Sam Klein: Building Libraries Any Agent Can Edit | Intelligence at the Frontier 2026

In the spirit of everything that Jeff was just describing, I want you to imagine a world where each of us, let's see, each of us has a posse. And the posse isn't just your friends, it's also the network of code and scripts and agents and small autonomous organizations that are out there in the world propagating and continuing work that you started. Uh in that universe, what are all the what all the parts of the ecosystem using as sources of truth? How are they recording cool information or knowledge that they discover? Uh how are they writing things to one another?

How are they sharing information? There's a lot of message passing. It goes a lot of systems that are describing themselves in terms of messages, and you communicate a message in the process of carrying a thread, and the thread competes. Uh we don't talk a lot about libraries. Not software libraries.

We don't talk a lot about the archives that are the long-standing evergreen representations of amazing things that you didn't just share once, you might have shared it many times. You really worked, you polished it off so you could reference this every time uh some silly flame war happens. It's it's an URL you vomit to your friend's chat where people start getting to be in the vibe about uh fluoride in the water or which shoes buy for their kids. So, let's imagine a different kind of planetary library. Like right now, we have we have a network of tens of thousands of physical libraries with a huge amount of beautiful architecture, and a lot of that was built in the last couple hundred years.

And they it was built not just as a way to store interesting published works, but to provide access. You should be You should live close enough to a physical library you could go and experience it yourself. Uh so what is the equivalent for a world that's full of agents? So where even when you want to go read something, you're also doing a bunch of other things. And so maybe in the middle of reading it, you need to go do something else.

Your agents will continue reading it. And when you're writing something, when you are polishing your monograph that you have just studied some beautiful feature of your local community and you write up a monograph for your friends, you're also publishing that with your posse. Where does the monograph go? Let's build libraries in the spirit of community libraries that are where the collection happens with the community. Libraries where agents are involved in every part of the process.

They are editing and curating the library, contributing to the collection. They are reviewing and commenting and it's their interactions that help to polish and, you know, update the little This was checked out cards that show you that something is actually used and loved. Uh so right now the archive culture is incomplete. Uh the existing models that we have for accessing it are all at risk of being closed. Uh most of them are going to start being uh at least overly your ads and almost certainly interlaced with paid place for paid salience.

So your model of the world is every now and then it's going to be a Oh. And And by the way, your ancestors also really like Pepsi. Isn't that interesting? Uh there are some prerequisites that we need for this kind of library. We need a better public catalog and we need a better authentication and curation network.

There are people at the Library of Virginia who authenticate source material at the source libraries. You need to do more of that. And then the meatiness right now is a bottleneck for most physical libraries. So we we can should talk about the labor questions, but in addition to the labor questions, there's just a lot of librarian work um that agents can help everyone do. We can clear a bunch of backlogs.

What would happen if we did? So, I'll just say this is inspired by the Internet Archive, which is the closest thing to this and any planet that I know has happens to be on Earth. You're glad you're on Earth. Uh, this is a nice estimate of all human knowledge by a mere teen favorite Wiki data. Antons, he estimates that just for notable topics in my current Wikipedia, there are about 3 billion concepts that should have the like they should have a monograph.

And we only have monographs for about 200 million of them. So, there's a lot of work to be done. Uh, there are some overlapping narratives. Just we need sense making in this in a sense of a shared library for a bunch of reasons. But I'll just call out a few.

One is to share senses across individuals. That's like messaging. Uh, one is across time. That's like passing things on to your descendants. Or if there's a war or a catastrophe, like let people know what killed you.

So, how do you capture important information even in times of crisis? And then related, um, how do you deal with noise? And in this case, like data set volatility. So, uh, think about system generating systems. What if some of the agents means your libraries would spin off other libraries?

The idea of a collection might actually create incentives for you passive agents to appear. Uh, So, you're a passive agent. How many of you have agents taking over your life right now? How many are buddy agents in the room with you now? Uh, all right.

How many of you are running more than a dozen agents? I see three hands. How many of you have acquired new hardware that is just for your agents to hang out on and have a party? Great. Yeah.

Okay. So, uh, at some point some of these agents will those sort of be a question mark. I see shows of nose. Once you have fully autonomous of things, your kids leave the house, you don't know if you're going to get an upsie. Uh so, one of the things that you might want are some models for safety or for evaluation or just for knowing direct currency to our systems are doing.

We want the same thing for our libraries. Uh libraries are being politicized while the internet was designed to be really robust. Uh our searching indexes and our AI models are now purpose. So, in addition to not being that transparent, they are the the design for robustness factor of three. Uh and that's not okay for not frustration.

So, we want to make sure that all of our sources are not just accessible, but also holistically forkable, commentable, and amendable. And like, maybe the maybe amending just means everyone can own a copy and you can make your local changes. That's That's a good place to start. But, we should all be able to carry the library that we hear about in our pocket. So, wherever we go, we have some some reference.

And then, if you think about your agent posse, each of your agents might know the kinds of knowledge that it needs to do its work. It should be able to carry all that knowledge with it. So, we need to be designing our libraries so that you can take a slice or a subset of the library and then say, "I'm going to be one of the 12,000 mirrors and, you know, co-curators because I use this every day. I'm going to do all my annotations in my library and then at some point, maybe we'll find a way to merge them back together or let everyone see everyone else's marginalia." But like, these marginalia don't just have to be for fun.

They could also be In your mind, it could be your annotation for the next reader. Which is always how I annotate my books. Uh So, we should also be keeping track of the forces that are always opposed to access to knowledge. That is a That is a universal in the history of knowledge That is right. It becomes out of fashion to share on us.

People like a photo of the people though. And uh so we get a system that is designed that is designed to help offset this. So, last I'll just mention some of the things that agents should be doing. If you're running agents forms, you should think about inviting them to learn just some knowledge curation tasks. Uh right now, if you think about the works are in physical libraries, we have maybe half a billion monographs and works that it would be useful to have good copies of and maybe translations in your language.

And there are different approaches to figuring out if copies exist. There's a lovely map of all my ISBNs that was commissioned for some very low range. It was like a bounty of $5,000. And the best visualization is really nice. It shows you for every ISBN block how many copies are known to exist because they're in the main catalog.

This is like both creating these interfaces is very agent-friendly and filling in the gaps that identifying a lack of ISBNs in this directory and a bunch of data that see is even referencing it in the internet archives very useful. We do the same thing for a lot of works little parts. Uh and every spiel could trick you. So, part of the spiel carry is really like conceals uh I don't know. Conceals We need We need We need the the attention.

Um But concealed transmission, you should be able to add something to a library before you figure out where before you get back to a local node. And uh when you have done a bunch of work, like a plurality of compute right now is being spent generating new knowledge. Monographs quality monographs style if not monograph quality to answer a very narrow question and then it gets distilled down to the response to some inter- like some interaction with a chatbot or a deep research bot. Uh we need places to publish all those things. So, make those places, even if all you're doing is publishing it locally inside your agent's home, publish those things.

Think of them as works and records that at least your ecosystem will want in the future. Build your local library, and then together we'll figure out how to federate those and make them into the global library. We need to figure out reciprocity. How you make sure that you're contributing to contributing to the shared library also contributes, like builds a commons. It just doesn't build emergent silos.

That is a good reason why you should be thinking about doing it within your own ecosystems, and not just on enterprise or iNaturalist. We need better access to public data. I know there's a lovely talk later today about public data for public data. And um And we need tools for writing from scratch. So, I'm a Wikipedian, so I mention lots of wiki tools.

We have a media uh dozen new faces in the wiki-verse. We have tools like iNaturalist that are very specific to gathering one narrow type of knowledge. But if you think about it, even something as boring as identifying photos with your phone uh can become its own entire library. And library the archive of self-identified animal photos in iNaturalist is now giving as the best available range maps for a library of species. Like this is fantastic.

This is what science library strives for. So, whatever whatever you're geeking out about, whatever your agent swarm is geeking out about, just invite it invite each individual process to record the small bits of information where if a million of those processes all did the same thing, the result would be of general general use. Uh other things agents can edit, you can check things like the Wikipedia backlog. There are about 50 amazing categories like this. A lot of them are just very boring but important.

And it's a lot easier to classify that something needs more references or something that needs to be evaluated for quality or is known to be too technical than it is necessarily to fix it. But you can imagine these most of these have an obvious mapping to an agentic skill. Rewrite something to be less technical. And then you can just look around in the in your extended library network and see does anyone have a catalog of tasks and start working through the tasks. Last thing I'll say is this can be really impactful, not just a fun experiment.

Like Wikisource, I mean, if you've ever used the Wikisource tool, it is the worst possible tool that could work for transcribing physical documents, but it works, by God. And it works in all languages equally badly. And in some of these languages, so there are over 2 million works in a bunch of languages, but in some of these languages this is the largest collection of digitized texts in that language. Uh and that's just because it ended up being easier to find people in this case human agents who learned to use a very crude tool to do a hand by hand OCR than it was to get the libraries in some of the countries this will be languages to fully digitize their collections. Uh That's all I have for you.

I would love to talk to anyone who's working on it their own agentic networks and figure out how we can link some of the texts. Oh, yes. So, on the 60th floor, we've been having some discussions with funders about the mechanisms of funding accounts, and you're all welcome to come up. There's going to be uh there's going to be a little mixer at 5:00 and 5:30 for networking guests. And uh we have we have people from a few different walks of life.

We're going to learn at 4:30 if you didn't already hear from the Buenos Aires economic director. We're going to hear about how they and Frank will engage with AI at C level. Excellent. Thank you, Esther.