Daniel Miessler - Anatomy of an Agentic Personal AI Infrastructure | [un]prompted 2026

All right. Thanks to uh Gotti and everyone for having me here today.

So uh I wanted to uh do a few different things with this talk. I started off wanting to talk about my personal AI stack and uh I

started thinking everyone has one because I submitted this uh months ago, right? And I'm like, well, everyone has

right? And I'm like, well, everyone has an AI stack, so I probably just want to talk about some ideas instead.

And uh, Gotti ended up saying, hey, look, um, you know, it is the name of the talk. Go ahead and do that instead.

the talk. Go ahead and do that instead.

So, I think I'm going to do a combination of the two. Um, this is the first one I want to talk about, which is companies become APIs.

So, Excaladraw, they released a thing.

They're like, "Hey, you can now just talk to your Excaladraw and it will produce the entire diagram for you." And

I was like, "Oh, that's pretty cool."

And they're and I'm like, "Okay, what are the instructions?" And I was looking for the agentic part. And they came back and they said, "It's easy. You just log

into Excalibur and you go into the tool and you type what you want to happen and the diagram will show up." And as of like two months ago, that is a

non-starter for me. I I can't like I'm not going to go to any physical tool and actually use it. So yeah, that that is pretty much a non-starter for me. I I

think the way this is heading is your company exists as an API and if people's AIs can't use your company in that way, you kind of don't exist. And I think the

way that is heading is towards essentially kind of like um IMDb or Rotten Tomatoes where you have ratings of software and they're all rated across

lots of different dimensions and your AI will essentially go out and figure out um what is the best service for this particular use case when you ask for

something. And and that's that's

something. And and that's that's essentially where I think this is going.

So that kind of relates to this one which is I'm sure everyone is seeing this uh custom everything. So

a big part of the previous idea is basically agents are doing more and more for us instead of us interacting with the world. We're interacting with our

the world. We're interacting with our agents. Our agents are interacting with

agents. Our agents are interacting with the APIs.

So they'll filter our news. They'll

create custom software for us. And

they'll do all all sorts of this stuff on our behalf.

And uh a big part of that is, you know, is it starts to separate what we see, right? And not only different software,

right? And not only different software, which is kind of crazy. You go into an environment to do a security test, uh you might actually find uh completely

custom software. So you you might have

custom software. So you you might have to spin up a completely custom stack to do the assessment. Um, the other way this is kind of strange is when you

think about that if everyone has their own AI filtering their interface to the world, they're actually getting different news and different ideas.

They're actually filtering reality into them and everyone's agent is doing that for their person. So, it starts to get into a situation where it's like everyone's experiencing a different

reality. Like, so you might go to your

reality. Like, so you might go to your next door neighbor who literally lives next door and be like, "Hey, did you see this new K-pop video? It's the best thing ever. I've watched it 95 times.

thing ever. I've watched it 95 times.

I've cried every single time I watched it." And what do you think? Did Did you

it." And what do you think? Did Did you see it? And they're like, "You still

see it? And they're like, "You still believe in Korea?

That's that's not real. It's it's like the moon. It's not real." So, I think we

the moon. It's not real." So, I think we could just be so separated into our own little worlds, which which will be really weird.

This next one is uh companies becoming a graph of algorithms. So this comes from my experience doing a whole lot of consulting and being inside of a lot of companies over a couple

decades. And the thing that I've kind of

decades. And the thing that I've kind of realized is most companies are just kind of winging it. There are many companies who aren't right like big banks, big

energy, military type places. They tend

to have a lot of processes but in general it's usually humans doing all these pieces.

and the the recommendation, the documents, the processes, they're more of like recommendations than actual rules.

And I think that's going to switch into something like this because I think this is what CEOs and CFOs are actually going to want. They're going to want to see

to want. They're going to want to see their company like this and AI definitely wants to see your company like this. Otherwise, it can't

like this. Otherwise, it can't understand it to be able to change it.

So, I I think this is a a big push that's going to come from the top down.

So, this next one is kind of like the last one, but actually bigger and crazier. And it's the idea that

crazier. And it's the idea that companies move from being human to being process. And I I don't mean AI run,

process. And I I don't mean AI run, which which is also true, but I'm talking about specifically about processes. So, I was embedded inside of

processes. So, I was embedded inside of a a large energy company for a very long time and I learned something extremely crazy there. Uh, when something bad

crazy there. Uh, when something bad happens, they don't ask who did it and like they're about to get in trouble.

They only have one question. Did we

follow the process?

And if the process was followed, they're like, "Okay, cool. That's fine. let's

find out what the problem was and let's get a whole bunch of meetings together and we'll make that slight little tweak to the process and now that's the new process and everyone follows it or you

get fired and so that that tweak I think is about to come to everyone right so this will be a matter of AI then being able to see all those SOPs and keep them

updated and one thing I find really interesting about this is actually what it looks for security, right? So security, we've had this

right? So security, we've had this problem of like we've always wanted to move left, right? We've always wanted to move left in the process, but what if

you could just build software correctly the first time using an SOP? Now, of

course, everyone's been saying that forever but I'm reminded that if you go to like a skyscraper site where they're building a skyscraper, there's no like place that you go and you're like, "Hey, who's

responsible for this building?" and not falling down.

There's no such person. There's no such department. It's literally built into

department. It's literally built into every part of the plan and they just follow that plan. Now, we've been trying this with humans. It hasn't worked out

well. But I think the big change here is

well. But I think the big change here is now people aren't really writing code themselves. The AI is already writing

themselves. The AI is already writing all the code, so why can't it be given a solid SOP to write against? So I think that's uh kind of where things are

heading.

All right. So now I want to talk a little bit about the stack that I use.

Um so these are some of the main ideas.

So I I like to have one single unified AI system and anything I do inside of AI I bring into it. If I learn something from some other AI, I bring that in as a

module. And the centerpiece is to have

module. And the centerpiece is to have the human the person at the center. And

it doesn't have to be like a super nerd like us, right? It could be an artist, an entrepreneur, uh a philosopher, whoever it is. And the

goal of this whole system is to kind of amplify and magnify humans. So it's to focus on your capabilities and what can make you stronger.

So the way that materializes is basically putting people at the center as opposed to so mine is cloud codebased but code is not the center of it. It's

more so the magnification of the human and essentially yeah trying to give everyone that advantage here.

All right so I don't think we have audio upgrade skill. So this is the PI upgrade

upgrade skill. So this is the PI upgrade skill that's part of the platform. And

what this does, I'm sure everyone in here has anxiety even just watching this talk. You're thinking, you're thinking,

talk. You're thinking, you're thinking, I should be actually using AI right now and I'm getting behind just by being at this conference. this pi upgrade skill

this conference. this pi upgrade skill what it does is it it looks at every single engineering blog from anthropic also openai uh I think maybe a couple of

others mo mostly anthropic though and uh it basically goes and finds all everything they've said everything they've said about what

should be upgraded um like new releases it looks at all their GitHub um releases it looks at their release notes for the current release um actually has a time stamp

from the last time we checked. So all

previous releases since we looked the last time [snorts] and it consolidates all of that but then it looks at my existing system and all the changes that

we've made and all the goals and skills and everything and then it matches up what it thinks would be best for us to upgrade so I don't have to track all that which is highly uh highly

stressful.

So this is uh some of the sources that it goes after and this is what the output kind of looks like. Comes back

with recommendations on we should change this skill, we should update this skill, we should add agents, whatever.

All right, this one is really cool. This

one is called council.

So this one actually spins up based on the task that you give it. It spins up like between two and 16 custom agents who are experts in that field and they

fight.

They debate aggressively on which direction it should go. And the

main agent watches them fight, watches the debate. And I one of the parameters

the debate. And I one of the parameters is how many rounds you want it to go for. So they debate back and forth.

for. So they debate back and forth.

um basically communicating like no it should go we should go with this test-driven development platform or spec driven development or whatever it is. So

this is what that looks like when they come back. They form their opinions.

come back. They form their opinions.

They make their best argument and then the parent uh agent basically says okay we're going to go this direction or actually currently it makes that recommendation to me and then I choose

which one and that's some more of the output.

Uh this one is first principles.

This one just basically reverse engineers um exactly what the root cause can be instead of like troubleshooting a specific cause. So we won't go into that

specific cause. So we won't go into that one. Iterative depth. This one is it's

one. Iterative depth. This one is it's it's a research uh paper that came out that basically says if you ask the same question from a different perspective

over and over, the AI gets extremely smart. It gets way smarter than it

smart. It gets way smarter than it should have. um just because it's

should have. um just because it's hearing the same question but in a slightly different way. So it's hitting the same sort of idea surface but it's covering it more thoroughly and I've had

really good results with it.

This one is really cool. It also comes from a paper uh by Zang uh and and the crew over there and basically what it

does is it creates wildly more creative results than just uh doing something like changing temperature or something like that. uh what it does is it's um

like that. uh what it does is it's um it's just messing with the probabilities for the output that it chooses and it's uh also quite effective especially when

you're trying to think out of the box.

All right, so this one is the absolute craziest one. This is the probably the

craziest one. This is the probably the most interesting uh piece of the whole pi ecosystem and this is the algorithm and the idea is that I'm trying to

pursue and keep in mind for all the academics here this is highly theoretical and uh it's experimental okay but I found that it works but I

know how hard the problem is and I'm not underestimating it the idea is software is moving so fast with AI Because it has

handholds, it's able to grab onto something which is does it work or does it not work? Well, what are you going to do if you say write me an essay about

mindbody transfer in Arkansas in 2038 or write me a short story? How does it know what a good short story is? So what

the system does is it attempts to build an ideal state for any scenario not just coding but any general scenario. So what

what it's actually building is a set of um criteria that that it can use for handholds and the ideal state criteria are the same

exact ones as the verification criteria and it's produced some absolutely insane results. So this is what it looks like

results. So this is what it looks like when it reverse engineers a request. So

keep in mind it knows everything about me. So I could give a single sentence

me. So I could give a single sentence and it can say well he probably meant this because of all this context and then it breaks apart this is what he definitely didn't want this is what he

definitely did want and it turns those all into discrete testable ideal state criteria and puts them inside the uh cloud code task system. So

it's being maintained by cloud code.

And this is some more of the output.

And this is what it looks like. So when

I run this thing, I could actually see tons of uh engagements. Basically, all the

engagements. Basically, all the different agents that I have working on things. I could see where they are

things. I could see where they are inside of the algorithm and how far they're proceeding.

All right. So this one here I haven't talked about before at all. This one is uh called Arbull, which is Spanish for tree. And I've been trying to do this

tree. And I've been trying to do this for like 10 years, and I was just doing it with like CLI stuff. And what this

actually is is it's a it's a discrete AI or any type of functionality. You

could put it into a single little container, a single little action. That

action can run locally on a command line. It can run in the cloud. I'm

line. It can run in the cloud. I'm

specifically using Cloudflare workers.

And you can then chain any action into any other action. You can combine them into a pipeline. And if you connect it to a source and an output, it becomes a

flow. So I could run a command line to,

flow. So I could run a command line to, for example, go find all the top level domains for Tesla and then pipe that

into get subdomains and pipe that into Naboo and get the ports. So, these are all independently discreet and I have like

a whole bunch of these.

Yeah, I've got tons of them and I I just I use them to build things. So, this is an example of something I built with it.

Uh this is called uh surface uh a thing I built for myself and what this does is it takes all my 4,000 different um sources. So, I've got like Intel

um sources. So, I've got like Intel sources, OSENT people, uh YouTube videos, mostly RSS, uh Blue Sky people,

and basically take um all of their sources, process it through, and then I have a thing called label and rate, and it rates the quality of this content um

and then it puts it into this interface for me, sorted with labels and everything. So, what this means is if

everything. So, what this means is if Mark Andre writes a shitty essay, it will not show up. And if a

nine-year-old kid from Idaho in whatever I don't know what grade that is, fourth grade, I don't know, whatever grade they're in, they write an essay and it's amazing,

it will show up. So, it's like it's it's basically uh democratized the whole the whole quality thing. Um, so you can find the best stuff. So just a use case for

all the different uh arbo pieces together.

All right. Yeah. So I would say the main takeaways for this are you want to make anything that you're doing

you want to have it be clear in your mind. You do not want to have anything

mind. You do not want to have anything that you care about be like an amorphous sort of blob because I feel like that is an attack point for somebody in AI,

somebody some McKenzie, a a giant team of smiling 22 year olds to come and attack that thing and extract everything out of it and basically um turn that

into a process which basically outsources you. Right? So the idea is we

outsources you. Right? So the idea is we should get there first. Right? The other

idea, main takeaway here is when you have a unified AI system like this, you can basically take everything you're doing and put it inside the system. You

put your context in here, you put all your tools, like all those different tools that I ran, um, the PI upgrade skill, they're all available to the algorithm. The algorithm instantly knows

algorithm. The algorithm instantly knows how to use anything custom that I built.

So what that means is you end up only doing anything once and then you incorporate it into your harness and then it's just part of your system. And

I I think this is the way to go. I I

think enterprises are moving in this direction and I unfortunately people like McKenzie are moving in this direction as well. And I think we want

to get there before they do. Basically

this is the project. It's called uh PI personal AI infrastructure. It's an open source project and just had a release and I definitely recommend you check it

out and thanks for your time.

Questions for Daniel.

All right, one over there about two three minutes of questions.

Hello. Um, I'm very interested in that algorithm, the climbing. Um, do your agents, um, draw inspiration from something like gradient descent? Could

it be like, uh, you know, are they converging at something? So, can you give us a little more uh about that?

No, I think based on that question, you'll be horribly disappointed. Um, no.

It it's more so along the lines of just imagining all the things that I probably meant when I when I made this request.

So, if I say, "Hey, build me an entire role playing game system. I want

history languages terrain combat system, all that kind of stuff."

I think what we don't realize is we're it's called writer's blindness where you have something in your mind but you're not communicating it clearly. And I

think this is like a central theme for all AI. Clarity of thought, clarity of

all AI. Clarity of thought, clarity of articulation is like absolutely everything. So what I'm trying to do is

everything. So what I'm trying to do is have the algorithm reverse that and pull that out. So it's looking at my context.

that out. So it's looking at my context.

It's turning those into ideal state criteria. And the the strict rule is it

criteria. And the the strict rule is it has to be discreet and testable because when we get all the way there's seven phases of the algorithm which is roughly oh I should have mentioned that the

algorithm is roughly based on the scientific method. So it it goes through

scientific method. So it it goes through like you have experiments you have the testing and the verification is at the very end and that ideal state criteria

is what you use for the verification.

So, real quick, real quick question. Um,

as all of this continues to grow, you all have your own personal AI systems, things like that. Do you think we'll ever get to a point where enterprises adopt uh bring your own AI, bring your

own AI assistant?

Um, I think they already are right now just out of a panic. Um, because all all their management is telling them you must do this immediately. Um, I I think this is the way to go. Obviously,

they're going to have guardrails for that. But, um, I mean, this is built on

that. But, um, I mean, this is built on top of cloud code, right? Guess what it is? It's Markdown just like cloud code,

is? It's Markdown just like cloud code, which means everyone's allowed to use cloud code at work anyway. So, you just you can just like put this on top and you're good to go. It's just context

that and skills that are adding that are bringing this.

Yeah.

Yeah. I love K-pop. Who doesn't, right?

Um I was wondering how much money you actually spent on this project of stock.

It's a lot of uh you know work there.

Uh not as much as if I didn't have the subscriptions. Uh yeah, you you get the

subscriptions. Uh yeah, you you get the uh Mac subscription and you have theoretically maybe more than one account. And yeah, that's pretty much

account. And yeah, that's pretty much it. I think I think I need to go.

it. I think I think I need to go.

Thanks. I'll I'll be around if you want to chat.

Um, before Daniel goes, thank you very much.

I want to say a couple words about Daniel. So, um,

Daniel. So, um, I think a lot of us try to keep up. I

mentioned it earlier, other people mentioned it across the day, but how we try to keep up, how we have to be our own researchers, how do we have to build our own tooling all the time. And my

bets aside on vibe coding, like vibe ops, DevOps kind of thing happening for teams. when any of us in this industry want to learn about what others are doing or not to do all the research on

our own, we go to Daniel. And Daniel

really felt awkward about coming on stage and showing you scaffolding and things you're probably doing on your own, which would be relevant in three months because everything would move on.

But he's the literally in the in the committee when we went through the CFP process, Daniel didn't really have time to submit a lot. And he's not one of these superstars who just put in a couple of words and be done with it.

usually puts in a lot of effort and we automatically were about to just we didn't know who was submitting what we said this really interesting but this doesn't put in a lot of there's not a lot of effort here right ask the please ask this presenter to give us more

because we want to approve this and Greg Notch who is here out there take instead of being in the room just allow going through tickets and doing volunteering work said I would accept this talk only

if it's from Daniel not know it was actually him so I really appreciate you coming on stage anyway and showing all thank