of guides on this as well.
Like I think we call it a composable AI e-mail agent
architecture. Like when you get a spam e-mail,
basically, you know, they basically figure out that this
(22:44):
is a spam e-mail. It doesn't do anything.
If you get a valid e-mail, like you know your API is broken, the
CM maker will create a response quickly and start writing the
response. Like, we've talked about this a
lot in our docs, right? But what Chai is doing here is
different. Chai has gone ahead and built
something for me. I'm gonna go and deploy it.
So while this is deploying, let's review what the code looks
(23:06):
like. So in the code, it has created
an agent. This part is analyzing the
sentiment. I can explain it and prove it,
whatever you want, right? So I can like just say, you
know, this part is not doing what I want to do this, right?
That part is there. So this is the step one, and
we've made sure that our workflow is extremely simple.
There's a step, there's an ID, and then there's a run function
(23:27):
where you just add whatever you want, and it's probably giving a
response. It's already created the UI for
it, by the way. It's just deployed it, right?
But let's go back to the code, analyse sentiment and it has
created a schema. This is more advanced than I
thought it would like, you know,sentiment or whatnot.
Like as a programmer, I would not probably do this in one go.
I would be like, you know, I'll get to that point someday.
(23:50):
And then it is feeding that too.It has decided to use GPD 4.1
Mini. So we have trained a lot of
things on different models and we are continuing to do that.
It can basically understand whatthe task is at hand and what
would be the right model for that.
And you can optimize that. You can ask, you know, make the
responses much better. So use a premium model or maybe
use something that I could use for free like from Gemini or
(24:11):
what not. And it has also has the ability
to create prompts. You can obviously improve these
prompts, right? But this was the first agent,
this is the second one. Summarize the e-mail.
This is the step where it is determining if the response is
new year. It has also created a schema for
that. If the response is needed, why
(24:33):
is it needed? The reason, the priority, right?
It's automatically doing this. You can completely change this.
You can just say, you know, add more labels here over now and
then I think it is just simply generating a response.
If the response is needed, thereyou go, right?
The Step 3. And by the way, you don't need
to look at this this way. We also give you a flow which
(24:55):
you can just see and this is what it's doing and it's gonna
process the e-mail sentiment, somebody determine if it is
needed, if it is needed this notskip log it and be done right
and you can regenerate it or what not soon.
We'll also give you the traces. You know, you can see how this
is running and if there's a memory required, the memory is
also created for this use case. Memory is not required.
(25:17):
So well, here's what the agent UI looks like.
Let's give it something like, I don't know what was the e-mail
here. I can just probably write
something. Your billing API is broken.
This is not good, right? So now it is processing all of
that for flow, right? It is figuring out if we need a
(25:39):
response or what not. And this is just a silly UI,
right? But I think what we have sort of
ended up inventing is nobody wasbuilding UI for agents,
especially not dynamic US. Like what would a e-mail agents
UI look like, right? So here's the summary, Here's
the sentiment analysis. Sentiment is negative.
You know, urgency is high. And it has also gone ahead, told
(26:03):
me the reason why it should be responding.
And it has also drafted a response, right?
And this is pretty powerful, right?
I just created this thing, right?
And I don't have to use this UI,I can just use it based on this
API, right? The Node Python call, I can just
send in my e-mail from whatever to and it will do exactly this,
(26:26):
right? So it is an actual agentic
workflow where multiple agents are doing multiple things,
right? And for a lot of agents, what we
have figured out is a simple UI is all you need, right?
So I don't know, you won a dollar 100 million lottery.
Sometimes it doesn't do the spamcheck really well and I have to
tell it like you're letting the spam pass through.
(26:50):
Let's see. OK, Summary emergency is higher.
It's pretty positive. Yeah.
And there you go, not respondingto this one, right.
I guess it's probably a spam, right?
Yeah, right. So sometime it doesn't do it.
It starts responding. And then I have to like come and
tell it like, you know, I don't want you to do this, right?
(27:11):
And then there's so many that I've created.
I think you have a question. Yeah, sorry, just to jump in.
So this super impressive click of a button or basic prompt can
generate all of this, I assume with the deploy button, what
not. This can all be hosted on Lang
base. So I can just kind of go in,
create a little army of agents and just by giving my prompt.
This has all the infrastructure needs, the ability to expose the
APIs, and then I can just kind of operate as I would anything
(27:33):
else that's just cloud host and serverless.
Yeah, that's that's the point, right.
So you don't have to worry abouthow to scale this.
You can send it like 100 millionemails a day and it will just
process it very easily. We are generally doing a hundred
150 million agent runs a month at this stage, right, So all of
this is deployed. Whatever we needed to work out
(27:54):
for this deployment happened behind the scene and also this
UI is also deployed. I know it's beta.
You don't see a URL right now, but we'll the first thing we are
launching this week is you should be able to share this
agent with anyone, right? Like you know what so source
graph actually uses Lang base todo this.
You know, AI bought in their docs, right?
(28:18):
So somebody in our docs team where they had an idea was, you
know, it's because it's always me saying something like this.
I don't know, like, let's figureout some feature pages.
Our feature pages are not great.So sorry everybody in our docs
team, right? So versioning, OK, versioning is
good, right? So fork, let's say I don't like
this page. And what I'm gonna do is I'm
(28:39):
gonna copy this link and I'm gonna go to Slack and I'm gonna
say maybe this page should have more content, right?
So what our docs team has done now is let's go over here
somewhere, right? So what they have done is the
moment I post something in Slackthat, well, this URL, this page
(28:59):
is not great. It should have like more content
here or what not. They think that this is a
prompt. They throw this to Chai
automatically. They have built a small dog
spot. Chai actually builds all the
agents needed to fix this because Sai has computers in it.
(29:21):
It is connected to GitHub, it can download data from GitHub.
It can figure out where the filefor this link is.
Our docs are on GitHub, right? And then it runs those agents
and that feedback to fix that problem.
And then it finally creates a PR.
Now this workflow is not new. You get a PR and you have an
(29:41):
improved doc. This workflow is not new.
You could do this with anything.What's new is the next time I do
this, they will not send, they will not tag Chai.
They will have a very defined particular workflow that will
work the way they want, right? It's like, you know when you do
one thing with Cursor, it's, there's no guarantee that the
same thing will happen next time.
(30:02):
If you add an API endpoint and you're really happy with it,
Cursor has not learned anything.You just have prompted a lot.
And when you're building something with Chai, you're
actually building real life agents that are there that you
can continue to improve. Like for example, I've built a
very mini version of API Route Creator.
We have a very different way of creating API routes internally.
(30:23):
I couldn't teach that to cursor.Maybe I could add cursor rules
or whatnot, which are just prompts.
Agents are agentic, right? So now I have a API Route
Creator agent through Chai that works locally.
By the way, it's not deployed anywhere.
You can ask, try to create something, you can take this
code and just use it locally, right?
So that is still true. You can use it with Olama.
(30:45):
We integrate with Olama, right? Maybe you don't like this UI or
you know you have some update inthe UI.
You can go here, ask some changes for UI, right?
One more thing that I really like is like I've been building
so many like I've not built thatmany agents myself, but right
nowadays I've been like, I don'tknow, like here's here's a
perplexity one, right? So a little bit perplexity.
(31:06):
I think I did one shot where I asked I make a deep researcher
like perplexity. And it basically tried and I
deployed it, but then when I ranit, I got this error, some
schema or something. I didn't even read it.
It's wipe coding anyway, right? So I ask it like, fix this
error, right? And in the second shot, I think
(31:28):
it got it right. So this is what we have now.
This is our deep researcher. I don't know, like let's what
did Open AI launch last month, which was April?
I don't know if they launched anything.
I don't know what it's gonna do.It's using EXA behind the scene
(31:49):
excise, you know, Internet scraping or search tool of
sorts. There you go in April 2024.
Oh, there you go 2024. Like it's obviously it's not
publicity, but it's working, right?
And like you can basically ask it like, I don't know what has,
(32:10):
I don't know if this will work on X lately, right.
So this is, I don't know if Twitter allows you to scrap
data, but I'm pretty sure their API is pretty down but exists
pretty good. It should probably bring
something up. It's waiting watch, it's taking
it's free time. There you go.
(32:30):
OK, some of my tweets by script.It's bringing up old data.
I think what's happening here isEXA does not have access to the
live X data and it's just when Twitter was open and they
allowed it to scrap. That is what they're bringing
up. But it's pretty good
architecture like you can see how it is happening, how and
what it is doing and people havebeen building so many fun stuff
(32:51):
like for example, a simple thingis as simple as like I don't
know, chat with PDF, right. So it should probably just build
a memory agent now where I should be able to upload APDF or
any file and talk to it, right? So this common problem, right?
While it is doing that, I can also probably show you a little
(33:13):
bit. I'm probably gonna zoom in a
little bit. So we don't see their names.
So there are usernames here. And by the way, in five to six
days, 35,000 agents have been built with Chai.
Wow. Let the sink in right?
People are building amazing stuff like this.
Is today's May 6th you look at the time stamps here it's almost
(33:35):
every other minute there's like an agent being built by somebody
right so PDF chat, Firefly summarizer, e-mail responder,
mail sorter, some of the. Most compelling agents that
you've seen developed, whether with Chai or otherwise?
I don't know, it's really up to you.
Anything. People are building so many
things. Like I saw somebody from
(33:57):
Change.org create a partition agent which scraps your
Change.org partition and helps you both write the content and
talk to people in comments. I'm like, holy shit, right here
you go Forex trader, like all sorts of fun stuff.
My biggest really, which I was that people are were gonna go
and use it like ChatGPT and thatwould mess it up, right?
(34:20):
Like, for example, I know if yougo to ChatGPT and you say make
me our deep researcher like perplexity, it's probably and
let's go over there. I don't know what it's going to
do. It's definitely not going to
build me an agent, right? So I.
(34:41):
Thought the mindset. I don't know what it is doing
right. So I thought people will start
using Chai as Chai GPD or cloud or something and that would
totally not be cool. But but unbelievably, like 95%
of people are like, you know, look at this like NAV centric
and this is just somebody going,you know, nutrition planner
(35:02):
mind. These people are building all
sorts of fun stuff. What they are doing is like task
refining. What it takes to build an agent
is automated now, right you I have, I've been asking people to
not use frameworks because they slow you down.
Use primitives. Now it doesn't even matter
because something else is writing code for you.
You have the complete control. This is not DSL as you can see.
(35:26):
Let me deploy it basically went ahead like this is not DSL.
This is not some weird weird configuration file that is super
vendor locked into LAN base. It is an actual code written for
you for the purpose of your agent that you are building,
right? It's, you know, like how would
you build a generate answer agent?
Like it has also created a really good, you know, prompt
(35:49):
for that purpose as well. So I've been having so much fun
just reading through what peopleare doing.
NER data scout agent. I know what that is.
Race engineer fit plan or whatnot.
Lightning code expert. A lot of people are actually
building coding agents. A lot of people have ideas of
never does this or love people never does this does that right?
(36:10):
They don't feel like they have the control over improving the
agent they are using. With Chai, you are actually
building the agent that does what you want and you have the
control of improving that agent over time.
As your project grows, your requirement requirements grow
over. So the people are building all
sorts of fun stuff here. Like I just like there are 54
pages of these. These are just workflows, right?
(36:31):
Yeah, actually a. Quick question on that, Yeah,
when someone makes their own coding agent, because there's so
many options out there and they want to make it best for them,
what would you say are some gooddeveloper experience principles
that people should follow or just be aware of so that the
tooling that they build is goingto be beneficial to them rather
than just something that's flashy?
I don't know if do you know about Tailwind.
(36:51):
Slightly familiar, but do you want to tell the audience?
Yeah, so Tailwind is this framework that was created by
two amazing designers and developers Adam and Steve in
for. For someone as technical as me
who always wrote CSS from scratch, I had like this weird
think about frameworks. Like I can't use CSS frameworks.
I'm a Bootstrap contributor, butI don't use Bootstrap.
(37:15):
It's just very, very weird when talking about it that way,
right? And for the longest of the times
I was like, I don't like Tailwind.
And then it hit me one day that Tailwind is not a framework,
it's actually an API, a primitive of sorts, which allows
you a defined way of building CSS.
And they have really pretty and really awesome, amazing docs
(37:36):
that anybody can read because everybody who's building, I was
writing CSS from scratch of their own ways.
I had my own way. But you on my team, if I hired
10 more people, they didn't knowmy way and took me time to teach
them. So Tailwind built really good
documentation, really good primitive that anybody could
learn. And I could just hire somebody
who knew Tailwind a little bit. They would create anything
(37:58):
because it's not a framework, it's just an API and you can do
anything with API, right? And there was this curve I had
where I was like I went from, you know, like I think like, I
don't know, there was a like a no to tailwind and oh, this,
this is pretty cool, right? Like my emotion emotion went
like this, right? This is the same thing that has
(38:18):
happened with vibe coding, right?
So vibe coding can both be helpful or very fleshy and
wrong, right? So what I see people do is what
they do is which is by the way, what I'm doing here right?
So chat with PDO. This is the wrong way to wipe
code by the way, it's fun like for example, it has gone ahead
created this PDF document memory.
(38:40):
I'm gonna for the sake of this chat, I'm gonna upload something
here. Let's upload this has some data
on my this is the home page of my personal website, right?
And I'm probably going to throw in another text file as well,
which has which is the API keys portion of our docs, right?
(39:02):
Just to ask these questions and it's probably going to go ahead
and process this in a minute while we chat, right?
There you go. These are ready.
Let's go to the UI and let's askme what is the name of Emmett's
company and how do we get API keys from that?
(39:26):
Fingers crossed. I hope this works right.
So I just added two documents, one that knows who am I.
So it should probably figure outthat the new other companies
land is there you go and you basically go to the settings
menu and there's this and there you go.
I just asked chat with PDF and this is what I got.
I can use it through this UI. So you can use this as an API or
(39:50):
UI. It really is up to you, right?
So what a lot of people do is they get this weird wrong idea
about vibe coding. So I think vibe coding.
So I'll take a step back. I think Vibe coding is going to
replace no code and no code. Right, like we always had like
no code, low code website builders.
(40:10):
We had green Weaver, I don't know, like 1990 or something.
You could drag and drop and build a website.
But even then lovable and bold exist and they give you code.
You don't need to use that code.But if you want to or if you
have someone who can help when you need it can help, right?
Like if you throw, you know, your no code to tool to
(40:33):
developer and they were like, Oh, I can't help with that.
They'll be like, you know, I'm going to build you something.
I'll charge you like Python bugsfor that or not.
Or you can just ask like, this code is doing this.
Can you just do this small implementation that I'm not able
to figure out? That's super powerful, right?
So for the longest of the time Iwas like, ah, white coding is
weird to a point where I figuredout, oh, white coding is
actually replacing no code and no code with actual code that
(40:57):
you still have to not put in a lot of effort to build, right?
That is super powerful, right? So I got over that hill where
I'm like, OK, vibe coding is great, right?
And this is coming from someone who's deeply technical, right?
But what people get wrong is, let's say you want to get to, I
don't know, 10, right? And what people try to do is
(41:19):
they try to do 5 + 5 or 6 + 4 equal to 10 with prompt
engineering or vibe coding. What actually generally works is
1 + 2 + 3 plus, I don't know whatever the rest is.
To get to that result right, Youhave to take it step by step.
It's similar to let's say you hire an employee or let's say
(41:41):
you're looking for a Co founder.You don't just throw all of the
context at them. You don't just say, you know, be
my Co founder and they'll be your Co founder or be my
employee and do all of the things that I expect.
You have to slowly build that relationship, slowly build the
understanding of the feedback loop of loop of you know where
(42:02):
now build this now do this now do that right?
Like a human in the loop of sorts, right?
That generally works really well.
So what I often tell people is start with something simple,
very simple. So you know, it works.
You get familiar with. So Agent DSS, the agent, agent
TSXS, the UI, which I can host anywhere.
(42:23):
It's a simple React code, right?It's conveniently hosted for
you. But if you wanted to take this
code and do whatever the fun thing you want to do with it,
yeah, be my guest, right? Then you slowly build this thing
up into something. I don't want to name names
because I don't have permissionsfrom them, but there's this
company that in the last five days, I think they have created
hundreds of agents, which I in there this the CEO is doing
(42:48):
this. They're pushing the limits with
China. And I'm like, I'm mostly looking
at their chairs. I'm like, oh, oh, it can do
that. I didn't know, right?
Like I built this thing. But they're like, you know, let
people upload their logos and from their logos.
It's a swag company, by the way,like their very common name.
Everybody uses them to build their T-shirts or stuff like
(43:10):
that. When you upload a logo, they are
able to figure out through theirinventory using an agent that
they are continuously building, which I that what can they
deliver? What colors by when and what is
like possible. This is like an extremely
complicated job. And I think that he's like, you
(43:30):
know, probably 3050 messages in one thread.
So now we are thinking about like, oh, this is too much.
We've just figured out like you cannot have longer as long chats
and shared GPT or Claude. But we we create a child for
these agents not to be formal. They are always going to be
there and you'll always try to improve them, but you cannot
(43:51):
have 10,000 messages here. So now we have this user
experience issue where we're trying to figure out, OK, maybe
we create some sort of done withthe last 10 versions, stash them
away. Let's summarize the context and
be here. We're calling it something weird
internally where we are very excited about that feature,
right? So anybody can use it.
(44:13):
It is just how you use it that matters.
You could be less technical and you could have much better
identic ideas than I do, which is what is happening.
I, I go look at the space and I'm like, oh, style guide,
etcetera. People are building things that
I didn't know we could build, you know, with this, OK, I think
I am on page one, which we just saw, but I should probably be
(44:37):
zooming in. People are building all sorts of
fun stuff here that I never likefought from Spring agent right.
So I don't know Swift UI could build a French wedding maestro,
right? Like anybody can build anything,
right? It's, it's the fact that it gets
(44:58):
you to, in most of the cases, 9095% done.
And then you have still the complete control over chatting
and fixing its behavior, improving it or starting fresh
or bringing in an actual expert,like, you know, somebody who's a
developer, who's a developer on your team.
Let's say you're a PM and you'rebuilding something.
(45:19):
It's fun, it's working. You have bookmarked it in Slack.
And now there's a bug that you just cannot figure out, right?
Maybe you used to code like 10 years ago and now you're not
brushed up on your React skill. You just tease your front end
guy a little bit or girl a little bit and they can just
say, oh, just fix this and be done right?
(45:40):
That they cannot do with a no core tool, right, right.
So Yep, what do you think? I think it's incredible.
I did have one other question for the For the last remaining
5%, if someone were to develop an agent gets pushed away there,
pull it into local Idi. Remember you mentioned that you
don't use cursor your VS Code because you have your own coding
(46:03):
agent, but what advice or tools have you seen in the AI age for
development that you think are useful?
Does MCP help with your dev experience, or are there any
other tools that you think people should start exploring as
they're trying to build to get these last few percentage points
over the line? Anything really that helps, you
know, back in the day. So I'm here today because I
(46:23):
started writing content and I started blogging when I was very
young, I think late 90s and all of my careers because I started
writing one day. I remember watching Cristiano
Ronaldo 2001 FIFA, I think Euro World Cup and I'm like, this is
awesome. I should write about it, right?
So you should just use tools, right?
(46:45):
No matter what, you don't know where what one tool would lead.
You're taught to. I love reading source code.
Like you can often find me like at 1:00 AM in the morning
reading the core code of Expressjust because, right?
I love to just read source code as in read as in just look at
it. I think my brain picks up on
weird things that later on and I'm like, I don't know how I
(47:08):
know this, but I know this, right?
So MCPS like this, right? MCP, every tool has pros and
cons, just having the knowledge,I think.
So here's my advice about that. When I was young, when I went to
college, I thought if I knew a lot of programming language I
would languages I would do really well.
(47:28):
So I started like, I don't know 26 different programming
languages. And for the longest of time that
turned out to be the wrong choice.
I ended up just getting really good at just a couple of
languages like JavaScript, TypeScript, Node, C, Python,
right? But now I really enjoy the fact
that I know more than 20 programming languages.
(47:49):
I know there is a solution in the lexler that can fix this
concurrency issue struggling with I know how to think about
it and if I give the right and wrong, these agents can do fun
stuff, right? So my advice from over the years
has now changed. I'm like read everything.
Potentially you'll have this is agents are just a new type of
(48:11):
code. They are going to be very
helpful and tools are going to be your augmentation.
I think humans have this capability.
So legs and arms are our tools. My father is a doctor.
I asked him how does our brain learn things?
And he's like, so legs and your hands are tools, but your brain
has the capability to create newtools.
(48:32):
Like you can learn to drive a car using these four tools,
right? And that capability is what they
call a process that he has this entire diagram of C1C2C3
cognition levels where you get effects, you understand effects
on C2, then you basically understand what to do or what
not to do, like reasoning. And 4th 1 is planning.
(48:53):
Like even a dog can reason like if it should, you know, sit on
your bed or not, right? But they cannot plan like
humans, right? I think we are going to have
planning tools. LLMS are going to get better.
And you, the human is going to be the best planner there is.
You will be augmented in the middle, right?
Like computer, human AI, you're the planner here.
(49:15):
You need to know all those toolsto understand the knowledge.
You may or may not use MCP, right?
But you may like something from MCP that you can ask an agent to
build for you, right? Or you can just use one MCP to
do that. One thing.
MCPS are cool. One thing I don't like about
MCPS are do you know what is a CLI like?
(49:38):
You know you have a command linetool and it does exactly what
you want. I think MCPS are poor man's
Clis. People who don't know how to
build a CLI can use MCP to potentially automate something.
The only problem I have with MCPis that it just consumes a lot
of tokens, right? Just to do a simple thing.
(50:01):
Like I was talking to Claude theother day about my Stripe
account. He did one analysis and he was
like, you're done for the day. Come back 5:00 AM.
I'm like, what? Right.
So I just only asked, like, you know, how many pro subscriptions
of Chai happened in April, in May, and it just consumed my
entire day's worth of tokens. That was unexpected, right?
(50:21):
I could have just built a CLI, one shot at it and done that,
right? So each tool, knowledge of those
tools will actually help you understand if that is something
you need or not need. And someday that knowledge will
actually really be helpful. Now you have all heard and learn
something because it will be helpful someday.
Now it can actually be helpful. Any knowledge of any technology
(50:44):
can actually be helpful because you don't actually need to learn
everything. You don't need to learn React
view and everything. But if you just know React, you
can kind of build something withview because these agents and
agentic systems and LLMS are getting really, really good at
it, right? So yeah, MCPS, well, definitely
try them out. So as someone who's super
(51:04):
interested in space and loves Team Open Source, could you tell
me a story about how you got your code running on another
planet? So that's a fun story.
When I was young, I wanted to bean astronaut and an astronaut
who was absolutely afraid of heights.
So I kind of settled for being AI, thought I'll be a scientist,
(51:28):
right? I'll invent things that would go
in space. You know how kids are, right?
And a lot of people say that, you know, two of the hardest
thing in computer science are naming things and cache and
validation. And I really like naming things
because I wanted to be a scientist, right?
So like, I wouldn't rinse stuff,right?
I think all of that led to small, little, pretty small
(51:52):
little contribution in the NASA Infinity Mars helicopter
mission. In fun fact, it was actually AUI
code that they are running in the Ingenuity helicopter
mission. Like, you know, who knew being a
web developer would actually, you know, you can, you can throw
your code to Mars, right? So, so yeah, and that happened.
(52:12):
I was like super excited when they told me like, you know,
this is the project where this code is going to be.
And then I later on and got a little bit more into it, you
know, got to see the project a little bit more from, you know,
like you know what that actuallycould have been even in an
alternate universe being espirant and also got down to
(52:32):
side down with Satya Nadella andtalk about it like, you know,
like what it felt like. Like I think about 10,000 or so
developers whose open source board actually made the Internet
E helicopter mission possible. And for me, open source was how
it happened, right? It's not like, you know, I got a
job at NASA and did that, but you know, I eventually got to do
(52:55):
the same thing. So it was super fun, you know,
So we're really excited that that happened.
I know my mother and father werelike, oh, so you do do important
stuff. I'm like, Oh yeah, yeah, yeah.
So, yeah. Fun story.
Love it. Yep, encourage ever to export
open source. You never know what happens.
Exactly. Ahmad, this was a blast.
(53:17):
Thank you so much for sharing your perspective.
Chai looks awesome. Courage everyone to try it
before I let you go. Is there anything you want the
audience to know? No, that was awesome.
Thanks a lot for having me read the source code.
Use your code for good. Absolutely.
All right. Have a great day.
Popular Podcasts
On Purpose with Jay Shetty
I’m Jay Shetty host of On Purpose the worlds #1 Mental Health podcast and I’m so grateful you found us. I started this podcast 5 years ago to invite you into conversations and workshops that are designed to help make you happier, healthier and more healed. I believe that when you (yes you) feel seen, heard and understood you’re able to deal with relationship struggles, work challenges and life’s ups and downs with more ease and grace. I interview experts, celebrities, thought leaders and athletes so that we can grow our mindset, build better habits and uncover a side of them we’ve never seen before. New episodes every Monday and Friday. Your support means the world to me and I don’t take it for granted — click the follow button and leave a review to help us spread the love with On Purpose. I can’t wait for you to listen to your first or 500th episode!
NFL Daily with Gregg Rosenthal
Gregg Rosenthal and a rotating crew of elite NFL Media co-hosts, including Patrick Claybon, Colleen Wolfe, Steve Wyche, Nick Shook and Jourdan Rodrigue of The Athletic get you caught up daily on all the NFL news and analysis you need to be smarter and funnier than your friends.
The Joe Rogan Experience
The official podcast of comedian Joe Rogan.