Agent Highways: Building a Reference Map for Agents Across All Your AI Tools

I think that some of the best parts of the conversation are at the beginning when it's a little unfiltered. And then some of my favorite moments in other podcasts are that initial riff around moment.

Yeah. Yeah. It kind of makes you feel like you're stepping into a conversation just as it starts without having to make it super formal.

Our lack of preparedness would be hilarious for people in this one. Typically we're quite well prepared.

I usually try to open up a scratch note and just write some questions and thoughts beforehand, but. I also like going into things blind too. All right, Bryan, you dropped a topic in a call with our team today and you used a term that for better or worse, I'm going to blame it on the fact that I was moving and probably missed the call you mentioned it, but you dropped this term and I think for us, we're developing this, this conviction that everything is a framework at Notion State, and it feels like a framework. And so I'm gonna come at it maybe more like, you know, somebody that's a listener on this term agent highway. You, you dropped that like it was, it was something that the team kind of recognizes and knows, and I think they do. So from, from my, my position of listener, tell me about agent highways. What exactly does that even mean?

Yeah, really excited to talk about this 'cause I think it's a really important topic in today's day and age and how we can use Notion and encourage Notion use across a tech stack. But before I go into the definition here, I'm going to do my new favorite thing and just see how good the title is. What, Josh, what do you think agent highways are based on the name and the context in which it was shared earlier?

I would imagine this has a lot to do with, let's boil it down to process, like work being just processes and, and a network of processes when you've got artifacts. And so you've got people traditionally before agents, you'd have people that have to interact across, you know, they gotta go get this resource, they gotta get this resource, they gotta do something to it, they gotta send it off for review, they gotta get it back, then they gotta send it off. And you know, now it's past their, their kind of stage of working on that artifact or whatever it is, you know. So if you introduce AI into work and process, what kind of highway, I suppose, or what kind of process path are you laying out for an agent? You kind of have to tell them, where to go and what to do when they get there. That would be kind of my thought is maybe defining the process of AI agents and what you want them to do. Let's go with that.

You are pretty close. Uh, honestly, it's like pretty accurate and it's an analogy or metaphor I've been using a lot lately is the way that MCPs function are like, you're just asking a question. Imagine you're in a Costco and you're at the front door of it. And you just ask a question like, go find me the pretzels. And the MCP is just going to look up and down and through every single aisle and everywhere until it finds it and then gives you an answer back. And in my mind, that's incredibly inefficient, even though they can do it quickly. It's not an efficient main means for doing it. And it's also variable because there might be different pretzels in different locations and it's going to return to you what it thinks is the best answer. And because I don't like that method of using MCPs, even though That's how Notion AI connectors are working, and that's how Claude and ChatGPT are using it when they're connecting into all your different software portfolio. So the idea of creating agent highways was how do we create a map to be able to give you the exact cross-section of your workspace for it to be able to work within? So something that happens a lot in, in our delivery models are we build around a few central core systems. So you have your project ecosystem, or you have your go-to-market ecosystem, your product roadmap in that ecosystem or a CRM. And each of these has a very defined set of databases and core objects. And each of those objects relate to one another in very intentional ways. And what the idea of the agent highways is, is to map all of that out in one page that serves as your reference guide that's loaded into your Notion agent instructions, that's loaded into your Claude projects, that's loaded into different ways that you're interacting with these MCPs to give it a very clear direction on how to get from point A to point B based on the prompt that you're receiving. There's a lot more about this, but that in a high level is what it, what they are.

My initial, my initial read on this was that it was more of like a GPS. This is how you get from point A to point B. And of course it understands the maps, but it sounds like what you're saying is it's actually more like the map itself of your information. It has no preference necessarily on how you get there. It's just telling you what's there, what you should expect there. And then of course, you know, when you think about the process of doing things efficiently, that's a, that's a different kind of discussion. It's more like a GPS kind of deciding the most efficient route to get there. And I'm sure, you know, there's maybe like a different framework, but it sounds like maybe what you had in mind was more of like the map itself of information. Like, what would you kind of consider that information architecture for agents?

Yep. I think that is spot on. It is very much so the map. It's not the map, but it's the highways on the map. Connecting the different islands or, or cities of data. Like, that's the intention of it. And the reason this is important is that— use our workspace, for example— we have one core company tasks database, but we also create and duplicate templates before we deploy those into all of our client workspaces. And having done a couple hundred projects now, we might have 300, 400 databases that have some variant of the word tasks in the database title. And when the MCP is looking, it may look across all of those databases before it identifies the right, right one. And the goal here is to be able to tell it, this is the one that I want you to be referencing. Don't even consider any of the rest of the noise.

So practically, how does this actually get set up in Notion? Like, I have an idea, but maybe tell us like, uh, on a tactical level, how do you connect this and, and actually use this concept?

Yeah, before we had the amazing power of AI, the way that we would've done this is we would've mapped out every single database that's a core database in the system, and we would have created a diagram probably in Miro or Lucidcharts or what have you that you'd be able to say, this is your ecosystem, your projects database that's in the middle, and it relates to tasks and files and status updates. And those are the connections that exist. And then a human would have read that and said like, oh, I understand now how these connect together. Uh, so that's that in concept of what we're doing is what we're doing here, but we're taking it much further to supercharge AI in doing this where we're saying, these are our core databases, identify all of the relations across all of these, create a diagram to help you map these, but then go even a step further and start to understand what are the core intents behind these. You reference tasks and projects when we're saying something needs to be completed associated with this project, or you reference invoices and opportunities when we're talking about pulling the invoicing schedule from a contract and loading that into our payment system. And you're giving it these key trigger phrases to understand and help the agent understand what the intent is from the prompt based on what these databases are. Because there's natural flows, a handful of natural flows for every core component of your ecosystem, and we're just helping create a guide for it so it doesn't have to think it on its own. It already has a, a map.

So is this one page that lives in a Notion workspace, you know, called Agent Highway or what have you?

It's one page per ecosystem that you want to create a highway for. So you might have one master agent highway in a workspace that's really well organized, like Notion State has one master agent highway. But if you have a 2,000-person workspace and you might have 10 core ecosystems, you probably want to create one master one for the handful of admins that are working across those, but you likely want one for each core ecosystem that collections of individuals are going to be operating.

And then as far as AI actually using this, do you set this up as a skill? How do you expect users to actually make sure that this is used? Like, how does a user, you know, utilize this tool to actually do something better?

Inside of Notion, it's actually really nice that you can just load a template into your Notion agent instructions. That's how we encourage it to be used. Basically, you create a preconfigured instruction page for everyone to use, and you have the Agent Highway preloaded inside of that. But when you start to work in other platforms like a Claude project or like a ChatGPT project, you need to load it into each of those projects. So you can build a skill in Claude or a skill in chat to reference this, but it still takes a little bit of manual effort to have that system use this highway. Yeah.

Yeah. I don't want to go too far into this because this could be a totally different episode, but in terms of like networking or like even best ways of setting up agent instructions and then in Notion, you've got Nozi, you've got custom agents, you've got skills, instructions, normal pages. It sounds like Your recommendation is that every person with their personal NOSEE and their NOSEE's instructions, that they would all reference a shared agent highway page. So the agent highway page is not duplicated across every person's individual NOSEE instructions, but you keep a centralized agent highway and then every person's NOSEE instruction just references that one so that if a change is made, every nosy already knows how to, you know, it's just referencing that with one singular place. Does that sound accurate?

That's correct. We don't want people to duplicate the highways because we want to be able to evolve the highway page as the system evolves. And we can't cascade updates through other people's private pages or their, their instructions if they're not using a centralized record. So we want to make sure that we're creating, creating one source of truth for this record.

Yeah. AI agents. And instructions in a collaborative environment, I think has doubled down on the, uh, the need for people to think about or to think in, in like a, I don't know how to say this, like a networked kind of framework. And what I mean by that is like when you network information, you, you have a single source of truth. We, we always talk about that as like a, a database. Like you need a single source of truth for your projects and single source of truth for your tasks. It's getting more and more important to get more granular and say, not only that, but you need to be, you need to have a single source of truth for this sort of artifact or document because it'd be super easy. It's, it's, there's, there's a bit of difficulty in setting up a database, but it can happen in Notion a lot, pretty easily. It's even easier to just take the agent highway document, copy paste it into your own Nozi and call it a day. And then finding that, or like, you know, recognizing that there's a disconnect there is even harder. So it, this, whole approach is just like further doubling down on the fact that we need to think in, in like a, the best way I can think about it is like a network, which, or a graph, which Notion is not a graph. It's not a network, but you almost have to think in that way to make sure that you don't duplicate information and cause problems that Notion was meant to solve.

Yeah, I, you're, you're absolutely right. It's, it's hard to avoid. Because we're so used to copying and pasting out of things. So it's hard to get used to, especially in Notion, subscribing to a page basically. But for you to scale the system and to grow it and continue to benefit from that system as it grows, it only makes sense to try to centralize these capabilities or these rules in one location and then reference them. Custom Agents actually introduced a really— and we talked about this in the Ava episode about the CAST framework, and a lot of what she's building is to reduce this problem, but Custom Agents created a really challenging new system for users to understand and get used to because custom agents are abstracted out from a normal database page and each custom agent page is basically its own custom surface to input information. And every time you can't subscribe to custom agents to personalize for yourself, you have to duplicate it. So you're basically duplicating instructions that only live with inside that custom agents page. And it makes it very hard to scale a solution like a pull task from Outlook for 100 people because you end up with 100 different sets of instructions. So using this concept that we were talking about, the agent highways, and using that page as a reusable resource that's just mentioned or referenced is the only way to scale the use across these, these different solutions today. Agent highways.

Agent highways.

I'm just going to give it one more, one more hit. Cause I think it's really important to narrow in on this. If you don't have a central system in Notion, an agent highway doesn't really work. You can still create your own, I guess, navigational resource for Claude or for OpenAI or ChatGPT to like find where it should deposit things. But it, it makes it very hard to do this. And what ends up happening, and we see this at organizations, even where we have good systems, if they aren't using this Asian Highway, they end up just using ridiculous amounts of tokens trying to have their LLM find where that data is and edit that data and where it should be, because the majority of the tokens are just searching and searching and searching. And this makes it much more efficient from a cost perspective, not just a speed and a usability perspective. And that's the real intention behind it. It's like, how do we make our tools more efficient for us? And that's why we created this to support our clients and our internal operations.

All I can think about when I think of maps is, uh, is my core memory of watching The Goofy Movie.

Are you familiar?

No? Oh man, it's my favorite movie.

I mean, I've seen a Goofy movie.

I don't remember the, uh, they road, they road trip and they have, uh, Goofy, Goofy's dad had his old like roadmap from when he was a kid. So then yeah, Max, his kid. He's got the roadmap out and they're like, you know, circling places and then I'm just telling you like what's in my memory for no reason whatsoever. Um, but yeah, Max changes it 'cause he wants to go somewhere else.

The map, I mean, the map example and this reference is exactly why I, I chose the term highway is just using city roads and local roads and, and back roads and country roads. You could get from Columbus, Ohio to Chicago in infinite number of ways. But the fastest way is to take Route 70 over, and it's a highway that we created to kind of connect it to. And that's like really the intention behind that specific.

I think I get it. I mean, don't, don't take this as any sort of like me pushing back, but if I was being, you know, if I'm being— if I care about the semantics or like the, the way it kind of read to me, I definitely read it as like more of this is the highways that you should take, whereas it sounds like it is more of just like a The entire map, like, yes, it should be like the highway map, but like the idea of a map to me says like, this is comprehensive. These are all your options with no, with no bias or opinion on what you should be doing yet. And then when I think of like, if you, if you told me highways, I'm thinking, oh, you kind of, you've kind of picked them for me, so to speak. But I totally, I get, I get the, I think the picture is, is good when you talk about like country roads and. And city roads like that, you could definitely get too much, too granular. And then you're talking about things that you don't actually— like, the agent doesn't need to spend its tokens understanding anyways.

Yeah, I think that that's spot on. I, I haven't articulated it well yet, but honestly, using the map and the cities example and highway paths is probably going to be a really good way to do this. And I don't actually remember the highways to get to Chicago and Columbus, but if I would have used Kansas City, 70 would have been right, and we could have also used 40. And both of those would've been two different options of getting to the same location in a similar way. Anyways.

Yeah, well, it's like, what is the, you know, you talked about like it has to be built on a centralized system and it's like, well, what is the picture if it's not a centralized system? And the first thing that came to my mind was like, well, you got like 10 Milwaukees, you know, your maps aren't like you know, I don't know if that's actually like a, if that actually translates, but because, you know, you'll have different projects databases for potentially good different uses, but a lot of times they're made in isolation. They're not connected, or it's like half the city doesn't exist on the map, even though it's there. You just can't see it on that map. It's erased or blotted over. I don't know.

Titling and the description could use some work, but the concept in practice is very effective.

Yeah, yeah, for sure. Frameworks for everything. You know what's funny is, um, when you, when you brought this concept, like this, this had been when this might've been before custom agents when Simon did his like multimodal agent thing that I, that I downloaded a big portion of his instructions included a table. It was a simple table and it's, and he had already kind of prefilled out on the left column, like projects, tasks, whatever. And then the right column, you were meant to link to the actual databases you wanted to use. And then in the instruction was already instructions for the agent to say, utilize this table to know which actual database we want, you know, to use for this use case.

Spot on. Yeah. The, I would say this is an evolution of that framework. And honestly, when, when you asked the question earlier and I was explaining what we used to do, it's that database diagram, but then we would also create a database glossary. And those are the two resources for humans to be able to understand how things happened. And the database glossary is what you're referencing in Simon's tool. And it's It's incredibly important, especially for onboarding people. And honestly, we create these agent highways as the AI version, and then we create a much more dumbed-down version for the human version now. There's way too much information in the agent highway doc. Like, users don't need to be told how to infer when a task should be referenced regarding a project. Um, but yeah, it's, it's very similar.

You know, um, back to like my little kind of pet peeve on the whole like AI seen where we call the computer ones agents. I wish agents was humans and computers, whatever we call them. Now I'm, I, I feel like we're gonna, we're gonna just land on Tron, which is like users. And I don't know what they call all the computers, but we've basically landed on agents, so we might as well call humans users.

Yeah, we totally can. I find it funny that we talk about humans in this way now. Like we've never referred to it. Like a year ago, two years ago, we never talked about things in this. I'm like, the human will do this while the agent will do this. We've never, we've never had that sort of language before. And now it's common speak.

Weird.

It is weird.

Gotta keep up. I feel like there's like this small possibility of a reality in my mind that thinks is like potentially real, which is like AI just completely, it's not like Skynet, but it becomes either you are above the threshold on AI to where you're, you're still able to create more, I guess, value at that leverage point where you found a place to be higher leveraged than AI, if that makes any sense. But if you're, if you're below it, you keep getting lower and lower, like the rich get richer, the, the poor get poorer kind of concept. And, you know, if there's a reality where that exists, it's like we're on the— I feel like because we were like, we have a business that's growing, it's around AI, you know, it's almost like maybe in 10 years or 20 years in hindsight, we're like, wow, we were super lucky to grab onto that thing while it was taking off. I don't know.

I think you're 100% founded in this, this thought and this concern. And the meme is the permanent underclass, right? Like I'm just trying to avoid the permanent underclass. I like to think of Altered Carbon. That future was really interesting to me. And have you seen Altered Carbon? Oh, it's this movie where there's a chip in the back of the sleeve's neck, and that is the consciousness of the human. And you can go from sleeve to sleeve, or you can make clones of your sleeve. So if you die, you're— they call it a stack, just helps you grow. So the really ultra wealthy people, they would have the same sleeve and thousands of clones of themselves in case that they permanently uploaded stacks so they could just continue to live uninterrupted even if they died. And so thus the wealthy people continued to grow their wealth and can continue to create a divide between them and everyone that was in the permanent underclass just kind of existed. And it's, uh, they're like, where we are heading right now is just there's this likelihood of massive inequality, rapid inequality widening because of the access to this. So like, we pay $200 a month so we can have super, supercharged clot. Like, 99% of the world can't, can't do that. And, uh, like, it just Yeah, it's hard.

Yeah, we'll see what happens. It's such a black box. Like it's actually really hard to know what will happen.

This is one of those situations where pulling threads from a previous conversation where we talked about humans are innately good, in my opinion, where the few will have the ability, the few that grow into bad will have an outsized impact on the entire populace. And like, that's what's scary about capitalism mixed with AI and where we're going. It's a very scary prospect.

Yeah. I mean, it is basically, it's basically like, wait, we talked about Silo, but you haven't read them or seen them.

I've seen the first season. Yeah.

Okay. Well, anyways, I guess it's like the, the more and more accessible, highly destructive information and knowledge and technology becomes to the like world-ending technology is getting closer and closer to the consumer, you know, not just the smartest people on earth that are just discovering it using, you know, particle accelerators. That's where it's like, it only takes one, right? It only takes— anyways.

The democratization of access to too much power. That's not necessarily my concern though. I think that that's probably our only saving. Grace in the future. And that's where Musk was heading with the whole Neuralink concept, whereas like, if AI isn't embedded in the human, then the few are going to control the many. That's why he was trying to supercharge every person with their own AI. But now, I mean, we'll see where that heads, but it's very much the, the Darios, the Sams, the Musks that have an outsized control over the direction of humanity at moment. And that's more my concern, not the, uh, individual that has too much power because of the tooling that they have access to.

Yeah, sci-fi becoming reality.