From Coder to Manager: Navigating the Shift to Agentic Engineering with Notion Co-Founder Simon Last

[MUSIC] >> Hi, listeners. Welcome back to No Pryors. Today, I'm here with Simon Last, co-founder at Notion. We talk about their new vision for Notion in the AI age as a platform for humans and agents to collaborate. How the engineering and product org and Notion is changing and these new tools for thought. Welcome, Simon. Hey, Simon, thanks for doing this. >> Yeah, of course. Yeah, it's really fun to be here. >> Notions at scale, amazing platform, lots of users. >> You did start quite a while ago. I think of Notion as one of the companies that has really, like, raised AI quite aggressively. I was told you first got your hands on GPT for at a company off site in Mexico. Is that true? What is the origin story of starting to work on this stuff? >> Yeah, I think that year that was 2022. I've been watching what's going on. In general, I've just been super curious about the technology and fascinating to try everything and think about how we can apply it. It wasn't until I played with GPT-4 that it became really, really real. So we got access to it. It was a proto-chetsubt interface and a mycofiner I've been and I both got access. It was just immediately clear. I would say two big things. One is that it was just pretty smart. It could follow reasonably complicated instructions. You could write things for you, you could edit things. The second big thing was that the scope of its knowledge was extremely interesting. Super, super deep and broad world knowledge. When we played with it, it became just instantly clear to both of us. The time is now to start. We think about how to apply it. It's only going to get better. >> We were talking about Mexico, GPT-4. You guys saw it was clearly the time. Did you start with a particular vision of what you should obviously be able to do with AI and Notion or just start calling people from different teams or recruiting people and say, "Let's experiment. How did you begin?" >> We immediately had a long-term and a short-term vision. I would say the all-store of the short-term one, the thing that was immediately obvious was, "Oh, it could be a writing assistant." So it could be in your document. You could select some text, rewrite it. You could have a write text for you, maybe look something up and then give you sources or more information. That was the thing that we immediately got to work on and we started a tiger team around it and then we were able to launch it in like two or three months after that. Then the long-term vision that we immediately had was like, "Oh, the thing that looks like maybe possible is more of a general assistant." What if you could just give it all the tools inside Notion that a human would have to be able to create its own databases, query, manipulate them, create documents, edit them, and we've all of these things together to do a longer-range task. We immediately started on both. The short-term one we were able to shoot very quickly and then the long-term one didn't really work yet. That took much longer to get working. Are there specific first launch of the AI-specific Notion futures and products was one last year? No, it was a February 2023. It was going to be a launch, yeah. My timelines are wrong. Are there a few specific learnings or breakthrough moments, do you think since beginning to release that are interesting? There's been a slog over many years, over all the years at this point with many learnings. I would say, yeah, I mean, just to give you a timeline of the arc of what we shipped is so the first thing was our ratings system. We called it AI Writer. That's the first thing we launched. It was easiest to get working as it's like single-step task, rewriting editing text. There's no retrieval aspect. It was just raw access of the model to write the text. The next big thing that we immediately started working on was Q&A, doing a semantic index of the entire workspace and then letting you ask a question and I can give you an answer that's grounded in the sources. That was also immediately obvious to us that that'd be super useful. We started work on that. That one we launched in, I think, was October 2023. So we started to beta before that, but our GIA was in October. That was a much bigger effort to get working. We weren't just plucking in the LM. It was actually doing this real time updating index. We had to get much more serious about the eVal is in the quality there as well. The Q&A has been a multi-year journey. Basically, what we did is as soon as we got the notion index working, it was obvious that, "Okay, we should index everything else as well." So we index Slack and Google Drive. We're launching new ones on a regular cadence. Now I would say fairly complete. One could argue that those are very difficult problems that those products natively have not solved perfectly yet. So how did you think about taking that on? I don't know if that's an offensive thing to other product teams, but it's not working yet. It's kind of true. This has been something we talk about a lot because it's almost like what right do we even have to do this. It turns out that mostly companies are pretty bad at making their indexes somehow. It's honestly kind of baffled us a little bit, but I think my take after dealing with all of this and working with the teams trying to get it working, there's a little bit of just AI-pilled savviness that's pretty important. I think most of it is honestly just a bit of a craft and attention to detail. Particularly with this indexing retrieval stuff, in order to really get it working, you have to be quite empirical and iterative and actually be like trying queries. Each data source is a little bit special. You can't just apply one size fits all to like querying Slack versus querying Google Drive, let's say. They're completely different kinds of information. We found that there's just a little bit of like craft and love that's going to in terms of like actually trying a bunch of different queries actually using it every day and constantly iterating and rethinking and tuning how the retrieval works. How do you think about the diversity of how people organize their workspaces? And just I mean, even notion is not use of it as not homogenous, right? I'm probably part of 15 workspaces as an investor, and so I look at them and I'm like mine's a mass and these people are really organized and the workflows reflected in how their notion works. Yeah, totally. I would say, I mean, the interesting thing is that with embeddings, it almost doesn't matter as much anymore. And the AI doesn't really care what the tree structure is, for example. All the AI cares about is that there's a snippet of text that has the context you need and then it can retrieve it. And so actually we kind of advise people now, like don't worry as much about organization, just find a way to get it all piped in and like thrown in there. You still make decisions that could change performance quite a bit like chunking strategy. Yeah, yeah, super important, but that's sort of not that sort of a transparent to the user and sort of independent of their particular method of organizing things. It just seems like still a difficult technical challenge given how different the content bases are. Yeah, yeah. Yeah, I think, yeah, that took a lot of iteration. Yeah, the chunk sizing, how retrieval works, the different like steps in the pipeline of retrieval. Yeah, there's a lot of iteration on that. Ivan said I should ask you how many times you've rebuilt notion and rebuilt your harnesses? Yeah, yeah, it's kind of a running joke almost. I mean, we rewrite our AI harness probably every six months or so. And the time to rewrite has kind of been decreasing just because I mean, like, like, progress has been accelerating. I think this is honestly a really key thing and something that a lot of companies get wrong is just like doing one thing and then just like sticking with it. You really do have to keenly aware of what the current state of the model and the technology is and then designing the harness and system and the product deeply around that. And it basically means you have to rewrite it every six months. I find it pretty fun. It's part of the process. You know, you get to you have to restart and rethink it. You know, we're working on, we're about to release a new version of our harness like in the next week or two. And then we're already thinking about the one after that as well. I think that leads to a set of questions I have for you on just like how does notion as an engineering and product and research organization work now that you have the power of coding agents as well. Because I imagine like you're willing this to rewrite the harness goes up dramatically. Like agents are going to help me do it. Yeah, that's extremely true. Yeah, I mean, yeah, it's been really fun to use the coding agents. I think the ambition of what I even consider building as it's going to blow up. What do you think is most dramatically changed and how you think about how engineering and product should work in notion over the last two, three years? Oh, yeah. I mean, it's definitely changed multiple times. I mean, in terms of the coding agents, we kind of went through multiple errors. There was kind of like the tab auto complete era and then we and then we got into sort of inserting, rewriting some code. But it wasn't really until the agents started working. I would say like early last year, we started to adopt the agents like I started using clog code, I think around April last year. That was a huge unlock. Like I would say the big shift there is that, you know, you can really push on getting these agents to end to end, you know, implement and verify and maintain stuff. But it requires pretty significant thought in terms of how you architect things and what is the verification loop. But the upshot is I think if you do it well, you can be much more ambitious about your building and also make it much more robust than you could have done with humans writing it. And then the flip side is if you do it badly, it's all slop. Does that change your lens of like what teams should look like at notion, like size, seniority, anything like that? Yeah, I mean, I would say, I mean, the fundamental effect is that, you know, everyone's individual impact in terms of their output. can be much higher. And your output increasingly depends on your ability and willingness to use the tools. I think that's the fundamental thing that's happening. And then, how does that play out? I think-- I don't think we've seen that much impact on the team's sides, really. I think we like to work in a small-ish tiger teams for the most part. I think if you can make teams small, it's almost always better. That was true before, and I think it's still true. Maybe increasingly a little bit, but not that much. I think-- yeah, the main thing is to just really harness the tools. Do you think something different happens to the median engineer and organization versus the 10x engineer or the engineer 10x more willing to use the tools? Yeah, I think the gap is bigger. You can be 100 or 1,000 x engineer of these tools right now. I think the gap is much bigger. The minimum bar has not changed, but the maximum bar has extremely increased. One impact has had internally, I would say, broadly, things feel a little bit more messy and chaotic. I would say, but I kind of love that. I mean, there's way more prototypes. People are-- for example, our design team made an entire Git repo, they call it the design playground. And it's essentially a simplified notion with a bunch of UI primitives in it. And they've made it really sophisticated. It has an agent in there. And it's pretty cool because it allows them-- all the designers can spin up super high-fidelity prototypes really quickly. And so it's no longer pointing out a mock and being like, hell, this will give you a URL to a prototype that's been deployed. And that sort of thing is true all the way up and down the stack for all of engineering. Just a little more chaotic, more stuff happening. All the PRs are more ambitious. Do you draw a line somewhere about stuff that is more dangerous to touch or sensitive? There could be a risk of data loss over here and not? Or is it-- you look at it all as it's fair game? We still do reviews on all the PR requests. And I would say-- and all the PR requests are now written by agents. They're often larger and more complex. That's the worst part. But the better part is that they're often much better tested, and we can demand a much better testing for the things that merit it. I never produce a PR that hasn't been fully untentested anymore. And so you can get to a pretty high degree of confidence that it works. But it requires you're not just vibe coding by saying the thing you want. You're sort of thinking carefully about, what is the change I'm trying to make and how can it be verified and how can it be deployed safely? And then enlisting the agent to help you with that process. When you think about where you said the general assistant-- doesn't quite exist yet-- what do you imagine notions, agents, being able to do over the next year or two that are still unblocked? They're still blocked by either capability or your harness work. We struggle for a few years to build an agent. And it always sort of works. But then wasn't that useful. Largy just-- it was too early. So we tried to build an agent, I would say, actually three or four times. And then we finally launched it last fall, so like last August, September. So the-- if you use Notion AI now, it's like a full agent that has accessed everything in Notion pretty much. So that totally works. I would say like a lot of the original vision that we had totally works now. And it's like fully shipped. Last August, August, September, I would ship to our personal agent. It was pretty much every user in Notion has an agent. And it basically-- it has access to all the things that the user has access to. It can create a database for you. It can update things, create documents, search the web, do research. And then the second big thing that we just launched last week actually was custom agents. So they can basically-- you can create a new custom agent, give it a name. And unlike the personal agent, by default, it doesn't have access to anything. So you have to grant it access. But then once you do, it can actually run autonomously in the background. So for example, you can give it access to its own database to file tasks, let's say. And then you can attach it to a Slack channel. And then it will start responding to people on Slack and a filing task. That's one use case. Another one is maybe you could-- you can give it access to a database of weekly reports. And then let it search the web, research your workspace. So a custom agent represents some work or job, some knowledge work tasks that you want to be done autonomously. One thing I'm really excited about this going forward is we want it to be extremely good at bootstrapping its own capabilities, basically from an initial kernel, allowing it to basically bootstrap itself to do anything. So even, for example, maybe building an integration that we don't support yet, deploying that and then using it. So you imagine that notion agents are actually the broader definition of agent where writing code is a tool. That access to. I think it's pretty key. Yeah. I think of coding agents as the kernel of AGI. AGI will be a coding agent. And code is just a really, really useful primitive for representing deterministic logic. The thing that's really exciting about it replying it to a knowledge work agent is that it can bootstrap the capability. So like I said, if integration doesn't exist, it can build it. If it needs to connect itself to a new data source, it can do that. Given notion is at scale, but is operating in a landscape of productivity and platform players that are at even more scale. Many of these will end up with their own agents, lots of people from the labs, the Microsoft World are trying to integrate other data sources, cross attempt to integrate and index. How do you think that plays out? What do you imagine that notion agents are best at? Or what do they have the right to go do? If you look at the landscape, I would say there's the labs. And then there's maybe the software platforms. And then there's maybe infrastructure. In terms of the labs, we see ourselves as the Switzerland for models. We think and our customers, they don't want to be locked into a certain lab model. They're always releasing new versions. Any given month one is better than the other. So we want to be a place where basically you can easily get access to all the best models at any time. You can easily switch around. Do you think open source plays into that as well? Yeah, absolutely. I think the open source models are actually getting really good. There's like the four different Chinese models now that are quite good. We actually just released one of them in our agent last week. And we're going to do all four for sure. They're actually quite good. And they're way cheaper than the frontier models. So I think there's a lot of these cases where you'd want that. We want to give that as an option. In terms of the other-- so we think of our role as sort of taking all the best models we can, creating really high quality state-of-the-art agent implementations, where people can easily and conveniently get access to them, and then making sort of a collaborative workspace that is really good for humans and for the agents to coordinate on. I think it's something that's very needed in the world. And we're just trying to do it in a really tasteful, well-executed way. You were describing-- you need the index to make the agents good. You give the agents access to the tools that we humans have in ocean. How do you think about the structure of notion and where it's useful or even not useful or relevant for agents? Like blocks and databases and such? It's all still pretty useful, extremely useful. There's been a challenge to sort of-- we want to make it really convene for the agent. I think that's a new thing that didn't exist. In the past, it was convened for humans, and then we also made APIs convened for humans writing code. It used our API. So we essentially have a new customer, which is the agent. At first, that was definitely a problem. So for example, our API uses this crazy JSON format for blocks that by default is crazy for both and horrible for the agent. But we basically took on that challenge and designed just really convenient APIs for the agent. We created a sort of a markdown dialect that looks like the default normal markdown, but it's sort of enhanced with all the notion blocks. And the models are really good at it. It works really well. So that's how it reads and writes the pages. And then for databases, we use a SQLite. So basically, it's the-- it gets to speak in SQLite, which also works really well. So the default thing did not work really well, but then we just took that on as an engineering challenge. And I would say now we have extremely community APIs that the agents are really naturally good at. How did you understand or figure out what would make the API better for agents? That's a good question. Yeah, I would say it's a combination of just trying things. It's very empirical. So we're just playing around and like noticing, oh, it's not very good at that. Oh, that's way too many tokens. How can we make this smaller? And then a little bit of just like first principles thinking of like, what is it the models are being trained on? And what's in their prior? What do they know? And what do we think it would naturally be good at? And like, how does the agent loop work? And like, what would be the convenient efficient pattern for accessing these things? And then just a lot of playing around. I hear user research where the user is actually agent and then ongoing, you know. Yeah, I mean, you just chat. with it. These are always there. It's ready to talk to you. Yeah, actually that is wonderful where you have infinite access to it. Even if it access to it. Yeah, and you can you can script and scale the access as well. I assume you have actually I know you do because you walked in you're like, "Hey, I need to get access to Wi-Fi. I need power. We can't block the agents while we're doing this." What are you up running right now? Tell me about your setup. I'm working on a new prototype and so I have a couple agents working on that. And then my setup these days is just either Cloud Code or Codex. I like the CLI tools. They're super simple and work pretty well. I'm pretty comfortable in the CLI. You don't need my generated game. It's a very cool idea. I would say my whole goal these days is essentially to just have as many running as possible and to run them all the time. So for example, every night before I go to bed, I'm like, "Okay, let's go guys." Yeah, basically, what I have to do is make sure that I've given it enough stuff that by the time I wake up in the morning, it will still not be done. And so I've maximized it. That's victory. That's victory. Yeah. So yeah, like I've done that. I would say less five nights pretty well. My personal record is that I've had a coding agent running for I think it was 13 days straight without stopping and just basically working through tasks. Well, well prompted. Yes, I admit to having woken up in the middle of the night at least multiple times this week and just being like, "Are you still going?" Yeah, I know. It's kind of nerve-wracking. I always like, there's always like, I'll check it one last time before bed and just really make sure that it's still spinning. What about on the notion agents? Do you have a workflow there that is core to daily work? Yeah, I mean, I use our personal agent all the time. So it has all the context about our company, and everything that's going on. So like, for example, last night I was asking it about how the custom agent's launch was going and like what the signals were getting from it. We're super useful for that. And then I have many custom agents that are running. My personal favorite is I have an email triage agent. So it has access to all of my work and personal emails. And it just wakes up every day and just archives all the stuff I don't need to see. I turned it over time to learn my preferences. Do you actually label data for it? It's pretty easy to do this actually. So all you have to do is you make the agent and then you give it access to your email. And then you can make a blank page. It's like it's memory. And you let it edit that page. And then you just say, "Okay, now look at my emails and then interview me. Ask me which things." So it sort of will propose things that it thinks that you're archive. And then you can kind of correct it. And then we'll use that to essentially generate a list of rules about what it thinks are correct or not. And so for the first couple of days I was sort of correcting it on things. After a couple of weeks or so, I dropped the approval entirely. And it just saw my thing archives. All things I don't need to see now. Wow. It completely solved my email problems. Because for me, I don't use email that much for work stuff. It's mostly in Slack. 95% of the personal emails and working emails out again. I don't need to see it all. And so it's just a waste of time. And so it completely solved that. So now I know my inbox. It's like only stuff I need to see. I've got lots of custom agents running. There's another one that I built that triages, customer, all internal feedback and bugs. So we have a Slack channel where people just just a post random product feedback and bugs. In the past, it would sort of sometimes get answered. But then sometimes like half-hazardly get ignored. Just because there's so many teams where things. So it's entire job is just to route it to the right place. And it uses a similar sort of like memory pattern where it sort of learns on the fly where it's supposed to file bugs. And then over time, it's built up like hundreds of roles that it just sort of like learn over time. So for example, there's a plug about the mobile app and there's the route to the mobile team and then a file task in their database. Do you look at that like the generated and updated memory? Because it's legible to you to say, like, does that make sense to me? I think I did it at first. But then sort of once you trust it's kind of working, you kind of ignore it. And then if it ever breaks, I'll go fix it. It'll break every now and then. But the benefit of not reading the email is just not read it. So yeah, I mean, generally I would say, yeah, the general pattern I follow is sort of I build it as a prototype. I haven't been sort of like in approval mode where I'm sort of watching it closely. But then after runs a bunch of times, you kind of trust that it's working. Is there anything you do internally at notion to make sure non-technical teams have the intuition for how to build agents or how to like express that productivity to? Yeah, that's a great question. I mean, we do sort of workshops and hackathons pretty frequently. So like, for example, like a month ago, I did a hackathon with the the people team and sort of sort of got them. The people team has been amazing. They're actually one of the highest adopters or custom agents. They do all these kind of workflows in slack and notion, kind of like manual work like that. And yeah, I would say yeah, like people are super excited to try it and sort of like maybe just need like a little bit of a push in, there's an intuition and like I'm starting. But then honestly, I've been super impressed. I think the concept is like kind of intuitive sort of like once you get past sort of a little bit of the technical barrier of like what is a prompt and like what is the agent and how does it get triggered and woken up and like how does that even work? But then once you sort of get past that, I think it's actually a very human like interface. Yeah, maybe the maybe the biggest barrier is actually just getting people to try and assuming it's going to work at all, right? Yeah. Yeah. You and I even originally met on the internet tools for thought community. It feels like you know, the tools we have for thinking are very different now. As you're like core conception of notion changed for the last few years because of all the AI stuff. Like what is the what what thinking does the tool do for you? Should agents do for you? What do you get to do? Yeah, I mean, it's I would say change quite a lot. I mean, we're all these speaking before AI, our goal is to create the best tool for humans to directly perform their work. And then now the goal is to create the best tool for humans to manage agents to do the work for them. That's a big shift. That's a pretty big shift. It's it's pretty fundamental. But it turns out that you need most of the same primitives. You actually all the primitives that we built are actually still extremely useful. It's more that we just needed some some new primitives like representing what is an agent and how does it interact with your pages and databases. But you still need the same primitives. You still need a document. It's an unstructured way to write stuff. Agents love to write markdown documents. So it's still very relevant. And you still need a database. It's he still needs structured data. You know, if you're working with your your swarm of like 100 background coding agents, you don't want to have 100 chat threads. You want a Kenban board. It's the same as before. Makes sense. You still need the coordination structure. What is one thing that just because you're ahead of the on this stuff and then trying to figure out how to bring. No notion and then users along with you. What is something that's really changed about how you personally like build even in the last six months? I mean, it's completely changed. I haven't written code since like last summer. I don't type code anymore. Yeah, it's it's it's complete shift. I mean, we went from humans type all the code to like we're still typing, but we like tab complete to sort of like we talk to the agent and it sort of does little task for us, but we are still in the outer loop. And then now it's more like I I design a end to end task that involves like making some change and end to end verifying it. And then I'm just the the outer the outer verifier sort of like like double checking at the very end that it that's correct. And if it's going off the rails kind of like monitoring it. So it's a it's a complete shift. You know, I'm I'm now like the agent manager is that of the quarter? Amazing. Well, thanks Simon. This has been a super great discussion about how we're all going to become agent managers and hopefully in notion. Cool. Yeah. Find us on Twitter at no priors pod. Subscribe to our YouTube channel if you want to see our faces. Follow the show on Apple podcasts, Spotify or wherever you listen. That way you get a new episode every week. And sign up for emails or find transcripts for every episode at no-friars.com. [BLANK_AUDIO]