Pious technical makeup is actually very

trivial and the core of an nested loop.

Uh you send an LLM request. The LLM

might respond with a final answer or

with tool calls, which means execute

bash, [music] read that file, whatever.

Then we execute those tools, give the

results back. LLM does this for as long

as it thinks it needs to and then

eventually spits out a final answer

without tool requests. That's the basic

[music] gist. There's nothing to it and

it's hilarious the amount of money I'm

being offered for something like this

compared to other deeds I did before,

which were way more technical and way

more deep tech.

I'm Mario and I'm the creator of Pi, the

coding agent.

The reason I built Pi is that I like

being in control of my tools. With Cloud

Code, which was the first coding agent I

used, I initially felt that its

minimalism was right where I wanted to

be. And over time the team started

adding a lot of new features with the

sheer velocity and the token maxing they

applied to Cloud Code,

there were also a lot of changes behind

the scenes that made the agent behave

differently for the same deterministic

workflow and I don't like that. I

already despise the fact that I need to

rely on a third-party provider for

tokens. It's just not the way I want to

work. I just can't bear the thought of

having a million lines of code that do

something that can be done in 5,000

lines of code because I think complexity

is the mind killer. So I was like, "Eh,

how hard can it be building my own

coding agent?"

I hail from a little town called

Deutschfeistritz in the woods

forests of Styria, a little part of

Austria and my beginnings with computers

were basically me loving computer games.

So I

went to my uncle, which had an Amiga 500

and later Amiga 1200 that let me play

games. And eventually I told my parents

I want to have a computer and they told

me, "Sure, if you work." So, we did and

I got my first 486 40 MHz with a turbo

button. This was very amazing. And

obviously I didn't earn that money. It

[music] was my dad mostly doing tiny

little illegal jobs after work. I'll

forever be grateful because that kind of

started [music] this whole insanity and

has become a big part of my life. And

thankfully computers back then came with

a programming environment in MS-DOS

called QBasic. So, that's where I

started out and it came with a lot of

example programs, two games, Nibble and

Gorillas. And once you have the

internet, you can learn anything.

The first open source project I got

involved with that was meaningful, I

guess, was called Free Basic. It was

actually started by Victor, the

preceding guy who who used to mentor me.

After being part of an existing open

source project, I thought maybe it would

be nice to start one myself and around

'29, '2010, Android had a little

burnout, so a lot of time, and I started

working on a game development framework

called libGDX because I'm very good at

names.

>> I'm Mario, benevolent dictator of

libGDX. Let's cut right to the chase.

>> Niantic Labs used it for Ingress, one of

the first location-based multiplayer

games on mobile or even desktop games

like Slay the Spire would be based on

libGDX. And lots and lots of tiny indie

games, some of which actually turned

into sustainable businesses for the

people who did them. And

>> [music]

>> the proudest moments of myself was going

to conferences around the world and

having people randomly come up to me and

say, "Man, this is amazing. Thanks for

building this because it now sustains me

and my family." That was cool. Like,

[music] that's how open source should

work, I guess. Be a multiplier.

The one thing [music] that libGDX was

lacking for the longest time was support

for iOS and it being geared at mobile

game development, that was kind of a

bummer, obviously. And that [music]

developed into a business relationship

eventually. We managed with a very small

team to kind of have feature parity with

Xamarin within a year and a half or so,

which made us an ideal acquisition

target for Xamarin,

who eventually acquired us and they were

immediately swallowed by Microsoft,

>> [music]

>> who then turned around and said, "We

cannot touch Oracle technology because

of licensing issues." So, they shut us

down and asked us to work on Xamarin

instead.

Such is life.

>> RoboVM is dead.

>> After the whole RoboVM and close

sourcing of an open source project that

I championed [music] as it's always

going to be open source, I kind of

soured on open source.

>> [music]

>> At least within

commercial environments. And I honestly

didn't have the energy anymore to

maintain any kind of community [music]

because that takes a lot of you.

>> [music]

>> My name is Armin Ronacher and I'm

co-founder and partner at Airndell.

>> Armin I met on the [music] internet 15

or 16 years ago. He goes around saying

he met me earlier at some game

development conference here in Graz

where he was still a [music] student. I

was very famous and ignored him or

something. I don't remember it. Sorry,

Armin.

>> But I don't think he sort of spend any

time thinking

about who about who he was interacting

with. But then we mostly interacted

through the internet. A combination of

Reddit and Twitter, I think. But Armin

is very much like trying to figure out

like how stuff works and then make that

work really well. Like he cares that

stuff works really well and then he also

cares that what he can build with it

is like morally good.

>> Initially I tried to hide Pi, the coding

agent, from the world by calling the

website shittycodingagent.ai,

which immediately poisoned it for any

enterprise deployments. I remember

somebody from GitHub telling me, "Can

you please rename it? We cannot like

[music] we cannot internally get

something through that has the website

shittycodingagent.ai." And I also called

it Pi because that's such a a name that

nobody [music] can Google. So, that

chances of anybody finding this thing

are very low. But then I did make the

mistake of

>> [music]

>> blogging about it and how I built it and

that it's actually not that hard to

build a thing like that. It's built on

the Node.js TypeScript stack. It's

basically four packages in AI package,

which is an abstraction across all of

the providers and models that are out

there with context handoff between

providers and normalizing all the less

than ideal responses that you can get

from a provider. Then there is a twee

package, German user interface package.

Then there is the Agentic Core package,

which basically implements this nested

while loop. And finally, there's the

coding agent itself, which just takes

those three packages and combines them.

So, the one inside I had early on is

that the coding agents can code. Why

can't the coding agent modify itself?

>> I really got an increasing feeling

throughout the last year that the future

of software, at least in part, will be

software extending itself. And with Pi,

it is exactly what I want. It's like I

say it needs to do this and it builds

its own extension or it builds its own

skill. And so, it was a very very quick

proof point of this magical experience

that a self-modifying agent does. So,

[music] I fully switched to Pi around

Christmas 25. Instead of actually

putting all this time into building out

Levo's, I started building more and more

stuff on top of Pi. And so, it was a

clear sign [music] for me like I have

developed an addiction not just to

agents but to Pi in particular.

>> And then, Peter decided to make it

[music] the Agentic Core of OpenClaw.

And then, yeah, everybody started using

Pi either directly or through OpenClaw.

>> Then I wrote a blog post in early

January sort of on the

when when [music] OpenClaw was was

really popular. It's like, "Hey, just do

away with the claw and just use the

underlying Pi instead."

>> [music]

>> Shopify seems to be running 50% of the

Agentic coding workloads on Pi now or on

OpenRouter, Pi is placing fourth now um,

front of CloudCode, beating out a lot of

other open source agents. So, that's

when I started understanding that this

is kind of serious, because most of

these issues and pull requests didn't

come from humans. They came from agents.

>> People in San Francisco are currently

token maxing, meaning use as many agents

and tokens as possible. That is

the direct opposite of what me and Armin

are doing, because what we're seeing out

of Silicon Valley is basically

extreme velocity, unsustainable

velocity. And I think a little bit of

pushback. So, if there's one thing to

take [music] away for best practices

with agent coding, don't leave your

brain at the door. While you can have

the agent write the code for you, it's

on you to actually read the code,

understand the code, and understand the

consequences of the design of that code,

because you will ultimately own the

responsibility for that code.

>> One of the reason we spent so much time

on agent coding tools is because they're

addictive.

And

if it just becomes performative art what

you do with them, and you actually don't

really develop anything of value, then

that's a bad future. And so, if

you can build something and you can

share it to another person within like

15 minutes, but you already shared it

before you even tried it yourself, like

there nothing of value was created

there. I I want the future to happen

where people that are really good at

their job and have enjoyment in it find

their job more enjoyable because the

boring part about it can be in parts be

automated with AI.

>> I realized that, first of all, Pi

apparently has commercial value in the

market. I'd like to build a team around

Pi that [music] actually makes it into a

sustainable thing.

>> It should be possible to build a

commercial company and also still value

this idea of like open protocols and and

and open source software [music] and and

finding ways to reclaim a little bit of

the space that was

otherwise held by large corporations

where like the main interest is just to

grow one closed ecosystem. The way we're

evaluating our our vision now is like

trying to see that we can play [music]

an an important role in keeping these

tools available to everybody. Whatever

we can do to to sort of keep the

foundation open, I really want to do

something for this. Pi fits very well in

the picture because it's an open source

product, but also because it is

it is the most minimal harness that

doesn't actually have a ton of opinions

about any model in particular. There's

no lock-in in [music] Pi. What a large

model can get away with inefficiencies

um because of just its massive scale,

um a local model cannot get away with.

Um and all of a sudden like the sort of

traditional approaches of software

engineering still matter. Bugs actually

need to be fixed and um [music]

behaviors actually need to be

investigated. Pi is just a very very

good starting point to to do all these

kinds of experiments.

>> Ultimately, what we want is

>> [music]

>> intelligence for everyone,

maybe even locally

uh with neutral [music] platform-neutral

tools like Pi or Open Code or whatever

you have,

allowing you to access that intelligence

[music] and make it useful. So, yeah,

it's that's what's up for that future.

>> The total number of humans did not

increase just because [music] we got

agents. Like at the end of the day,

really we're we're just here to service

humans.

>> I don't have [music] any grand news

legacy um visions myself.

I just like to build things, [music] see

what sticks, and contribute um

to the way a lot of people use [music]

new technology.

That's basically my only wish.

>> [music]