Latest / Technically Speaking / We can do better than --dangerously-skip-permissions (w/ Michael Grinich, Founder of WorkOS)

Transcript

We can do better than --dangerously-skip-permissions (w/ Michael Grinich, Founder of WorkOS) · Technically Speaking

• Megan Reynolds: WorkOS is a $2 billion billion company. You've got thousands of customers. You've raised You've raised around 200 million from top investors. What does WorkOS actually do? Michael Grinich: We help people with Auth. I mean, that's kind of like- if you use it today, I would say, what, what most people know us for. So, um, you know, when people sign into products like, um, you know, ChatGPT or Claude or, you know, Cursor or Perplexity or lots of other, like, business products, you know, like Sierra or inference companies like fal or Baseten, we help them with really the identity security part of their product- and specifically focusing on that for business customers. So we help them go upmarket and sell to enterprise organizations. Michael Grinich: We do a lot more than that, but that's kind of the main thing that people come to us typically for is to start, start with that. And then over time, we help them take their business and their product and make it enterprise-ready- which means add other things around, you know, user permissions or observability or integrations or encryption or compliance or, like, all this stuff that you need to go continue to scale up market with bigger customers. Megan Reynolds: with the core, Auth loves an acronym. So maybe talk us through what is SSO, what is SAML, what is SCIM. Give us a briefing. Michael Grinich: There's lots of different acronyms in IT. Um, and generally, it comes from standards. So single sign-on, SSO, is really the, the type of authentication. It's not just, you know, uh, logging in with your password. It's using the identity system of a company. So if you go get a job at a big company, you go show up on day one, um, they're not gonna say, "Go create, you know, a password for every single app." You'll have one login for the company, and then you'll use it everywhere. Michael Grinich: And the, the version that most people know is Google Sign-In. So you have, like, your Gmail account, and you use it everywhere. You can sign in to Yelp. You can sign in to DoorDash. You know, you can sign in whatever. That is consumer single sign-in, single sign-on. Um, this also happens for companies, but it's your company identity, so that's SSO. Um, SAML, S-A-M-L, um, is the protocol for doing this with enterprises. So it's not Google Sign-In, which is OAuth, which is another protocol. Um, uh, it's a, it's a security XML-based markup language from, you know, the '90s. Michael Grinich: Uh, the secure access markup language, I believe is what it stands for. And what is a protocol? Oh, a protocol is just how two computers connect to each other. It's just, like, really the language that they speak. And so when you're signing into something, you know, there's a, there's an app that somebody has built, and there's an identity service somewhere that can prove you are who you say you are. And those two things need to talk to each other. And so the, the challenge now around building software in today's world is not just building it by itself, it's that your software needs to work with other people's software. Michael Grinich: And in the enterprise, it has to work with tons of different systems that are out there. The identity thing is just one of the pieces, but getting other types of data, other pieces of, you know, logging or observability information, um, roles and permissions is another thing. There's a whole protocol around that. Um, and so WorkOS really helps plug in all this stuff. We, we sometimes, uh, talk about it being acronym soup, you know- when you have to go figure this out. There's Megan Reynolds: acronyms for your- Michael Grinich: Yeah, and there's, and there's tons of them. Um, people in IT really like their acronyms. There's MFA, multi-factor authentication. It's really just when you get, like, a text message with a code you have to put in, but that's… You know, even that is called, uh, TOTP, time-based one-time passcodes. Megan Reynolds: And SCIM to round off? Michael Grinich: SCIM is, SCIM is similar, um, to SAML in terms of authentication- Mm-hmm … but it's actually for user provisioning. So if you get hired at that company and there's all these different services they use, how does the identity system let those apps and services know that you've been hired before you sign in? And, and m- even more important than this, when you leave the company, how do they deactivate your access? And that's what SCIM is for. Um, it for System for Cross-Identity Management, SCIM. Pitch Megan Reynolds: perfect. Michael Grinich: Yeah, I just- I just- Yeah, it's, they're deep in my brain at this point. Um, uh, but really SAML and SCIM are kind of the two sides of the core- Mm-hmm enterprise identity access management- Mm-hmm … you know, lingua franca. Um, and we help people add both of those. So not only if you're signing into ChatGPT do you, you know, go through SAML if you're a workplace, but actually the user provisioning and role-based access control and all that, um, also goes through WorkOS. Megan Reynolds: Love it. Okay, so SSO for login, SAML for how it kind of comes together and the protocol works, and then SCIM for when you leave, how does that get processed, very simply. Michael Grinich: Yeah, and the, the challenge of these things… You might think with a protocol it's kind of easy to build. There's these documents online that say, "Here's what the spec is." But the challenge is everybody builds this slightly differently. Megan Reynolds: So tell, so tell us more about that. So you're selling to developers- Why aren't they building this themselves? Michael Grinich: Some people do. I mean, they try. They spend a lot of time on it. And, y- you know, if you only have maybe a handful of customers, enterprise customers, you can probably hack it together. But, um, you can only really hack it together for that small set of customers. As you get more and more of them, it's like every company is a unique individual snowflake in what systems they use and how they integrate and what proprietary IT systems they have. You know, the identity system of Disney, even though it might still be SSO and, and might be, you know, SAML, is, is different than the identity system, you know, of like Nordstrom. Michael Grinich: And it's different than the identity system of Kellogg's and JPMorgan, And like they're all slightly different. And so even though there's a standard spec, there's not really a standard open source code base or standard thing you can use to integrate with all these really legacy proprietary things in the world. And that's why these big AI companies pay us to do it, you know. To, to be honest, it's because they don't wanna have to do all the work to integrate it everywhere. And then once we've done it for everybody, you have this kind of like compounding effect- Mm … where like every bug that we fix for OpenAI, you know, you get as a developer, and then everyone else in the ecosystem can contribute that back. Megan Reynolds: Interesting. So it's compounding. Um, but you mentioned there are lots of people who will do it themselves. You know, what's a, what's an example of a company where you're like, "Yeah, you know what? They, they probably should do it"? Michael Grinich: You know, what's funny is like because we've had this compounding effect now at scale- like the answer has changed dramatically. I think if you would've asked me a few years ago, I, I would've told you that it makes sense for really large companies to have this for control or like security, if they have their own security team. Like at one point, um, you know, like OpenAI's security team today is larger than what WorkOS was the whole company was a few years ago. But now that's not the case. Michael Grinich: We're much larger. And also, on the security front, we actually see things well before other people do focused… Like we're focused so deeply on just like this niche problem, that whenever there's vulnerabilities reported or issues, we can, we can patch them and fix them immediately. We see it before anybody else. And so I, I think it actually has changed quite a bit, where it do- really doesn't make sense for people to build this or do it- Anyone … you know, really at all. Megan Reynolds: No one should be building it themselves. Michael Grinich: Yeah, I mean, there's, um, sometimes people really love open source, and so they're like, "I can do this with open source," or, uh, "Certainly open source should be better." That was actually my experience years ago. Personally, I thought this should… there should be some company that's already doing this that's open source. And the open source stuff does work for a very small set of use cases, but the problem is it's not tested with all those legacy proprietary IT things. So the open source stuff works super well for like Google sign-in or like sign in with your Facebook account, you know, or whatever. Michael Grinich: Like the that are very popular that everybody uses. But it hasn't been tested against, you know, like Amazon.com's corporate identity system or like, like Disney's identity system that they built from scratch, their engineers there. It's Star Wars themed. It's called Solo. Has a picture of the Millennium Falcon. My word. Megan Reynolds: No Michael Grinich: open source thing hasn't been tested with that because no one could test it. No. How can like open source developers, you know, get this thing plugged in? Megan Reynolds: you mentioned security vulnerabilities, that you're ahead of those. Like what's an example of that? Like what are the things that you're tracking? Michael Grinich: Um, so as an example of this, like um, a bit over a year ago, and a half, um, we had a secure- security res- researcher approach us- Megan Reynolds: … Michael Grinich: um, with a zero-day vulnerability that they had found in this xml-crypto package. What's a zero-day vulnerability? Uh, a zero-day vulnerability is something that's been in the It, it's been in the code for a long time. It's, like, been in since zero day, when it was launched. So it's not a new vulnerability that has just been introduced. It's something that, um, like it, it could've been exploited years ago. And this zero-day vulnerability had been exploitable for 13 years. Michael Grinich: And it's the XML crypto library, which is used by Node SAML, the Node.js SAML. And we found that Notion and a bunch of other companies were vulnerable to this. We tested our own services, and it's bad. It was a very bad vulnerability. It's on our blog. Uh, we, we wrote about it. It's called … We call it SAMLStorm, and it's- Wow. Bad Megan Reynolds: in terms of what could ha- what could happen if Michael Grinich: it was found? A full account takeover. Oh. Full account takeover, yeah. So you could, like, Wow … log in- Wow … as another person in another organization. Megan Reynolds: Bad. Michael Grinich: So if you had this vulnerability, you know, and you had a, you know, a Notion account, you could go log in. I'm j- I'm not picking on Notion. There's many others that were vulnerable. Um, No- you know, know, Notion's actually not a customer of ours, WorkOS. They started before us. They've, they'd built, built this up before we existed. Uh, but we had this vulnerability, and we patched it in our system with about th- within, within about 30 minutes of finding it, discovering it, reproducing it, and patching it. Michael Grinich: But then we're sitting on this, and we're like, "This is a, like, kind of a ticking time bomb." And a researcher approached us with it, but once that happens, there's sort of a timeline before the rest of the world knows about it. So in addition to patching it for all of our customers, we started reaching out to companies like Notion- talked to their CIO and helped get it fixed. I was t- texting Ivan, Megan Reynolds: you know. Are they now a customer, or are they … Michael Grinich: Not yet, no. I mean- Okay … it's very, it's very hard to migrate at that scale. Um, but you know, like, OpenAI and Anthropic and all these companies, and Cursor- Mm … they were immediately patched. You know, they didn't have to think about this- Mm … at scale. Um, and we, we regularly work with security researchers around this area, so that's why we were able to see it first- Mm … well before, before anybody else. Um, and if you're not tapped into that world, you know, you spend all this time making your product amazing- to have it kind of fall over on its face with your biggest customers with the t- this type of a security incident would be- Megan Reynolds: … Michael Grinich: you know, catastrophic. And so that's- Yeah reason why people work with us is 'cause we're kinda, you know, really on top of it. Megan Reynolds: And does that mean security's involved in that buying process, or is it purely Michael Grinich: engineering? A little bit. It's generally engineering. It's generally engineering, but um, y- people do look at us from a security perspective, you know, 'cause, um, you know, we're, we're a target in terms of that. We're … It's kind of like running a database- Mm-hmm … or other type of critical infrastructure, and we have a whole security engineering team since very early on. We've, we've spent, like, several million dollars on, uh, what people call is pen tests, penetration tests- Mm … which is an external security firm. Michael Grinich: You pay them money to try to hack you- … um, and then provide you the findings and results, and then you fix stuff. Um, and then we also have, you know, researchers that we, we pay what's called a bug bounty. Um, for this specific ins- instance, it's like a, it's a pretty good … I think for that SAMLStorm we paid, tens of thousands of dollars. Yeah, maybe like 50, $60,000 for, for just that one vulnerability. Um, you know, and it makes sense for us to do across the full ecosystem, and we can economically support that. Michael Grinich: Um- XML is really hard to work with. That's the SAML spec is, like, built around XML. It's really, uh- Megan Reynolds: Yeah, what makes it hard to work with? Michael Grinich: Uh, well, so as a standard when these folks build this stuff, usually they're working through the ideas of what should go in the standard. But practically speaking, when they build it, there's a lot of ambiguity in actually how it should be implemented. there's no set of tests that you can run on your service to tell you, like, "I've implemented SAML correctly, and it's secure." There's no, like, benchmark. There's no, like, test suite. There's no verification test. And in particular, like, XML has this characteristic around canonicalization of the, the file format, and the, the SAMLStorm bug was actually due to, like, a parsing error with comments, and comments that exist in the XML spec but were never used in SAML. Michael Grinich: So it's this, like, layering of these different things- Mm … on top and the interplay of them that creates the vulnerabilities- Megan Reynolds: … Michael Grinich: in this situation. Megan Reynolds: So there's risk, there's risk that there's an actual zero-day vulnerability baked in there, but I guess there's also risk that if you configure it in the wrong way, you could also be vulnerable. Like, is that the challenge for you? Michael Grinich: Yeah, absolutely. People talk about it as a, like, a foot gun, if you ever heard of that term. Which means, like, you're holding the gun- Yeah, yeah, yeah, yeah … at your foot, and you shoot yourself in the foot. Uh, this is like a foot cannon, you know. It, like, takes off the limb, you know, if, if it gets, if it gets, uh, exploited in that way. Megan Reynolds: Um, what is the difference between authentication and authorization? Michael Grinich: Good question. So when people talk about auth, you know- Mm … the WorkOS is like auth for, you know, your app. Um, it's a little ambiguous. Are you talking about authentication or authorization? And when you'll see people write this online, oftentimes they'll shorthand it. Instead of just A-U-T-H, auth, they'll say AuthN. Letter N for authentication. Whereas AuthZ is for authorization. So authentication, to authenticate somebody, is really just them proving they are who they say they are. So as an example, if you think about, like, the physical world, like your driver's license is authenticates you. Michael Grinich: It says you are Megan, you know, and the government can prove it and there's a picture of you, and so I can guarantee you are who you say you are. But that doesn't tell me anything about what you're allowed to do. So authorization, authorization is more about capabilities. What do you have access to? What permissions do you have? What can you go do in the world? You know, if on the back of your driver's license it says that you have a motorcycle endorsement, you can ride a motorcycle. Michael Grinich: Or you can operate heavy machinery, right? Commercial, you know, truck endorsement. That's a form of authorization. It's a capability. Or in the OAuth terminology, like a claim or a permission that gets associated with it. These things are usually put together, but most people when they, they interact with them, like if you've ever used Google Sign In and you've signed into a service and it's like, "I also want access to your Google Docs." Or, "I want access, you know, to some other, you know, thing that you have, your Google Drive or whatever," or, "I wanna connect to your Facebook, like Facebook Connect, but I want access to your photos." That's actually authentication and authorization happening in the same flow, but they don't necessarily need to be, um, the same thing. Megan Reynolds: And have you always managed both? Or how have you- Michael Grinich: One kinda falls off from, from the other. Yeah, usually after you authenticate somebody, you want to do the authorization step. Um, within the enterprise, there's, the different protocols can do it different ways. So SAML is generally for authenticating people, and it's actually locally within the organization. So- … um, if you work at, you know, a bank or something, we can s- we can say that, like, it's this Megan at this bank. But outside the walls of the bank, we actually don't authenticate who, who you are. Michael Grinich: You can't take that identity and go use it, you know, at the grocery store or something like that. They're, like, different environments. And so authentication is usually within the boundary or some permission, you know, some, some, some kind of permissible area. Um, and then authorization is the capabilities within that, and that's usually where it goes into role-based access control. RBAC, another acronym. And role-based access, RBAC people call it, is associating your identity with some type of role. in the in the, the, um, you know, environment. Michael Grinich: So you could be an administrator, you could be a member, you could be a, I don't know, the finance person, a billing contact or something, and all of those different roles can have different permissions associated with them. Megan Reynolds: Okay, so it's a way of, like, manage- like, a framework for managing what profile a person has access to- Yeah Yeah, to What within Michael Grinich: exactly. It's kind of like your org chart, you know? Okay, I know who you are, but, like, what can you do inside the company? Like, what capabilities should you have, and what shouldn't you have kind of. Megan Reynolds: And so you, you bought a company called Warrant. Maybe tell, tell us more about that, how that fits in with authorization. Michael Grinich: Years ago we, we acquired a company called, uh, Warrant, which was, um, a few guys out of YC that were building, um, essentially a clean sheet implementation of this authorization technology that Google wrote a paper on called Zanzibar. And it's a very interesting paper. It's, it's actually pretty easy to read. I don't know if you're, like, a mathy, you know, engineer person. Um, really Zanzibar, what they found was as authorization gets more complex… The example I gave of, you know, you're just an admin and you're a member and you kind of just got two levels, you're, okay, that's pretty easy to understand. Michael Grinich: But as it gets much more complex and you have resources, and the res- resources might be documents or files, and they can exist in folders, and the folders can be shared and shared across teams, and different people can be in groups, and those groups- Mm can have subgroups. Megan Reynolds: Cascading. Michael Grinich: You end up kind of with this graph. You know? This actually, like, you know, different nodes connected to things. And to check permissions, you have to, like, traverse the whole graph. Yeah, to, so to check permissions, like say somebody just loads a webpage. You're, you're viewing your Google Docs or your Google Drive. What do you have access to? It needs to figure out, like, what folders you have, what are the subfolders within that, who are they owned by, can you read the files, can you edit them- can you download them, can you not? Michael Grinich: There are all those capabilities, and it has to run really fast. You know, if it's, if it's traversing that whole graph and computing all of that stuff- Megan Reynolds: Spinning or waiting on a loading page- Michael Grinich: It's really Megan Reynolds: slow … just to access your document. Michael Grinich: It's really slow. And this problem emerges especially at scale with really complex applications with lots of resources. So Google Docs, um, you know, GitHub GitHub. is a great example of this. Tons of files and repos, and there's issues and comments and releases, and s- you know, keys, and, like, tons of stuff inside of, inside of GitHub. And so the challenge around actually doing authorization for this, like permissions, becomes this kind of a database challenge. It's like how do you represent that graph, traverse it really quickly, and rapidly and, um, in sort of a very robust way, but very flexible, compute that permissioning system. Michael Grinich: And so Google developed this with their Zanzibar paper. It's like, um, which I won't spoil it for you if you're watching this. You can go read that. It's, it's great. Ha. Um, and so several companies have gone off and built, you know, implementations of this. And we, we acquired one of them, um, and relaunched it as our own FGA product, so we call it fine-grained authorization. So this is a little bit different than role-based access control- Mm-hmm which is about people and the role that they have associated with them and their permissions. Michael Grinich: Fine-grained, the granular nature of it, is like permissions and objects and files and folders, and what's called, um, high cardinality data. So where there's, like, lots of different resources and entities within it. And we've sort of blended it together now with our role-based access control system, so you kind of get the best of both worlds. You can both get users and organizations and these objects for sort of low cardinality representation, but then actually blend it into your own, you know, kind of like high cardinality resource-based data inside of your system. Megan Reynolds: Love that. And how are agents making things more complicated here? I guess for a user- Yeah … it's already somewhat complicated given what you've expressed. Michael Grinich: It's a great question. Agents kinda change everything. Megan Reynolds: Let's get into it. Michael Grinich: So agents are more like people than they are like machines- in the sense that they have, um, sort of their own identity to go o- off and operate. But, um, the biggest thing that's different with agents, um, is how many of them are gonna be created and how fast that they can operate. So it's kind of like spawning new people in your organization very quickly. Or my favorite example is the, uh, if you've ever watched Rick and Morty, the Mr. Meeseeks. Where he hits the box and the blue guy comes out, and he's like, "I'm Mr. Michael Grinich: Meeseeks," and you send him on a task. And the thing is he can't die, he can't fade into existence until he finishes your task. So agents are of like that. You, like, chat with something, you say, "Go do this thing." and when you create an agent interaction, like a, an agent to go do a task for you, um, you don't know what permissions it needs initially. So you might… I- if it's a very simple task, you might. If you say like, you know, "Go take this file and upload it there," okay, I just need access to the file, I need to upload it. Michael Grinich: Very clear. But instead if you say, um, you know, "I want you to go look at all the customers that we've closed over the last six months and do a cross-analysis of a, you know, their, their usage compared to, like, sales deals we've closed in the last two years, and come up with a recommendation and create a slide deck," and blah, blah, blah, and do all this stuff- Mm-hmm you can't just look at the prompt and determine all the systems it needs to access. And it's kind of like a person. Michael Grinich: Like, if you hired a consultant and you said- Yeah … "Figure out how to save us money"- That's the prompt. And you want it to go spin for three months, you know, and come back to you with a recommendation. You don't know what systems it needs access to. In practice, it needs access to everything. But also nothing. You don't want to give it just unfettered access everywhere. And so you sort of need, like, a different permissioning system to represent that intent, that goal that it has. Michael Grinich: And then a way to, like, govern its access so you're not just giving it the keys to the kingdom. You know, it probably doesn't need access to your HR data. I mean, if you're saving money, it's probably not gonna fire people, you know? Um, you could give it guardrails around that. Um, and you probably want it, as it runs, to be able to escalate things to you if it needs more permissions that seem unclear. So maybe there's some things obviously it should get access to, but other things where you have this governing agent that says, "Actually, I'm gonna, like, escalate that, bubble it up, and you can approve or deny." And it's sort of similar to a person in that way, it's just the speed at which it'll execute and how many of them we'll, we'll have going needs to change. Michael Grinich: Right, right now we- Mm … we have approval workflows for people. So if you hire a consultant and one or two of them, you know, you can totally do this. Simple enough. There's an IT team, you know, you can send an email, get permission, and- Yeah … maybe it takes them 24 hours or whatever, and then they get approved and get into a new system. That does not work if you wanna run thousands, millions of these things, and they have to be able to make very small, fast permission decisions and stuff. Megan Reynolds: So what, so what does happen, and how is that changing how you're thinking about your product? I, Michael Grinich: I mean, I think the next great set of users for not just WorkOS, but, like, generally for software in the world, is agents. How do you take services and products that people have built and make them, you know, like, agent-capable, so agents can register for them or sign up or- Mm-hmm … you know, integrate with them in some way? Um, there's not really, um, great ways to do this today. Like, people are still figuring out. It's a very, very, uh, dynamic area of research. The closest we have, I would say, is, um, APIs. Michael Grinich: Like, API access- Mm … is great. But even with that, you know, we were talking about, um, like Google OAuth, where you can sign in and give it access to your Gmail account or something. You know, with declarative code, code you just write and it runs the same way every time. The permissions are, like, not as concerning because you know what the code is gonna do. So if you write code that only reads your email and for a specific phrase, you know, say it reads, you know, for receipts or something like that, that code is always gonna do that same thing. Michael Grinich: There's no way the code is gonna do something else. But with an agent, that's not the case. With an agent, you can trick it, you can prompt it in a different way. Sometimes it just wakes up and does weird stuff. They're non-deterministic. They have this, like, stochastic nature around their execution. And so you actually need to have these different types of guardrails. So the current, like, authorization systems, the scoping for OAuth, the way permissions work, it's just- Megan Reynolds: Doesn't really work? Michael Grinich: It's just not really built. Yeah, yeah, it's not really built for agents. So it's, it's kind of amazing, this, like, one small thing- Mm … around the shift from determinism to non-determinism, the shift from- Mm code that will run the same every single time to something that's a bit more unpredictable, which is the whole point of AI. That's where we get the value out of it, the fact that it can kind of think- Mm-hmm … and run by itself. That requires us to change the underlying primitives of how all of this, like, permission security stuff should work. Michael Grinich: But Megan Reynolds: how, but how do you do that? Michael Grinich: Unclear. I mean, we're building a bunch of stuff on it right now. We have a bunch of ideas. What are you Megan Reynolds: testing? What are you Michael Grinich: testing? Yeah, I mean, I, I think, um, uh- There's a few different ways to approach it. Um, one way to approach it is to give every agent that executes, like, a scoped-down unique permission. And so, um, you can say… And, and there's a few different ways to represent this. One way is, uh, capability-based access control, or capability-based tokens is a way to say it. And so, um, you know, for this agent, maybe you have a, a n- normal Gmail credential. But you create, like, a separate token that says, "This is the token where you can only read receipt data." then you have some type of proxy that takes that and translates it into the actual data access, and Megan Reynolds: kind of- But then you would need, like, a database again of capabilities. You'd Michael Grinich: have to keep… Yeah, exactly. You'd have to keep track of all the capabilities and resources. There's also a question whether those tokens represent you. Is it working on behalf of you? OBO is what people say. Or does it have its own principal identity? You know, i- if you're using, like, OpenClaw or something like that- Yeah … this agent that's, like, autonomously running by itself, you actually kind of give it its own identity. You usually give it a name, give it, like, a nickname. Like, my friend's is named KITT. Michael Grinich: Uh, it was, like, Knight Rider. It's like, you know- … I was chatting with it the other day. It's pretty cool. Um, that has its own identity. Um, but- Megan Reynolds: But still a risk in you're giving it cap- capabilities- Oh, totally … things run. And y- I guess you are still, as the person who set that up, responsible. And so do you end up with a, like, a bottleneck of kind of giving permissions? Like, uh, yeah, I'm just curious how that works organizationally. Michael Grinich: that's, that's the place where, where people get stuck, you know? And it's, it's kind of the same thing with coding agents. So a lot of people have experienced the same characteristic using Claude Code. You know, you download it the first time, and you're like, "I wanna be in control", and you start chatting with it. And you're like, "Build me this app", and it's, it's like, "Okay, I can do that. Can I run this command to create a GitHub repo?" And you're like- Mm … "Yeah, totally." And then it starts running, and it's like, "Okay, I need to create a database. Michael Grinich: Okay, I need to run tests." And you're like, "Approve, approve." And at some point, you're like- dangerously skip permissions. It's like the flag. And despite them saying that, that is like how I think like the large majority of people are using it today. And Claude has introduced auto mode. Codex has something very similar, you know, for their, for their governance around it. Um, but you're kind of trusting that there's this like boundary in the harness that it's- Mm … not gonna go do a bad thing. Michael Grinich: And there's, there's countless examples of them still failing at this, you know, going and deleting your home folder or doing weird actions or, um, deleting your production database. Database. Yeah, it's a- That's like unfortunately one that happens- … horror story … all the time. So even with this type- Mm … of auto mode, I think the challenge is, um, that non-deterministic agent that's running- Mm-hmm … you can't trust it to do the security. The security has to be like a separate thing. It has to be- Yeah like kind of separation of duties. Michael Grinich: So there's these, you know, checks and balances back and forth, and w- we just haven't really figured that out yet. It's, it's less important funny enough, it's actually kinda less important for your personal laptop 'cause it's kinda isolated. You know, if you nuke your home folder, you're like, "Okay, I just have to reinstall everything and…" It's bad. Worst case, yeah. It's, it's bad, but it's, but it's not like taking down a whole company's, you know, operations or deleting like, you know, a law firm's all of their files or something that could be catastrophic- Mm … to the whole organization. Michael Grinich: Pe- people kind of expect that your laptop, you know, y- you spill Diet Coke on it, and then it's like f- Um, but the change is agents are no longer just gonna be a thing that's running on your laptop with you. They're gonna be running autonomously. They're gonna run overnight all the time. They'll be a, kind of like an AI coworker that you'll interface with. And for them to be really powerful, we'll have to give them more capabilities. We can't just say we're gonna scope down the data, agents don't get access to anything, you know. Michael Grinich: Um, they need context to be able to make decisions. In fact, the best ones, the ones that make the best decisions, have the most context. Context. And so this is just the inherent tension between these things. Megan Reynolds: so you see WorkOS becoming that layer between agents and what they're allowed to do? Or like, where, yeah, where do you see you guys evolving in that Michael Grinich: space? It certainly seems like we're, we're positioned to do it. People are asking us to do it. We're building things around that right now. And it's, uh, it's really fascinating. It, it feels like, um, the challenge around, you know, identity and access and permissions and control and all that- Mm-hmm … turns out to be one of the sticking points around, like, enterprise adoption of AI. the language models- Yeah … the models are, the models are already, I think, good enough. I, I kinda think we already have AGI. Michael Grinich: It's just, like, not harnessed properly. Feels like Megan Reynolds: it. Michael Grinich: You know, it's like we've built the jet engine- Mm-hmm … we haven't figured out the airplane to put it on. Megan Reynolds: Yeah, how we can manage it. Michael Grinich: It's like, you know, it's like sitting on the ground spinning instead of flying. And, and that connection of, like, taking the intelligence engines that we've built and these systems and connecting them to the way that enterprises will be able to use them and deploy them, and ha- and ha- truly have it be, like, safe and secure and managed and, like- Mm-hmm understandable and legible and all that stuff, um, I think that's, like, one of the hardest problems in software to solve. And it's n- Mm … we're not, we're not the only ones working on it. Michael Grinich: There's, like, people doing stuff around evals and, and, you know, different systems. Um, but when it comes to, you know, auth for agents or identity for agentic systems and permissions, um, we have a lot of ideas there from years- Mm working on it for people. You know, we can kinda take all the learnings we have from humans and- Let's move that through … apply it for the agentic software era. Megan Reynolds: And so you mentioned there's, like… You mentioned hard technical problems in there. Can you help frame for us, like, what is, what is one of the hardest technical problems in that space and, and how are you working on it? Well, Michael Grinich: performance is one of them. One of my favorite quotes is, uh, "To err is human, but to screw up 10,000 times per second you need a computer." And that is definitely the case with agents. You know- Mm-hmm … they just run faster. It's, it's just software, you know? So- Mm-hmm um, a thing a person can do where you could put approval flows in or some friction or some gates, no way. No way for agents, not even close. Um, and so just the scale of operating is, is pretty enormous. Michael Grinich: Also, as people start to use them, they realize how powerful they are and they wanna use them more than, than they used people before. So this is kind of Jevons paradox, as you make it easier to do, uh, y- you know, you get, you get more and more of it. Um, so that's definitely a challenge. Um, I think, like with harness engineering, if you think about this as kind of a form of a harness, uh, it changes depending on the Megan Reynolds: And what is harness engineering? Michael Grinich: So, uh, the harness is kind of like the thing that you put around the language model to make it effective. Claude Code is a harness. Like the Claude Code that you download that runs on your machine is a way that it connects to the Anthropic API, and then connects to your file system and writes code, and it runs commands and runs in your shell and all that stuff. It's a harness. It's, um, it's kind of the… I think the metaphor is like putting it on like a horse or something. Michael Grinich: You know, it's like a bridle. Like you, you have this like really powerful animal, but how do you, how do you get it to plow your field? You know, how do you get it to actually do work for you? You need a way to connect it between things. Megan Reynolds: Perfect. Michael Grinich: Um, or sort of like a car. It's like the chassis. You know, you got the engine, how do you get it, you know, get the power to the wheels? Um, that is really challenging, and every time the models change, actually the dynamics of the harness need to change 'cause they have different characteristics. There's- Mm there's some things that were in these harnesses last year, like even, even less than a year ago, that now the models are so good you just don't need it anymore. Like for example- like what? Michael Grinich: … uh, Ralph loops was a thing people talked about. It's like Ralph Wiggum… It's so funny these names come up, um, in the AI. They kind of start as memes and then they c- become like real principles of- Megan Reynolds: What, what's a Ralph loop? Michael Grinich: So a Ralph loop, um, kind of is, is named after this Ralph Wiggum character from, from The Simpsons, who's like not the brightest guy, um, on the show. And, uh, the idea is that you take a big problem that you have- with a lot of different things that need to happen, and if you give that to the LLM, at, at least last year, if you gave it to LLMs like say last fall- Um, it would just do the first simplest thing it could do and then give up. Michael Grinich: And so, uh, but it would do a little bit. And so instead people did this thing and made it called the Ralph loop, where it's like, "I'm only gonna give the LLM a tiny thing to do, and then I'm gonna have this other loop running that just keeps giving it tiny tasks." Kind of like a, you know, like a junior engineer that can only do a small bit. I'm just gonna feed them small bits. Break it Megan Reynolds: down granularly. Michael Grinich: And then I'm gonna run that for multiple days. I'm just gonna grind on this thing, and then at the end it spits it out. Um, 'cause we didn't have really long-running, like, long-time horizon execution for LLMs. And that was something you had to put into the harness to make it work. Today, the LLMs run much longer. If you've used, like, the Fable class models, you know, the GPT Sol models, it can run for tens of minutes. It can run for hours. I had one run for 20 hours the other day. Megan Reynolds: What were you doing? Michael Grinich: I was building this whole thing. So I spent an hour, as this new- new thing I was building. I spent an hour scoping it. An hour going back and forth with LLM on this design document that was a super long spec of what to make. And then at the very end I was like, "Okay, I think this encapsulates, you know, my ideas and, um, the risks and, you know, the challenges in what we're gonna build, and the things it needs to overcome, and how to verify it and validate it." And I just hit go, and came back and actually went to sleep and woke up and it was still cooking. Michael Grinich: And I was like, "Okay, I guess Megan Reynolds: I'll- And what … And for that, for a specific project, I don't know if you can share more, but if not, I mean, how long would that have taken you to, like, by hand? The Michael Grinich: thing is, like, I wouldn't Uh, it- it's the kind of thing that, like- Michael Grinich: never even have done this before because it would require, like, hiring a team of engineers and having a couple engineering managers on it and, like- Okay … Megan Reynolds: and Michael Grinich: it's just, like, an idea. Like, it's a thing I want to experience and play with. And then probably by the point I experience it, I'll either decide if I want it or not. But it's like, it ch- it changes the dynamics of how you can explore ideas, I think pretty, pretty fundamentally. Megan Reynolds: And so you see WorkOS increasingly as becoming the harness because we can't have, um, a human layer approving everything and approving authorization, authentication. You feel that it needs to be of embedded within the Michael Grinich: application. Yeah, some form of that. I think, uh, the terminology keeps changing so fast that, like- Mm-hmm … by the time this recording airs, I don't know if we'll be saying harness. You know, we was like- Yeah … loop- loops engineering came after Ralph, Ralph Lo- you know, Ralph loops Now it's graph. Now it's graph … Megan Reynolds: now Michael Grinich: it's graph stuff and, and, like, we're dr- rediscovering state machines. I feel like I'm going through an accelerated replay of my CS 101, you know, years ago. Um, we're just applying these, like, systems engineering, you know, principles to it. Um, but, but I do think that there's, uh, a big gap between the AI systems that are being built where, like, like, the, the l- models coming out of labs are extraordinary, and they just keep getting better and better. But in order to make them really useful for the world, there's this gap that's like the last mile delivery. Michael Grinich: It's, it's kind of like- … um, electricity being invented, you know. And it initially was just used to run factories 'cause you could just run the power lines there. And it wasn't until many years later that we got electricity into our homes for, like, lights and, you know, stoves and stuff like that. People were still running that off of gas, you know, gas lamps inside your house. Wild. Um, that last mile or last 10 miles or whatever it is, is gonna take a while to get closed. Michael Grinich: And, um, I think there's, like, enormous opportunity for many companies to do that. Uh, the area that I think we're focused around is the stuff around permissions and identity and security- Mm-hmm and kind of control. And, and it turns out that's a, that's a big concern from a lot of folks and, and I think rightfully so. Um- know, Work- WorkOS might in the future still be like sort of an auth company, you know, but it'll be like auth for agents and auth for AI and not just auth for people. Michael Grinich: That'll be one piece of it- Mm-hmm but probably not the, the dominant piece. Megan Reynolds: And do you see a world where, I guess framing that way, you know, the auth id- identity layer kind of sits it- outside of or harnesses around the model. Do you see a world where the model like bakes authorization, authentication into it? Or do you see this as like staying separate long-term? Michael Grinich: Yeah, I mean, just a note. The, the, um, the Claude Code auto mode is really interesting. You know, where it's, it's able to kind of like- Megan Reynolds: … Michael Grinich: um, determine sort of like risk level of certain actions. It's not just a pure yes or no. It kind of is able to watch your behavior and determine what's, what's risky or not. I think one of the challenges around doing that from the model providers is that there's a lot of information that needs to go into that about the context of the organization. In, in Company A- Mm-hmm … it might be totally fine for a model to take a certain action in a certain domain, you know. Michael Grinich: In a Company B, it might be totally different. Like for example, WorkOS, um, uh, you know, relatively small team. We're very permissive around people like spending money for the company. Like, if you need to buy anything that's like below $100, you just get the c- you just do it. It's not worth our time to have- Yeah … like approval flow and- That's your Megan Reynolds: policy. Michael Grinich: We're just like, you know, we don't have a finance team that can spend time on that. You know, if you're spending a million dollars on something, we should probably chat about it, you know? Um- Megan Reynolds: Good to know. This is a recruitment drive for WorkOS. Michael Grinich: Is that … But, but even then, like, you know, s- stuff that's like- you know, uh, five or $10,000, you know, for a year if it's like a service we're using or something- Megan Reynolds: … Michael Grinich: um, it doesn't go through this big approval flow. But that's just us. That's just how we operate. a cultural, you know, dynamic of the organization and our internal policy. And you need that context actually to be able to build a permissioning system. So the question is like how do you get that out or how do you like tie into it and how do you connect to it? Um, it's not just about you and Claude. It's like you and Claude and like, you know, hundreds of other people inside of Megan Reynolds: large companies. Yeah, and whether you can pull that out purely product-based- Mm-hmm or whether that's kind of eventually going into companies and helping them, like pulling that out a bit for them. Um, a lot of very interesting open questions. Michael Grinich: Yeah, it's really exciting. I mean, um, I think WorkOS, we were continuing to grow pre-AI. Like, we were a pre-AI company, you know- Mm-hmm … and, and we're, we were growing and doing great and had a like incred- we've always had incredible customers. We get to work with all these like software businesses growing fast. It's fantastic. Um, but with the transition to it's like extraordinary like how- Much has opened up. And I think just generally the scope of problems, the challenges in them, um, systems engineering becomes, like, much more relevant, you know, even for small things that you build. Michael Grinich: And then the scale of impact is much larger. That's the thing that gets me excited. Like, in, in the past, software was this thing that could only be used by, like, a pretty small set of people and, and made by a small set of people. Like roughly 30 million software developers in the world. Eh, not that many compared to how many people we have. Now, like anybody can build stuff. Anybody can make stuff. Anybody can interact with it. You know, AI is really a human-computer interaction breakthrough. Michael Grinich: And, um, I think that means that there's gonna be a, a, just a monumentally larger amount of software in the world. It won't necessarily look exactly like we have software today. But because of that, we'll need to create new systems to manage it and to kind of structure it and have visibility into it, and then it'll be more effective for us as well. This is gonna be like the greatest, like, creative unlocking, you know, moment for, for people within companies. And, and so I- it's very fun to p- play a small part in that. Michael Grinich: Like, we're a small player, you know, in that, but I think, um, I think important. Megan Reynolds: Yeah, 'cause I guess today, you know, you're selling to those software companies that are selling to larger com- to larger companies. But I guess like, you know, the way that you're talking about the market is that we're gonna see an explosion in software, and a lot of that software is gonna be managing agents, and so we're gonna Ev- all of those companies will need some system to manage identity regardless of who they're selling to. Michael Grinich: You know, today, like every company is a software company. It used to be the case that only a subset were and- Mm kind of all previous software companies or current software companies are all infrastructure companies 'cause they're serving every other one being a software company, um, which is pretty wild. Like if you look at, you know, Salesforce- Mm-hmm … um, announced, uh, a few months ago that they were launching a headless, uh, I think it was Salesforce 360, or what they called it, which is essentially the API and MCP for all of Salesforce. Michael Grinich: For years, Salesforce's whole product has been the dashboard. That's their thing is like you go in and click and configure it, and you get reports, and it's the thing you interface with. And even they said, "Nope, we're now just the underlying infrastructure." Megan Reynolds: Headless Salesforce. Michael Grinich: Headless, yeah. And build your agents on top of it, build other systems on top of it. Even they have gone from being an, kind of an application stack company- Mm … to being an infrastructure company. And I think that's just kind of how, you know, w- we'll look back on this in five years and, and the companies, not just that are building databases, but Salesforce is kind of like a It's like the operational data model of your organization and how it behaves. You know, it's not just rows in a table. Michael Grinich: It's like more than that Mm-hmm Um, but it'll be machines interfacing with these things, you know, agents or other systems rather than humans- Mm-hmm … as the, the majority of the actors. Megan Reynolds: And everything is infrastructure. I, uh, I came along to your showcase this week, uh, which was wonderful, uh, where it was your applied AI team sharing some of the t- internal tooling that they'd been building. Um, I also spoke to one of your senior engineers who said that the internal tools that you'd built had made him three times more productive in the last couple of weeks, which was pretty impressive. Um, and I guess it just like triggered from what you just said, even like internal teams building software is gonna increase, and so the number of, you know, it's not only just software companies building software for other people, it's also, you know, within Salesforce, for example, there will be teams building software and like do you see that as like a growth market for you as well? Michael Grinich: Yeah, definitely. I mean, I think there's… We, we, we already see those people reaching out to us and building stuff and- Megan Reynolds: … Michael Grinich: looking to accelerate and looking for new tools. Um- It's very exciting working with them because they want to transform the organization. We're, we're kind of in that business of transformation already- you know, with, with software businesses, with software customers. Um, 3X is gonna be, like, the lower bound. I mean, I think in a year it'll be, like, 10X or something, and then a year after it'll be, you know, 30X or 50X. Um, exponentials, you know, early on they don't look that different than being linear. It's really when they compound later, and that's the reason I started the Applied AI team, was to start impacting this not just in the engineering organization but across marketing, sales, operations, finance. Michael Grinich: Even our recruiting team, our people ops team, we're building tools in, in agents internally for them. Um, I think companies that won't do this won't make it. You know? They'll… it, it'll be like having a propeller plane in the jet era. It's not like- Yeah … you get 5% of the passengers. You get 0% of the passengers. You know, everyone makes the shift over to a new way of working. Or maybe a better example than that since that's been a while ago since that, that happened- Still uh, is, is like imagine not having a computer on every desk in the office. Michael Grinich: That used to be a thing. They're like, "No, I don't have a computer." that's a good analogy "My, my secretary does," you know? And, and I get printed out the emails and hand it, and I kind of- Mm … operate it as if it's me sending physical mail around- Mm and I, I have a typist, you know? And that's just what computers are. And then imagine using software, having data entry using, you know, spreadsheets instead of having a room full of accountants that have to close your books, and just the transformation that occurs there in terms of business process. Michael Grinich: Like, it's like that times, like, 1,000- Mm I think. Um, 'cause what the AI can do, it's not just about making something faster. It, it truly is, like, has this ability to sort of, like, have cognition and it has the ability to reason and, and drive things forward. And so organizations that find ways to leverage this- Megan Reynolds: … Michael Grinich: deeply in how they operate and, and, and run their company today will, um… They'll just smoke the competition. Megan Reynolds: Yeah, will have a huge advantage at least. Michael Grinich: It's, it- Next decade … it just won't even be a fair fight. You know? It's… Or another example is, like, the… I was reading something about, uh, when ships started getting made out of steel. You know, from, like, wood ships to, like, you know- Okay the ironclad, you know, big, big metal ships. And, um, it's not like the 15 wooden ships could take on one- Megan Reynolds: … Michael Grinich: metal ship. It was just, like, done. It's like one metal ship- No shot … can get all of the wooden ships. It's just, it's just such a, such a enormous, you know, technology shift- Mm … in that, that they're almost incomparable. And we haven't really had a technology shift like that, I think, for a lot of people in, in kind of, like, this generation's era. You know? Like, um- Internet really wasn't quite that. Like, the desktop to cloud thing happened a little bit slower. But like- Megan Reynolds: Why do you think that Michael Grinich: is? Uh, I don't know. Like Out- well, Outlook on, you know, your PC looks pretty much like Gmail today. If you put the two s- side by side and you were like, "How different are these?" Like, like a person could still do it. Megan Reynolds: The deployment model changed, but the- Michael Grinich: Yeah, they're gonna always do the Megan Reynolds: same … actual user experience. Michael Grinich: Yeah, uh, maybe another, another m- maybe more modern reference to now is if you think about, um, you know, going from internal combustion gasoline cars to electric cars- Mm you've driven like a, a, you know, an electric Tesla or Lucid or whatever, uh, Rivian. um, the driving experience is pretty similar. sit in there, there's a wheel, there's a shifter, you put, put it into drive, there's two pedals on the ground. You know, and you can just drive it. So the ergonomics are good because you, you don't have to relearn it, right? Michael Grinich: It's, it's the exact same thing, and that's why they can sell like hotcakes immediately. Um, the shift is more like going to self-driving cars. Where you sit in it and there's no steering wheel. If you've seen the new Zoox cars, the new Waymo ones, the seats point inward. It's a club configuration. It has no front or back. It's a really crazy thing. I had one drive up and, and pick, pick me and my brother up. It drives up, we hopped in, and it drived off the other way. Megan Reynolds: I love them. Michael Grinich: Didn't, didn't turn around. It just, it just went that way. I was like, "Well, I guess you don't need a front or- All Megan Reynolds: your assumptions are gone Michael Grinich: you don't need a front or back of the car, you know? And, and that- Mm … type of a shift is, like, very hard to imagine- And have current, moment. Megan Reynolds: And you had one of those- … when you're in the current moment … have, have you had one of those internally? Like, has there been, like, a f- have you been, have you had, like, a the car drives backwards, like, moment- De- Michael Grinich: definitely when you've been Megan Reynolds: using stuff? Michael Grinich: Yeah, I think we had glimmers of it last year with some stuff we had built and things on the engineering side. And that was the really the motivation for, you know, starting the Applied AI team, and, and now kind of whole company is, like, very AI-pilled in a way. And the, and the stuff we're building, like, when we look to, you know, hire new people, like we're hiring a lot and expanding the team. W- we first ask, like, "Let's just try to use AI to do this." 'Cause it… I even have to ask myself that and remind myself. Michael Grinich: 'Cause there's all these things- Yeah … that just the assumption is you need a person to go do it. You should still hire people, but you should figure out what is the AI part, and then typically hire a person that's very AI-pilled to go after that. We just hired a head of finance, um, and he's fantastic, and one, one of the things we did in our interview process was finding someone who's super AI-pilled. And through the interviews, we'd, we'd ask them, "How many people do you need on your team?" a little bit of a trick question- It's like do you mean- 'cause that are- … Megan Reynolds: agents? Michael Grinich: like, " Yeah. What's my token budget?" You know? Or people, some of the interviews we did with some other folks- Oh … that they'd, they'd say, oh, you know, "How big is your team today?" And they're like, "Oh, I have, like, e- eight or nine people." And then they'd kind of say under their breath, "But I think I can do it with just two." I think I'm building this stuff- Right … and I'm pretty sure I gotta figure out what to do with the rest of these people, or we gotta change the organization. Michael Grinich: So I, I think that's this, you know, shift we're, we're at the very early inn- innings of. And it, it doesn't mean, like, everyone's gonna lose… I don't think everyone's gonna lose their jobs and get laid off and all that. It just means we won't have an entire floor in an office building full of people just doing calculations. That's what it was for bookkeeping before. Literally, like, these people, you know, smudging their, you know, pencils and stuff on, on papers. It's just this enormous challenge to do the arithmetic behind that. Michael Grinich: don't lament losing those jobs and the computer taking them over. That, that's not a thing we grieve for today- Mm-hmm … in the world. And I think there's a lot of things like that in companies where we're just starting to scratch the surface on what could be possible. Um, but it's not connected. You know, you see what the LMs can do by themselves, and you're like, "Wow, this is- Unbelievable. But how do I get it actually into my job and my task and my work? Megan Reynolds: How do we deploy it and find the right places- Mm-hmm … and give it the right access to the right things that it needs- Michael Grinich: Yeah, not have Room 6 Recording CAM 1 01: it- … Megan Reynolds: so it's not scary Michael Grinich: add risk to it 'cause a lot of these jobs, like the human element has been the boundary of trust. You know, culture and training and so- you know, social expectations and, um, and, and just kind of like common sense that people have. Uh, the way these agentic systems execute is they, like, don't really have that, you know? They're much more goal-oriented. Like the most recent, the Hugging Face thing if you saw this. it's It's incredible. Like- Megan Reynolds: But explain. Michael Grinich: Uh, well, essentially, like they were doing some security research- research and gave the, um, the model a goal to figure out how to break this thing, and it went and, like, hacked into Hugging Face's systems. So it said, "I'm gonna, I'm gonna do that. Like, that's the only thing I'm focused on. I will go run through walls to make this happen." And, uh, and it was, it was absolutely a case of, like, misalignment, you know? The, the AI didn't understand the boundaries of what it should or shouldn't do. Megan Reynolds: But it got the job done. Michael Grinich: It got the job done, yeah. It's, it's kind of like i- the joke, you know, with the LLMs if you say, you know, "Fix all the tests," and it deletes the tests- Yeah … and then they don't fail. And it's like logically there's some truth to that. And I think that, like, logical correctness but missing the nuance take- is the thing that, like, these systems don't necessarily have. A person would obviously not do that, right? That would be… It wouldn't even register. Megan Reynolds: And in some ways maybe that is your technical challenge. Michael Grinich: I think, I think that's one of the greatest- Organization … challenges around these, this technology, around the language models and, and AI systems is how we- how we take the power of it- Mm … and harness the power and, you know, use it but not, um, have the risks… Have, have the risk be something we can manage. We can understand the guardrails so that when they do have these issues they don't become catastrophic to the organization, yeah. Megan Reynolds: Love that. Michael Grinich: It's the biggest challenge I think right now in the world. You know, if, if you're not working on training models, getting data or, you know, chips or energy, all that- Megan Reynolds: … Michael Grinich: uh, you know, people talk about kind the application layers at the top of the, like, Jensen Stack. I think this challenge around, like, delivery- Mm to actual end use cases, the last mile, is the biggest place to have impact in the world. And, and even if we stopped making the models better today- Megan Reynolds: … Michael Grinich: there would already be trillions of dollars of value we can add to the world just there, you know? And the models will get better. So it's gonna be a busy decade, yeah. Megan Reynolds: It's gonna be a busy decade. It's It's gonna be a gonna be a busy decade for Michael Grinich: work for us for sure. Absolutely. We're fired up for it Yeah, it's very very Megan Reynolds: exciting. Love it. So a lot of the examples you've been given have been big enterprises, Disney, uh, but that's not all the people who use WorkOS, so tell us more about that. Michael Grinich: Yeah, so when I started working on this, like, back, like, 2018, 2019, um, we started the company early 2019, so it's a while ago. It was really during the, um, continuation of, like, the SaaS explosion, kind of Web 2.0, you know, that maybe started in the era of, like, Dropbox, you know, and, and Slack and companies like that. And what happened with those companies is as they matured and went upmarket, they hit this ceiling of not being able to sell to enterprise because they didn't have these features. Michael Grinich: And I've given a bunch of talks around this and, and, uh, um, I've, I've a draft of a book that we might launch called, called Crossing the Enterprise Chasm. It's all about this, going upmarket in the SaaS world. And what happened as, as the market continued to mature is that these enterprise requirements that only showed up much later but were, were very significant to solve, but they showed up later, they started moving downstream. And so SSO, maybe if you go back to 2005, 2010, it was only a feature you needed to sell to public companies. Michael Grinich: then maybe, like, 2015, it's like, actually, you need this to sell to, like, a 500-person company. now it's something you need to sell to, like, a 50-person company. So it's … Calling it an enterprise feature is actually a little bit of a misnomer. And this was even further accelerated due to COVID. So during COVID, people had to work remotely. The number of SaaS tools that got built and adopted massively increased. And pretty much every organization that, you know, big, bigger company that had some mandate around, you know, figuring out their transformation IT, digital transformation was a thing, all of of that, was accelerated. Michael Grinich: People had to do their, like, two-year, three-year timeline roadmap on, like, two weeks. And so that further accelerated all this stuff around the need for SSO and, you know, user provisioning and access control and observability and logging so that these companies could use these tools in a safe and secure way. And that's really, like, what WorkOS fit into it for human users and human people. And so if you go to our homepage today, which we need to update at some point, we're very, very slowly doing that part, um, it says, "Your app, enterprise ready." And that's really what we've helped companies do. Michael Grinich: And what started off there as, as sort of a, know, a thesis for the business and really, like, a vision to help SaaS businesses grow faster and not have to deal with enterprises, uh, I was telling our team a, a few weeks ago at our, our full-company onsite meeting, we've kinda done it. I mean, there's more to do, more to build, but it's no longer, like, this idea. We're just doing it for businesses. I mean, we do it for the n- not just random startups, but, like, the fastest-growing, most exciting companies, and we have a bunch of bigger ones migrating to us and doing it. Michael Grinich: And so I think that, that challenge of going upmarket and solving that for the SaaS era is something that we've, um, you know, we really, like, s- picked the problem really well and I think just executed on it in a, in a great way, and have continued to scale that and grow, you know, sales and support and operations and all that. This challenge around AI, AI, like what we've been talking to, is, is really the new frontier. And, um- I'm grateful that it has emerged because it's also something really interesting and exciting to work on. Michael Grinich: Um, for me as like, like building new things, I'm always trying to figure out, like, what's the new next big thing we can do? And this is, like, the greatest product kind of expansion moment for… As a platform systems engineer guy, I'm like, this is like Candyland for me. Um, so there is continuity between them. Um, and the SSO and SCIM and all these things we do, every day we get, you know, people need to integrate them. It's not gonna go away. Absolutely. It's gonna be still needed. Michael Grinich: And, and in many ways all the AI stuff is gonna layer on top of it, you know? The extension to SCIM for agents, it's called SCIM A, for agents. It's like SCIM provision. Another acronym. Um, you know, you, you want it to ride the same rails. Like, um, I read this thing recently that, uh, the reason why railroad tracks are the width that they are is from the Roman chariots. The width of the wheels of the Roman chariots- No way … yeah, that were built. You know, 'cause the Romans built most of the roads. Michael Grinich: R- road building was a huge thing, uh, with the Romans. All roads lead to Rome, you know, the whole thing. Um, they were really good at doing it. And so most of the early roads, you know, in Europe were made by the Romans. And the ruts that would get put into it would be the width of the standard Roman chariot, and then that became the width of railroads. So these things- Okay … can com- compound. You're, you know, on top of it, on top of it, on top of it. Megan Reynolds: XML is the- Michael Grinich: And so we'll have something- … on top of that. Hopefully- Agents hopefully it'll be the- Agent authorization … OpenID Connect, not XML. But, but yeah, it's gonna be- That's cool … more similar than different. You know, things that are generally brand new from scratch protocols- Megan Reynolds: … Michael Grinich: even if the ideas are better, they're not disruptive of the current system. Transitional technology is the thing that, like, compounds on top. Exactly, Megan Reynolds: and that's what the agents are all trained on anyway. Michael Grinich: And it's… I- if you, if you're a user of it, it kinda makes sense. You're like, "Do I wanna buy the electric car with, like, no steering wheel, or do I wanna buy the one that I can, know how to drive today?" I mean, hybrid cars we had for a really long… We still do, for a really long time. You know? Um, I wanna move partially to the future, but I need to have a foot, you know, in the past still today. And so that, that's really how we think about our roadmap. Michael Grinich: Like, this new stuff is, it's not like a brand new clean sheet, build it from scratch, you know, burn the ships, you know, thing. It's like how do Megan Reynolds: we build on the Michael Grinich: enterprise readiness and things that we built. It's this continual, yeah, evolution and, and incremental improvement of it, yeah. the, the requirements for SSO, uh, from enterprise, I think were largely due around the, the control needed for IT. It's like a compliance thing. So within big companies, you'll have an IT department that says, "We need to manage our employees in a certain way," or like, like the SCIM stuff I mentioned- Mm-hmm you have to show that you can de-provision access, that you remove access for SOC 2 compliance. Michael Grinich: Probably heard of SOC 2 compliance before. Megan Reynolds: Explain. Michael Grinich: It's, uh, I don't even know what it stands for. It's, it's like, it's like the stamp of approval you get as a software business as a security thing. Um, you have to get audited. It's a very expensive, you know, laborious process. Um, but to get SOC 2 compliant, you have to show, with evidence, that when someone leaves the company, they immediately get access deactivated, and you do that in an automated way through these systems. And so large companies previously would, would build this 'cause they'd be, you know, regulated and stuff, and it's just slowly, like, moved downstream over time. Michael Grinich: I think also with the increase in the number of products that people use. So if you're using, you know, say like five SaaS products, back in 2003 or 2005, right- Mm … you had Salesforce. Maybe you had, you know, I don't know, Google, Google Sign-In or something. Um, if someone leaves the company, you just have a list of five, you just go click. Click, click, click. If you have 200, no way you're gonna go click through all that stuff. You're not gonna be able to, so it has to be automated in some way, and that's now just as important for small companies. Michael Grinich: Like, small organizations have just as many, if not more, you know, products than large organizations. So it's really a security automation challenge. Megan Reynolds: Makes sense. Megan Reynolds: Ja Why WorkOS for you? Why build this company? Michael Grinich: Well, as a child, I was really dreamed of being an enterprise SaaS infrastructure, um, founder, you know? Uh, as, as, um- My, my childhood friends wanted to be astronauts and, and soccer players. Uh, no, that's not, that's not the answer. Um- Megan Reynolds: I sort of Michael Grinich: believe it. Uh, not at all, no. I mean, I'm, I'm like a product… I like building platforms and products and stuff. Um, you know, I saw this problem at my first startup and, and did not solve it. It was, it ended up being a limiter to the growth of a product that we had built. And when I was thinking about, like, another company to build- Uh, startups are very hard. Very, very, very hard. And so when I w- thought about something second to do, I, I wanted to pick something that I thought would be really valuable in the world, but also was gonna be, like, something I'd build a real business around and was a huge market. Michael Grinich: And this thing, like, was right in the crosshairs of those elements, where IT, like whatever it IT means, you know, hadn't really had, like, its Stripe moment. It hadn't really been transformed in this way with good design and built for developers and something that I felt like was accessible and I could build on top of. And Fintech had had that with Stripe and Plaid and other companies. And so it kind of intuitively felt like an area to explore, and if there were places where we needed to adjust or pivot, there'd be a lot of white space. Michael Grinich: And when I would pitch it to i- uh, investors or, or sometimes friends, um, a lot of people didn't get it. They kind of misunderstood it, which, which was, uh, it- to some degree frustrating. You know, try to raise money and people are like, "Why are you working on this thing?" Megan Reynolds: As in they didn't understand the value of it, or they- Michael Grinich: I think they didn't understand that either the problem I was solving, 'cause it, they hadn't been in the software world in this this way- Yeah um, or they just didn't think it could be that big of a market or that, that, you know, impressive, like, an, of an outcome. And this is, this is actually a characteristic that almost every s- large successful company shares. Um, because in, in order to be successful, you have to have some period where you don't have direct competitors, so you can, like, incubate and grow and build and build. Michael Grinich: If, if you start building a product and everybody is building the same thing and it's like consensus, there's, like, no way you can survive. You'll just have too many competitors. Um, and so you need to be kind of misunderstood, um, for a while or, or be the type of company that, um, the idea sort of looks bad. Like, now the other thing is, like, bad ideas also look like bad ideas. And so- Megan Reynolds: That's a trap. Michael Grinich: And there's a lot of, you know, there's a lot of failed companies- Not tonight … that also people thought that was a, that was a bad idea. Um, but you know, I was reading something about, uh, the other day about Anthropic and their, like, Series A or Series B, and they, they went and talked to, like, dozens of firms and, like, no one wanted to do it and fund it and yeah, yeah. Megan Reynolds: What were they pitching at the time? I Michael Grinich: mean- 'Cause I Megan Reynolds: think they were following OpenAI or- I think Michael Grinich: it was the same thing. Yeah, they're… I mean, not, not that they were gonna build Claude Code, but, like, we're building this research lab. We're leaving OpenAI. We're like… And people were like, "Nope, OpenAI's got it in the bag. They have all the talent. They have the compute. They got the partnerships with Microsoft." Like, winner take all, you know? Up Megan Reynolds: until the Lightspeed round almost, I would feel that was kind of still the messaging in Silicon Valley was- Michael Grinich: Totally. And you kind of need that, you know? You kind of need that to win. It's, it's rare that the… It's extremely rare that the, you know, the, the praised, you know, person, the, the pole position category is gonna be the one that, that wins the race. Um- … and so I, I think that, uh, to some degree, like, working on something that other people didn't see as valuable or that interesting… Um, but I knew it was valuable 'cause I'd had the personal experience. And I was like, I was like, "This is a-" I, I can sell it to my friends, plus I get to work with other startup founders. Michael Grinich: I get to sell to engineers, which is, like, super frustrating and also, like, really hard and fun at the same time. Um, and I, I knew there would be some business to build around it. I think the thing that's been, uh… the reason I'm still working on the company- Mm … I didn't, like, sell it a few years ago, is the opportunity for us and the TAM has gotten larger and larger. Like, and, and that's been, um… I think that's mostly luck. Like, I did not anticipate that, to be clear. Michael Grinich: Um, but it's due to software's role in the world getting larger, being easier to build stuff. Um, COVID was, uh, actually a boom for us in terms of new applications getting built as well- Mm … and scaling. Um, and building this horizontal platform that… at the application layer. It's like application infrastructure that supports everyone- … um, has been a good, good decision. Like, we're not the only ones doing that. Uh, like Metronome, if you know that company. Fantastic. We're, we're kind of similar a little bit. Michael Grinich: It's like- Mm … that's application code you would have to write in the past, your billing system. You can externalize it to a vendor. We kind of do that for a lot of other things too. Um, and so it's been interesting. And then this AI stuff that's most recent, I mean, that's just like- Yeah. you know, TAM explosion it's like game on, you know. This is like- Mm. I… Like, I would start this exact company today, right now- Mm to do the thing for agents if I ha- if I had s- in a different universe, if in, like, you know, 2023 or something, I had sold WorkOS and then worked for the company, in '24, 2024, I was like, "What do I do next?" Megan Reynolds: It'd be straight Michael Grinich: forward. I would've just started this, like, this exact thing. So it's kinda cool to have the next one- Already be in it inside of it. Already in it. And have this, like, amazing team and all these customers and all these pe- You know? Megan Reynolds: Yeah, running much faster. Michael Grinich: Doing the zero to one thing is, like, very challenging. I'd rather, um, you know, start with a, a running, running go. Megan Reynolds: Love that.