Okta, Inc. (OKTA) Earnings Call Transcript & Summary

Dave. So Ely, Harish, Dave, thanks for being here, Janice as well. I very much appreciate you all being here to hopefully get a technical session for people. And go into a topic that I think is very interesting. So Ely, the Chief Product Officer; Harish, SVP and GM of AI Security. And then we all know Dave and Janice on the IR side of things. So I'm going to throw it over to Dave to hit the safe harbor, and then Ely and Harish have some prepared slides and topics to hit. So I'll throw it over to you, Dave.

Great. Thanks, Eric, and thanks for hosting us today. Really appreciate it. And just Janice is going to throw the safe harbor up. You guys are all familiar with the safe harbor language, just a reminder of that. And just to remind you that today's conversation is really about the technology. We're not going to be taking financially related questions today. So with that, I'll throw it over to Ely.

Thanks so much. And once again, my name is Ely Kahn, Chief Product Officer here at Okta. I've been here about almost 6 months, but been in cybersecurity my whole career, first inside the U.S. government, U.S. intelligence community, doing cybersecurity, then a startup called Squirl that was acquired by AWS, spent 4.5 years there building products. Most recently before Okta, I was Chief Product Officer of SentinelOne. And yes, excited to be here to talk a bit about our vision, our road map and our strategy. Okay. So just as a refresher, I thought I'd start with a reminder about what is the Okta platform. And really over the last 3 to 4 years, we've been building out a complete identity platform that covers all the major use cases that you need to protect your identities, preauthorization, at authorization and then post authorization. So maybe first starting with the preauthorization focused products. We have our identity security posture management products. This will be an acquisition that we did 3 years ago, a company called Spara. This gives us the ability to discover misconfigurations, identity-related misconfigurations in hundreds of your SaaS and other types of applications. So generally speaking, we're looking for overprivileged or misprivilege administrators or other identities, including administrators that haven't been -- accounts that have been used in a period of time or other examples of misconfigurations that increase your blast radius. Next up, we have our identity governance products. This has become one of the fastest-growing products in our portfolio. And it just recently has crossed over 2,000 customers, which puts us on par now in terms of customer adoption with other IGA products like SailPoint. And so this gives us the ability to ensure that those users that you've onboarded to Okta have lease privilege, allows us to access requests and access certifications to ensure that those users continue to have the right amount of privileges over time. This is also complemented with Okta Privileged Access, which I think about as the flip side of the coin to our IGA products. With Okta Identity Governance, you're provisioning access to sort of standard sorts of applications. And with our Okta Privilege Access, you're provisioning access to high risk, high trust resources, things like databases or Amazon EC2 instances. When we do that, we can do session recording. We can also plug into on-prem resources, ensure that even if you're using a long-lived credential or API access key, we then short-term federal tokens to ensure that even if there is a compromise, the blast radius is small. Next up, at authorization time, we have a couple of products. Of course, our flagship [indiscernible] product around access management, which provides secure cross-platforms fishing resistance authentication. And then we sync that up with Okta device access so that you can use the same login processes both into your device and into your applications. And then lastly, Identity threat protection. This is our Identity threat detection response product. This gives us real-time signals that can be used to look for misbehaviors and anomalies during the session. And the real power here is these aren't producing just another set of alerts. We can tie this into other products so that we see users above a certain risk threshold. We can do things like universal logout or step up authentication to really contain that risk within that session. Now these products, they're not siloed. The power here is that they're part of a unified identity security fabric and they allow us to cover a broad set of users and devices, including employees, partners, contractors. It can plug into both modern SaaS applications and legacy on-prem infrastructure. And now most recently, we've extended the platform to support AI agents as a first-class identity in the platform. I'll get more into that, and Harish is going to go real deep into that. In terms of that Unified Identity Security Fabric, let me give you just one example. So let's say, as part of our ISPM, our Identity Security Posture Management product, we've identified that a particular administrator is not being used. That account is not being used. We can use that information then kick off an access certification in our IGA product to officially recertify whether that user needs that administrator account? And if not, revoke. So we have a bunch of out-of-the-box integrations across our products. We also have a low-code, no-code solution called Okta Workflows that lets you stitch together any other type of custom integration between the products that you might want. And then lastly, and certainly not least important is our uptime story. So we're at 4 9s as our official SLA. This year, we're actually at 100% uptime. And over the last 3 years, this means that we've had roughly 69 minutes of downtime. And compare that to Microsoft that's had over 2,000 minutes of report uptime. This is a key differentiator for our platform and one of the reasons customers love us. Next up. Okay. So a big move for Okta has been that we've been really pushing towards this identity platformization story. And we've been working with some of our largest most complex customers on this story where they're making a big bet, and they want to take a very complex set of legacy solutions and simplify. So this is a customer that we've been working with for the last year. We'll complete this platformization story this year. And they're very typical of a large Fortune 2000 company. In this case, they've used a legacy solution for identity management. I think they are using SailPoint here. And they're using another vendor for their access management, in this case, Ping and then another vendor for privileged access management, in this case, CyberArk. And they have specialized teams for each of these vendor solutions. And -- it's not only a cost issue for them, but it's a complexity issue. And that complexity also translates to security, meaning the more complex your stack is, the more likely there are gaps in which adversaries can compromise. And so they wanted to solidify down into a single vendor, Okta, not only to reduce costs, not only to reduce complexity, but also to increase their overall security posture. And we're well on the way to this. We'll finish this complete replatformization this year, and the customer will be seeing significant -- they're estimating somewhere around 40% economies of scope gain in the form of cost savings. All right. So maybe backing out now in terms of our 2026 product strategy. And this closely relates to that last slide there in terms of the platformization play that we're making. So we have 5 main elements to our product strategy. These are not the only things that we're working on, but these are the top 5 most important things that we're putting our efforts around. The first one is our Okta for AI agent story and ensuring that we have a first-class set of experiences across all of our products to support AI agents and to protect the identities of AI agents. I'll get more into that in just a second. Next up, second and third, maturing our IGA and PAM solutions to be world class so that they can [indiscernible] replace legacy vendors like a sell point or like a CyberArk and even just win head-to-head in greenfield competitions where the customer may not even be using us for access management. And we're well on our way to that as well this year, and we have a number of wins where that happens. Identity Security Fabric. I'll actually talk about Identity Security Fabric a little more detail in the next slide through the lens of Okta for AI agents, and how all of our products are coming together to provide that unified Identity Security Fabric for the AI agent use case. And then lastly, Okta, like a number of companies is all in on AI native product development. And we are aggressively using coding agents across all of our product teams and seeing 50% gains in product life cycle development which means we're shipping products and features faster to our customers while still maintaining the really high bar around uptime, resiliency and security. All right. Let's jump into the AI agent story here quickly. I'm not going to go through this in detail, but I saw there were some questions and the pre reads about what type of agents does Okta protect. This is a common question, and we actually built out an AI agent taxonomy to help with this. At a high level, the 4 major categories of agents as we think about them are embedded agents. These are things like your Zoom companion agents. Now embedded agents are the one type of agent that we do not try to protect. These agents are essentially hard-coded into SaaS applications. They don't expose APIs, so we can't really manage them. Next up, stand-alone SaaS agents. So this would be something like a Devin or a [indiscernible]. So this is an AI agent that exists as a SaaS agent, we can onboard these types of agents into our platform to govern and protect them, agents associated with automation platforms, where the agent built into a boom or a torque or times. That's also a type of agent that we're working to protect. Homegrown agents as well. So if you built out a fully custom agent using plan chain or an agent builder platform like Amazon Bedrock, those are also agents to support. The other dimensions of agents that we oftentimes look at and have some unique characteristics in terms of how we protect them is the delegation pattern. So is this simply a human prompting agent? Or is there a set of subagents that are spawning. And then how long is that agent exists? Is it a very short-term ephemeral agent or a more lasting one. These are all considerations that we've gone through as we built out our Okta for AI agents platform. Let me talk about that quickly in the next slide. Okay. So this is a busy slide here, but let me talk through it real quickly, and then Harish is going to, I think, touch on some of the details here in more depth. So when we talk about our Okta for AI agent story, at the top level, we think about it through 3 simple questions. Where are my agents, what can they connect to you and then ultimately, what can they do? And so what you're seeing here is our reference architecture. Most of these things are in place, not all of them yet, but these are all the things that we're building towards really over the next -- either we have now or building over the next 6 to 12 months. And it starts with discovering your agents. And so we built integrations with a number of different tools to directly import agents from those tools. This could be things like Amazon Bedrock or our sales force control tower. Any of those types of agents we can directly embed them into our registry through a direct import by calling their APIs. Also if you build a homegrown agent, you can use our SDK to then import that agent directly into our registry. But we also have the ability to discover unknown agents. And the first capability that we launched earlier this year is using our browser extension to discover all of claims between users and agentic tooling and then giving the users a workflow or the administrators the workflow to register those unknown agents in our registry. Once they're in our registry, then they can be registered with a human owner or they can be registered as a fully autonomous agents. You can apply additional meta data to them to ensure they have a full identity and they can come under user entitlement reviews. So is this a user that should have access to that agent and doing access requests and access certifications. Also, does this agent have the right access to resources and able to do access certifications over time on that. From there, once that agent is registered in our agent registry with a first-class identity, you can start giving it permissions in the form of scopes. So these are considered core screened permissions. And so you might give it sort of provisions to Slack or to GitHub or to Jira in terms of being able to read versus right, versus update, et cetera. And from there -- and this is really important because you want to give those agents the least amount of privilege possible, but enough privilege to fulfill their goals and objectives. Now we've also shown some -- we put some blocks on this recently, how you can apply fine grain access controls to agents. So not only can you apply these core [indiscernible] grant scopes. But if you also want to give it a more fine-grain access policy like say, "Hey, this agent can only view not sensitive data or they can't write sensitive data." Those would be additional fine grain access controls. This is where we bring in some of our of 0 capabilities in its FGA products to combine with our Okta for AI agents solution. All right. So once you've defined the access policies, then it can start accessing resources can access resources through our MCP servers, authorization servers, our SaaS authorization servers. We also have a new agent to agent capability and ultimately, we are able to monitor these capabilities, produce logs that can be shipped to a SIM. And perhaps most importantly, if things start going wrong, if that agent starts misbehaving, acting suspiciously, we also give you a kill switch that allows you to revoke tokens. -- and ensure that those agents cannot continue accessing resources in a malicious way. So like I said, there's a mix here of things that we've delivered and things that we're working on. But this is really the complete story that we're working towards and will ultimately help customers not only identify their agents but ensure those agents operate with lease privilege and have a small blast radius if they're compromised. All right. So with that, I'll hand over the mic to Harish.

Yes. Yes. Thank you so much, Ely. And I see some questions trickling in. So I will keep this brief, so we can make sure everyone's questions going to address because that is what is most critical. For introduction, I am Harish Peri. I run the AI business at Okta. So I spent majority of my time with our customers in the field, making sure our everything to do with our AI security story resonates and that we're growing this business. So that is why I'm here. So some of the content I have is more around our positioning our capabilities and how we tell the story to the broader market, which I think will also be very helpful to this audience. So if you go to the next slide, the goal that we have for every one of our customers and folks that are not our customers that are embarking on their agentic journey is every company will become an agentic enterprise in the next year, 1.5 years. What that means is many pieces of their product development, portions of how their employees interact with their tools, how they connect with their end customers, how they connect with our partners, suppliers; all those are going to become agentified. There's going to be agents in the flow there. Our appeal to all of them is you could be anagenic enterprise or you could be a secure agentic enterprise, which means you have a centralized security offering to be able to control the identity and access and authorization of everything those agents are connecting to. So that's the broader kind of hook that we have for the market, which is to say, how do you not just be an agentic enterprise, but how do you do it in a way that's secure? Because the difference between building agents and rolling them out into production is security and scalability concerns. So that's really how we're starting our position. We go to the next slide, please. And since we've embarked on this journey, we've engaged with hundreds and hundreds of customers. I mean we're talking CISOs, CIOs across some of the largest companies in the world. And the thing is the tongue in cheek way of putting this is the cat is out of the bag, everyone has the agents they're rolling out. The question is, where does the cat go, right? And that gets into something -- some of what Ely was touching on, which is discovery is a really huge issue. But the common themes are, we are -- everyone in our company is going to become a developer at some point. How do I know that they're developing within the right guard rails. How do I ensure that all of my Claude code or my open AI or my cursors or all those development assistants are actually being control in the right way and how do I ensure that they don't execute judgment and delete something important overnight with no human oversight, right? So this problem is very real. It is very top of mind for pretty much every one of our customers and every new company that we're talking to. So it is very much in the zygus, the topic of securing the agents. If you go to the next slide, please.. Now I want to I want to spend a little bit of time on maybe some broader topics, which is -- I actually spend a ton of my time educating our customers about just the nature of the overall problem. And one of the topics I get pulled into a lot is, "Well, I have an AI gateway. I have guardrails around my LLMs." And to that, my response is that is necessary, but not sufficient. When you think about what it takes to secure agents, you really have to think about models, which is, of course, the brain behind the agent that is making decisions in response to a prompt and deciding what tools it can be accessed. But there's a second part of it, which is how do I actually secure the identity and the authorization of what that agent is actually accessing. And that's a lot of what Ely was touching on. And so model is one piece, identity is another piece and the third piece is the data itself. So for those agents that are accessing your data at large scale, either for [indiscernible] generation or as part of what their -- just their daily operations also getting down to a level of is this agent authorized to access this data at this moment in time given all environmental signals is a third big piece. So we like to think about model identity and data and where Okta starting right now is this notion of how do we control the identity, the authorization and the access of every one of these enterprise agents that are running around your organization. So this is an important concept because people now get hung up on the model side, but that's only 1 out of 3 things after a bigger problem. So if you go to the next slide, please. And this is just an example. This is a completely hypothetical example, but this is -- as we see agents being rolled out, this is kind of a calling pattern. One of the fundamental issues you're dealing with when it comes to agents is unpredictability. This idea that in response to a given prompt. So this is a hypothetical finance agent that's going to help your procurement team with contract or PO approvals. In response to the same prompt, there's going to be tools, there's going to be memory, there's going to be LLM that are accessed and it might caught with different responses in each situation. And I bring up this slide because it's very critical for people to understand that this is actually -- you're not dealing with a deterministic piece of software. You're fundamentally dealing with something that at scale will get you a lot of efficiencies, but has a level of nondeterministic to it that does require a level of human control, which is where dent comes in. So again, very powerful but also has a very nondeterministic aspect to how they work. If you go to the next piece. And there's a builder, so we can just kind of click through this. In the interest of time, we'll pause here. So when you take an average large enterprise. Once you think of their agentic roll out at scale and you think about their users, you think about their [indiscernible] functions, you think about the various applications and resources that this agent is accessing. When you start to do the math and you multiply through the various MCP servers, the various other agents that single agents are calling. When you think about the external services that they're accessing. This is a little bit of a large number, but you could -- when you do the math, you're talking on the order of billions of access decisions that have to happen per day to get a truly agentic enterprise to work because, again, when you multiply the users, the applications, the resources, the agents, the agentic agent chain, the complexity increases tremendously. And each one of these access arrows is a potential risk factor because what you're dealing with is not just a deterministic access claim, but it's an agent that is making a decision at that moment in time and could be exercising adjustment. So that's why the notion of the attack surface becomes even more complicated when you're thinking about agents. And as you multiply that across the entire estate of technology that your average large company has, the complexity is quite mind-boggling and requires centralized control. So if we go to the next slide. There's another thing that you would need to think about here, which is -- so I touched on nondeterminism. There's also a notion of autonomy. Agents are going to take actions on their own, right? And traditional Zero Trust was not made for a world where you have these autonomous actors that are nondeterministic. So you really need to think about how do you control that. The other important piece, which there was a question in the chat about differentiation is this notion of delegation chain. So every regulator, every auditor out there is going to come knocking on a company's door to say, can you prove to me that a specific resource was accessed by a user, was accessed by an agent or was accessed by an agent acting on behalf of the user through a multi-agent access chain. And being able to centralize and track the delegation chain is the other important concern that breaks kind of Zero Trust as we know it today, which was made for humans really. And then finally, there are more and more models emerging ephemeral latents, which is agents that spin up other agents that are short-lived that might execute a specific task. It's an entirely new paradigm of technology and software, and we really need to think about Zero Trust holistically and why we are so bullish on this is, the one common thread across all of this is that context is what was the intent of the original user that gave a prompt to an agent that then either spun up an ephemeral agent or call the subagent and then access to resource. It's the authorization and the identity that's the core of all that to be able to say we can actually control the chain of what is happening. So I don't know there's a little bit of zoomed out, but it does require the industry at large to rethink what Zero Trust is all about and identity and authorization end up being at the core of all of this. So if you go to the next piece. So again, when we think about what Okta for agents is, is really the direction that we're heading in and actually not even heading in. The GA capabilities that we have in market today are really providing this notion of a control plane for every enterprise agent in your organization accessing every resource on behalf of any user in the organization. Centralizing that, governing every call, binding into user context, ensuring that access is provided just in time with lease privilege and then creating an immutable layered audit trail, so you know exactly what resources access, by who, on behalf of who at what moment in time is the core of the platform that we're -- the capabilities we're adding on top of the Identity Security Fabric that Ely touched on earlier. So it's very powerful stuff. It's in market today. And in the interest of time, I'll go to the next slide, which is which is actually this. So just I want to leave you with one -- maybe something a little provocative. So I said maybe a minute ago that I spent a lot of my time educating our customers. There's a lot of its that are popping up that we end up doing myth busting around, things like, "Oh, I have an MCP gateway. I think that's good enough." And the answer is that's good enough if you're not interested in full user agent to resource chain mapping and auditing, which is something that every company is going to need. So that's some education there. The other thing is "Oh, I use Bedrock Agent Core or a copilot studio today. And that has some identity stuff built in." But think about where Okta is today. We are an independent neutral central broker of human identity to every resource in your ecosystem. That's exactly what we're doing with agents. Separating the identity from the core agentic platforms actually gives you not only a level of control, but also a layered defense and depth strategy that reduces your single points of failure. So that's the other big myth that we spend a lot of time on. And the third is, well, shouldn't -- this is just acting on behalf of a human. So why do I care just going to carry the humans credentials? Sort of, but not really. When you have agents that really are helping your organization, these are multiplayer agents to use a term from anthropic or these are enterprise agents that are accessing multiple resources. They do need their own independent identity that can be intersected with the permissions of the invoking user to then create the perfect permission chain. So this is just an example of some of the education we're doing. The market is moving very fast. We are right at the forefront of it. We're spending a lot of time educating our customers, but there's tremendous, tremendous momentum on that front. So I'll pause there, actually, I think we can we can call it a day on the slides because I want to make sure we have time for questions. And I don't know, Eric, if we hand it back over to you.

Yes. Thanks, Harish and Ely, for the comprehensive both review. So a lot of things I want to dig into. Maybe Ely, just coming back to where you started or one of the things you started with was the 4 different buckets of agents. Today, can you just high level give us like set the ground or just lay the framework for us in terms of what's the maturity of AI agents today in organizations that you're talking to? And secondarily, what are the most like prevalent types of agents that you're seeing.

Yes. One quick AI joke here. So we had a large financial services customer in our customer executive briefing center a month ago, and I was asking them how many production agents they had. And they responded with over 2,000. And when we were first doing our pricing analysis for Okta for AI agents, we assume there will be somewhere around 20 production agents in a large customer. And so we -- that was done back in October. And you can see like how fast the market is moving just from that one customer anecdote. But in reality, the most prevalent type of AI agents that's out there, our customers right now are coding agents. So there's certainly a lot of -- and probably more sophisticated customers are building their own agents. But almost every organization we walk into has a Claude code or a codec or another coding agents. And so these are top of mind. And with those agents, it's not really about the number of agents because typically, you can think about that as like one agent for the entire development team. But it's the breadth of usage that's happening and the types of tools that those coding agents are able to connect to, and it's those developer to agent tool interactions that needs to be protected.

Right. Okay. And then, Harish, one of the things that you said that was interesting is I think you called out that discovery is the biggest issue that customers are facing right now. So can you just help us understand what Okta does today from a discovery of agent's perspective. And really, I mean, I think there's a lot of emerging vendors out there that we are familiar with and see you getting a lot of momentum just -- what are they doing that's unique and novel, and like how differentiated is this capability that either you or the emerging vendors have when it comes to discovering agents?

Yes. I think -- so more broadly, Discovery is -- it's not enough to say I run something and I have a dashboard that shows me a list of 10,000 agents. Like that's fine, but then what do you do. And so where Okta is coming in is to say, where most likely are the most critical agents is going to be living and how can we find them the fastest. So that's what Ely touched on, there are known platforms like AWS, Copilot Studio, Azure foundry, agent force ServiceNow. There's a whole host of them that they're already partners of ours, and we could easily detect agents in their agentic run times and then import them doctors. That's one kind of discovery. And the point I want to hit on is Discovery is half the equation. You also want to register those agents in a central identity directory and give them their own identities such that all the tool calls then get routed through Okta. That's the real value there is Discovery is part of it, but giving them identity and then controlling the flow of access is the other part of it. So there's the no platforms. There's browser plug-in capabilities for a variety of browsers because one of the -- one type of agent is also sort of the SaaS agent that you're accessing through the browser. The other one are local endpoint. And so the good news is we already have deep partnerships with the major cyber providers to be able to tap into their APIs and detect agents that might be running on the machine. But again, I want to emphasize that finding them and showing them in a dashboard is 50% of the equation. The other part of it is, say, now what? And that's the part where the full connectivity with Okta matters is to say, "Okay, now I'm going to give you a unique identity and ensure that there's no MCP calls, there's no tool calls that this agent is making that is not authorized and vetted and brokered by Okta." It's really becoming in line and the flow is the end goal that we have here.

Yes. From a product capability perspective, is Okta on equal footing with some of the emerging vendors out there when it comes to discovery? Or do you think this is an area for development, if you will.

Yes. I mean I'll let Ely take that as he's deep in that on a day-to-day basis.

Yes. So right now, what Okta supports is one mode of discovery, which is browser-based discovery. This is for unknown agents. We also have those direct imports for known agents. But for unknown agents, we support browser-based discovery. You can see in our strategy slide, we'll be adding additional forms of discovery. I do think that there are some really unique aspects of the Okta platform that will allow us to differentiate something that folks forget is that Okta is on every endpoint. Every Okta customer has an endpoint agent called Okta Verify, and this is a very powerful capability that can give us broad visibility into what is running on that endpoint. This is -- I mentioned Okta Device Access. It's how we discover what's running on that device to ensure that you have a secure locking experience. So we'll continue expanding our discovery capabilities, including using the full power of the Okta platform to do so.

All right. Staying on the governance topic for 1 more question and maybe switch at 1 time. But on the governance side of things, I think 1 of the maybe things that 1 might theorize in terms of who has an advantage when it comes to the identity stack is -- the IGA vendors, the traditional ones have a lot of understanding of all the entitlements, MediConnect or all these applications, and they have deep visibility into all those different fine-grain nuances that you need to know in terms of establishing access in the application. So what's your perspective on the IGA vendors having that unique advantage? And I understand that Okta has IGA, so it's a little bit of your in both camps, but what's your perspective on that being a strategic advantage or moat?

I mean my take is that we caught up with the legacy IGA vendors, in it's proven out through our customer adoption. With over 2,000 customers using Okta IGA, we have a very powerful IGA product, where some of the biggest customers in the world are looking at us from an IGA first perspective. So I think this is an area where we've invested significantly. We're going to continue to invest to close some of the long tail of feature gaps. But I'm very happy with where we're at on the IGA side. And yes, it's a critical part of our overall products to secure AI agents, but I'd say it's one component. And frankly, it's a really important component, but it's not the most important one. The most important one is ensuring that the AI agents themselves have a least privileged identity that has a small blast radius. And the IGA piece can help you ensure that, that blast radius stays small over time, but we see it as an add-on capability, not the core central capability you are offering.

Yes. What we hear from customers across the board is they're looking for a system of action versus like a system of record where you can keep track of rules and things. They need a system that can enforce the most critical thing, which is, is this agent allowed to execute this specific action on behalf of this user at this moment in time based on a variety of signals, some of them being entitlements in the government system but there are other signals like network signals and endpoint signals and other risk signals and that ability to do dynamic real-time authorization, managing the entire chain of custody is the true ask that the requirement that the entire market is converging on. And IGA is a portion of it, but it's not the all and all. And then that's where the rent piece comes in is because we have a platform that can stitch together a user agent resource and that entire chain.

Right. Maybe moving to the authentication runtime side of things. So we talked about the gateway, and I think there's been lots of mentioned the gateways from all sorts of vendors, et cetera. But can you just double-click on what you guys mean by agent gateway or access gateway, if you want to delineate the language there? And what role that is playing because we did this in our piece, but Okta's role from the authentication side of thing is like, okay, we need to narrowly define the scope of what the user and the agent can do. And if we do that right, the simple side of you might say, okay, it's all established in the token itself when they go to authenticate. But I think you and the market are saying, "No, you need the agent gateway on top of that." So just help us understand what the agent gateway is adding incrementally that's not defined in the agent's token.

Yes. So the agent gateway does a couple of things. One, you can think about it as a single endpoint or access point for a variety of agentic tools. So if you're using Claude code or Codex, the agent gateway is the access point for how those tools interconnect into our platform. And then the gateway itself, once those agented tools are connected to the platform, serves as a policy enforcement point. And so it's where we enforce that those agents are accessing the tools that they're authorized to access. It's also where we provide the inventory of tools that those agents can access. So onboarding various MCP servers, each of those MCP servers have a variety of tools, will allow folks to mix and match those tools into lease privilege virtual MCP servers that then those agents could access, and then it serves as the real-time policy in portion points to ensure that they're only accessing things that they're authorized to.

Can you just be more specific, Ely and Harish, with like real-time enforcement because, I mean, to be can, I think we see that in a lot of like product announcements and press releases, but just be a little bit more specific, if you could, on like what real-time policy enforcement is. Because to me, that seems like it is where the market is really trying to focus its efforts on when it comes to AI agents.

Yes. I'll start, Harish, feel free to jump in.