Allie Howe: All right, welcome to today's episode of the Insecure Agents Podcast. I'm your host, Allie Howe, and today we have Animesh Koratana, founder and CEO of PlayerZero, here with us today to talk about context graphs. Animesh's X article on Context Graphs got over 2 million views and helped us understand exactly how powerful our agents could be if only they had the right infrastructure. Animesh, would you like to introduce yourself

Animesh Koratana: for sure? Thank you so much for having me. Excited to be here. But it was Animesh from the founder and CEO of PlayerZero. PlayerZero is an AI production engineering platform. So we spend a lot of time thinking about what happens to code after it's actually left the IDE. How do you maintain it? How do you operate it? And more importantly, also, how are decisions made in organizations to actually maintain and operate software in production as we build agents that that run SRE, QA, software engineering and more. But yeah, excited to be here. Excited to talk context graphs and the dig in.

Allie Howe: Yes. Now I'm super excited to have this conversation because I remember seeing your X article around the new year and I was super excited to learn more about it, ask more questions. It seemed like something that would enable like enterprise AI as Italy. And one of the parts of your article that I thought was super interesting was that you said that basic context graphs are the missing infrastructure that can help us predict futures capture behaviour and help us understand why decisions in the enterprise were made. Why do you think it went viral? Or like, what was that like, seeing all of the attention around it?

Animesh Koratana: Oh man, it was, it was surreal. You know, I think that the story behind context graphs from from our side was, it's something that we've been working on for, for a long time. We actually did the the word context graphs was like a as it was a new thing, but the idea behind it was something that they gave up for for a very long time. They're kind of funny back story to this was during Thanksgiving, actually, there was a couple days when, you know, all our customers were quiet and, you know, team was at home, you know, hanging out with their families. I spent a few hours actually just writing up, you know, like what is PlayerZero and you know, what makes it different?

And a lot of it actually came down to, you know, building the central understanding and how production software actually works. Part of that was actually understanding the organization of the decision making behind it. So I wrote this up and I kind of sent it out to to a few of our investors. And since that and our all of December was just talking context graphs. And you know, I think what we realized was for the first time we put a finger on what was the gap between what agents were really delivering in the enterprise and what the expectation was right. 2025 was supposed to be the year of agents, the enterprise and for some reason a pleasant and I think we spent a lot of the year, you know, thinking about live, but nobody actually put a precise finger on it with context graphs, with decision traces and all the banks that create cable day.

We actually had to take an opinionated stance on how to close that gap and, and make agents truly productive, right for for really complex, you know, multi persona enterprise workflows.

Allie Howe: Yeah, that makes sense. Yeah. I remember then 2025 where people said this is the year of agents, but it's related to like the next decade. And then people are calling 2026 that you're enterprise AI, but there's a lot of missing infrastructure to enable that. Which context graphs is one of them. Exactly. For someone that maybe is unfamiliar with context graphs, has never made one, Like, what is it actually, like physically? Like what kind of files? How do you store? Like what tangible is a context graph?

Animesh Koratana: Yeah. So in order to understand a context graph, there's actually a a primitive before it. That's really important to understand, which is a decision trace. A decision trace basically is the connective fabric between the that that explains the why behind a particular decision. So a good example of this it's or let's take some of our systems of record that we're used to today, right? Salesforce or ServiceNow or Linear, something like that take Salesforce because you know everybody in the world uses that in Salesforce. We have a record that talks about for every single deal that we've closed, right. What's the ELK contract value, who negotiated it, when did it end up?

But a lot of the decision making about why that that deal was closed to the way that was, when was it closed? What was the negotiation actually like? You know, all the context about, you know, this was an end of quarter deal. You know, this customer had a budget constraint. You know, if these are the specific teams and here's exactly how we need to grow the contract, All that stuff ends up living in the mind of the, you know, Rep, the SC and the few other folks that were in the conversation when that was happening. But what's saved at the end of the day in the system of record tends to just be the outcome.

The decision trace is all the work that led up to that. And in order to actually be able to truly, you know, have agents that mimic the way that humans actually operate in organizations, we need that decision context as much as we made the outcome as well, right? So that was kind of the core insight, right? If we can capture these decision traces, the path to actually getting to the output Reg record showing up in the the system of record, you can actually build an institutional memory of an enterprise that allows agents to actually operate have the same efficacy that humans do in really large complex, you know, non deterministic messies or even violence.

Allie Howe: Yes, that makes sense. And that resonates from what you talked about in the article, which was this concept of the two clocks problem where we have done really well capturing 1 clock, which is state what we have spent years now capturing the event clock which you're talking about. So how interesting and powerful would it be for AI agents to be able to uncover the events that led up to a deal close were said about Salesforce or something, and then you can predict how likely already to close the sale again? What are the criteria that's needed to close a deal like this given all of these different variables?

And the most interesting thing is that the context graph, and correct me if I'm wrong, seems like this living, breathing Organism where it's constantly updated as the agent discovers new information, reverses it and finds different like new branches or new nodes within the enterprise to accumulate. So what is that process like of helping it grow over time?

Animesh Koratana: Well, I think a really important insight here is, you know, agents as they own more and more authority to make decisions in the enterprise naturally end up kind of viewing and encapsulating the decision trace itself, right. So imagine an agent that's, you know, negotiating a sale right now. What makes a difference between a Rep that owns the deal from first conversation to negotiation versus an agent that's doing one small part of the word right? The difference is that the that the Rep actually has the full context about, you know, every single conversation that they're having, every, you know, emotional burden, how somebody could get promoted, all those different things.

As an agent starts counting more and more authority to actually do the same work, then that context naturally ends up being, you know, a view into the the trajectory with that agent over time, right? And so that's actually one of the core insights that that we had right where it's the trajectory itself actually represents the decision trace. And as we actually park attack the organizations that, you know, end up adopting agents, as we architect those organizations to have agents actually have the authoritative decision making ability, then they actually naturally end up creating decision traces. And those decision traces can actually be in committed back into some sort of central understanding that can also be compared to one another.

And that adds up kind of, you know, embodying that institutional memory for the context graph that that we're describing it.

Allie Howe: That makes sense. So it sounds like context graphs are the missing infrastructure that is going to help us create this event clock.

Animesh Koratana: Yeah. So, I mean, the event clock is the agent's trajectory, right? I think that that equivalence is actually it's it's a bit of an odd one they had to think about, but it's actually naturally what happens, right, Right. You take kind of PlayerZero, for example, we have agents that are, you know, working through the entire decision lineage for remediating an incident, right. So an incident comes in, you know, this service is down. People can't check out. That's where it starts. And then you go talk to SREs and SREs have a, you know, playbook for, you know how they, you know, mitigate bleeding and maybe in second to restart the service.

Maybe it's, you know, A to they'll look for changes. There's context that you're navigating along this whole process. At some point you're, you know, talking to support people. When you're talking to developers, you're talking to QA and these agents are actually working through that entire decision lineage to go from here's what broke, here's how to fix it. Let's actually go fix it. Let's go follow through all those different things for SRE, support, QA and more. And that every single, you know, probably saw the production ends up becoming a decision trace for how do we solve problems at production. And so the next time we end up seeing a similar incident or a similar support ticket or something like that, we can actually recall, right, you know, past decisions and say, not only that, Hey, we just reverted the the change.

We also understand why we would reverted the change. And that context becomes really important to being able to do, you know, really hide judgement work at the enterprise.

Allie Howe: Super interesting. Can you also use that context to predict OK? Like based on this scenario that happened before that we capture this event tree, this change looks like it's potentially going to break down the server OK.

Animesh Koratana: You absolutely, yeah. So this is this is actually where you know the step after context graphs becomes really, really important. You know, context graphs are are really meant to model kind of like decision dynamics within an organization. And if you model these things really well, actually become closer to world models than like static graphs. And what world models allow you to do is they allow you to simulate or to have predict how things would behave under certain conditions. I'll give you an example of how we do this at PlayerZero, right? We have this, this idea of code simulations, this what we could take is under specific set of changes, we can simulate the behaviour of your production infrastructure and your production software.

So we can say if you would have changed this particular area of the code and then we, you know, go and simulate this incident with this particular ticket that happened, you know, 5-6, ten months ago. We think that the changes that you're about to introduce are going to reintroduce that particular problem that happened to get rid of it, right? And so we can actually do this reverse mapping as well. And you can't imagine, right? Like they can actually introduce new forms of a verification, new forms of working right to these workflows just because we can model all decision dynamics in a way that you just never really could before.

Allie Howe: Yeah. That's amazing. I, I feel like that's such a state change in terms of like technology today. I mean, that's just revolutionary in terms of like how which better we'll be able to make decisions as an enterprise and once we unlock this functionality. So you all feel seem like you're very far along with this. Obviously at PlayerZero, you're leading the charge here. What is the difference between where you see PlayerZero today and the average enterprise? Do they have this technology? Are they using context graphs? Are you? Are they getting the most out of AI agents today?

Animesh Koratana: So I think most enterprises are not getting the most out of AI agents. I think actually a great kind of comparison for this is we're both start up people think about how we use AI and if we didn't have AI, how many more people would need in our team to actually be productive, right? To get the same amount of output. I have to talk about this eternally, right? We have a team of about 3035 or so folks. If we didn't have AI, if we didn't use PlayerZero, if we didn't use a 0 Claude Code and all the other great things out there. I think for engineering team, we have to be something 150 to 200 people range.

Now, if you went to enterprises, you asked the same question, you think that they would answer the same way? I don't think so. I think they're probably maybe 20 percent, 30% more folks. They're getting some no benefit from this, but they're not seeing the kind of step function that, you know, that we see as kind of the AI native start-ups, like kind of building building into the space right now. And I think actually this comes down from to like philosophical difference of, you know, is AI additive to an individual retributor's work? So does it make an individual more productive or does it actually have re engineer or rethink entire function?

Right, and I think as an AI native startup, we often think about what is the AI native version of this function look like where agents actually own the authority to do the work and make decisions. Whereas an enterprise today, I think largely just because of this trust barrier and hopefully context graphs best start eroding S up barrier. But at least today, what we see a lot of the enterprises and think about agents being additive to an individual, right? And so there's a ceiling to how much more productive you can make an individual just because they have a minute amount of time of the day, right?

They they, they can only, you know, do so much and you can only manage so many Claude Code agents and all those kinds of things. We really need to be thinking about what abstraction Hireright, what do kind of multiplayer look like as they navigate different, you know, organizations after as they navigate different functions right to different people and your organization to actually, you know, create a certain outcome.

Allie Howe: Yeah. So I've had some enterprise power stations too through different discovery conversations or potential customers from my startup. And I'm definitely seeing that trust issue as well. And the power of context graphs and enterprise AI is very alluring. But today it's blocked by AI security. It's fundamental component of enabling this like enterprise AI dream where we can predict events and understand exactly like how events took place. I think one of the reasons why AI security is such a blocker for enterprise AI, What I frequently see is very similar to what this blog from Keycard Labs not too long ago that they posted. So basically like this article says there's this 3 trifecta pillar of capability, autonomy and security.

And you can, you need all three in order to actually unlock this trust equation, but most enterprises today only have two. So I'll I'll give you an example. So with capability and autonomy, agents run wild in God Mode and they've got no security whatsoever. And with autonomy and security, if you're missing capability, then your agent is very narrowly scoped and it's not able to actually do meaningful work. And then if you have capability and security without autonomy, then you're stuck in this human in the loop situation where your humans are dealing with consent, consent fatigue. And that's not productive or very meaningful either. Enterprise are stuck on this.

How do I trust AI agents? How do I actually enable them? What infrastructure do you think is missing for AI security or do you see these issues happening? What are you? What are you saying?

Animesh Koratana: Yeah, that's a, that's a great question. I mean, it's, it's, it's funny to draw it out in this way where, you know, there's this pendulum between, you know, if you give AI access to everything, it can do some really cool things. But that's just not a practical thing, right? For for for most enterprises. And then I think there's like another layer on top of this when we start talking about things like context graphs, where what we're saying is actually there is a, a system of records, so to speak, that is actually outside of the systems of record that are already governed, right? That's why these decision traces, like all of those different things are actually like derivative sorts of data that live in between these different systems of record.

And so like, how do we actually govern, right, data that comes out of decisions that are made in between these different things? I think for a long time, these things actually used to just be, you know, relegated to people's minds, right? This is these two humans, right? That used to have access controls and all those things that you're attached to the identity of the human. But then like the decision making happens within their heads. And so it's kind of, it was a natural and easy answer for a long time. There's been kind of like 2 important ways that we've been thinking about this. One is to think about what is the identity of the agent, right?

And I think this is something that I've seen, you know, more kind of AI native companies lean into than enterprises. But the same way that we manage identity for every human, there's also an identity for an agent. And that agent, you know, has access to certain systems of record. It can take certain actions, something stuff to be escalated and and obviously approved right by humans. But the agent has its own identity. And I think that's like an important kind of mindset shift because you can actually start structuring agents to work in the same way as humans. If you had considered the identity equal, I think the other direction has been, you know, thinking about, well, how do you infer the governance of certain objects or certain records based on, you know, primitives that already exist?

So like very simply, you know, I have access to a certain set of Salesforce records. You have access to a different set. You know, I might have access to the code base. You might have access to, you know, the ticketing system, but maybe not the code base, right? Like whatever that might be, there's governed objects that, you know, we've already built governance for right in our enterprise through access controls and all those different things. As we build a context graph, we can essentially go figure out, you know, for a particular decision, what are the different governed objects that that touches? And then if that essentially mapping our, you know, access controls to separate objects and then using that as a way to kind of limit to or not limit the, the access that an agent might have as it's actually doing work.

So this is kind of like two different approaches to this. But I think to like zoom out a whole bunch here, like security is incredibly important, right? I mean, there's like, there's no way that like we had sidestep this as a, as, as, as a technology or like as a, as, as a product, something that needs to be kind of built in from first principles in the way that these AI products are actually built. It's not, it definitely can't be an afterthought and it definitely can't be something that that would take for granted, especially given that like, you know, a majority of the value creation that's about to happen through AI is actually going to happen in the enterprise.

And you're saying,

Allie Howe: yes, I think agent identity is definitely a key piece to enabling context graphs as well, especially because these context graphs will end up having different nodes in them that we did not anticipate at agent inception. So the agent needs the ability to sort of walk around the enterprise, build this data ontology with nodes that it may not have been granted access to originally, but my age to request access to, there needs to be some sort of like methodology there. It reminds me of an article I read recently around some goal based authorization where I made this analogy to AI agents being kind of similar to humans in the way that like maybe the CEO goes and hires an assistant or something and then that assistant comes into the enterprise and they're wearing a badge.

It's kind of like a proxy to the CEO and they can walk anywhere within the organization and maybe they need to like, I don't visit one part, but they're able to do that just because they're proxy of the CEO. And so we basically just like trust that, but we also need to keep like a system of record around like what did they access? Why did they access? And we need the same thing for agents and with agents too, we need to sort of do like a DIF between not just the permissions the human has maybe to the Salesforce records that you talked about, but also what the agent has access to.

And so how do we combine those to create some sort of context aware authorization model? Very difficult problem to solve.

Animesh Koratana: Very, very difficult problem to solve. Yeah, I mean, that would be that would be kind of like the the, the dream state. I think making sure that it's interpretable and like transparent. I think it's going to be the hard part there. But yeah, I mean like that, that would be incredible if you could actually inherit permissions as you are, you know, navigating the organization. For example, like one thing we often say internally is like, you know, for agents to be truly effective in the enterprise, it has to navigate not only context but also people at process. Right. So it's like people processing context and then like a part of navigating people in processes personally because, you know, different people in the organization have different context just by virtue of what they have access to, right.

And so, yeah, if there's like an inheritance model or something where as the agent itself is navigating through different people in the organization, it's equal to accrue context and access that that the people in the organization may have implicitly the that'd be pretty badass.

Allie Howe: Yeah, that would be super cool. Yeah, maybe that's on the way. We'll see. Yeah, for sure. Have you seen other different types of context graphs? So for instance, I know your article came out in the beginning of this year and I was so quickly and this was not one of the topics that just like went viral and died. It's definitely still in conversation and it is building. So what I've seen recently between now and then is people publish skills graphs, graph DVS for agents. There's so much work people are doing and to help AI agents do work more effectively. What people are finding is if you give them these graphs, you're given these schemas, then they don't get lost.

They're able to follow and navigate more effectively, which I feel like is key for creating a future where AAI agents, right, all of software, which I think is where we're headed ultimately. Have you seen any of those articles around? Like, what do you do?

Animesh Koratana: Yeah. I think it's interesting. It's like ever since the article came out, I think everyone in their mother thinks that they've built context graphs. And, and to be honest, like there isn't like a prescription for exactly what a context graph needs to look like or exactly how it is to be built. What I can kind of safely say is, you know, the way that we built context graphs inside of PlayerZero look significantly different than I think most of the other articles that we've seen out there. For starters, actually, it's not in their graph. It's actually not it's not a graph. There's no notes. There's no it's it's it's it's a different kind of state space, a different type of tool and whatever trained ends up looking and feeling very different.

But I think graph is at a good mental tuition for like how to understand how we're actually trying to model. I think actually the the like the, the data ontology or like the way like a context graph. Actually looks is also vertical specific because the way you actually capture decision traces, like the bottleneck in a lot of this is like, how do you actually capture the decision traces effectively? And that if if that's the bottleneck, then you actually have to engineer everything underneath it to go and, you know, work around and optimize that particular bottleneck. And, you know, for us again, around like it's a response or ticket response or QA.

And like these types of functions, we had to build a certain UX, a certain interface for these agents to actually go and operate with the people process context, right? That's what we talked about earlier for, you know, somebody building in like the AI sales space or somebody building in the AI security space, or, you know, like these different functions in these different roles will actually look like the way the context graphs materialize will be different just because the, the way that you'll actually capture the decision traces will be different. And so I, I think that's actually the core insight here. Like there is no prescriptive way to do this.

I, I strongly think that there actually is of like a universal context for, for, for every like for every, you know, enterprise in the world. More reportedly, not only there's, there's no universal contact for every function in any and in every enterprise either. I do think there's a convergence of any different functions. So for example, like, I don't think you know, five years you're going to have SRD in support and you know, QAB different things that they all operate on with the same kind of fundamental understanding of context, right through building an intuition of like exactly how does production soft rocks would work? But like, you know, the sales context graph and then the engineering context graph are are different to fundamentally and the way you capture decision traces in those cases are different.

And therefore the price that own it and the infrastructure that owns it will also be different. That's probably like my, my quick take on, you know, all the articles that have come out. There's definitely mirror to that, but it looks very different than for the way that we built it internally.

Allie Howe: That makes sense. So yeah, it's very interesting to think about for sure, 'cause I feel like I saw one example, a Nanograph photograph DB, and they were modeling different, like Star Wars characters. I guess. Minecraft it was just a conglomeration of files so just kind of like a, you know a database schema about like this is what I would store from a Star Wars character who's what's interesting to me if you like personally, but what's interesting to me not not being distant to you or someone else that builds the same thing

Animesh Koratana: I. Mean like I think the context graph can like be broadly divided as like, you know, a model of the decision dynamics for a particular function or like, you know, group of people, how you choose to model it and whether it be a bunch of files that you just shove into the context with though, whether it's, you know, custom training models, whether it's, you know, something in Neo4j or like, I don't know, like there's many different ways to actually go about it. I do have opinions on like what's going to scale like what's going to work with you. Well, I don't think it's going to be a graph in like the most traditional sense.

I think it actually creates a bunch of security problems too. You try to model them in a particular way. But yeah, I mean, I think at a high level that's that's the the definition. So the implementation could be over charred.

Allie Howe: Yeah, it's very interesting to think through to which way to model it. Would be better or worse for security the trade-offs there because I'm imagining that two of those actually like a graph data structure underneath Then you have to think about and you probably still have to like file like which agents can go to this visit these nodes or this part of the graph. And so you have to do some sort of Asia identity or author on every single load, every single traversal, every single walk of that graph. And then also like, where are you storing the graph with it's all files? Do all the agents have access to that?

And if you're running agents like in sandboxes, how can they communicate with that? And you add something to the graph that maybe you shouldn't have because like a Prof injection came in and then you've got poison memory or poison graphic. How do you deal with that problem? Like so much, you have to think. About

Animesh Koratana: it's, it's a whole can of worms. I think there's, there's like another can of worms around like, well, how do you keep the, the context wrapped up to date as like decision dynamics change and then also as the underlying records change, right? So if I had the, the, the, the record at Salesforce, right? Like it needs to somehow, you know what, how does that actually affect our decision making executions? Yeah, I think there's there's only a worms there as well, because at the end of the day, somebody needs to have access to that record or they were not to synchronize it.

Allie Howe: Yes, death of line and what this skills graph that I saw that somebody made that spill a lot like MCP to me because they're just they're saying that this graph enables the agent to make better decisions around like which skills are available a console accident like what context like it basically can have in any given time. But I feel like that's very similar to MCP.

Animesh Koratana: Yeah, it is. It's like there's a lot of cute ways you can like try to put these things together. Agents are extremely expressive and extremely versatile. And so the fact that like, you know, things I'm at work is incredible. But I think again, the point of the context graph is to find a way to accrue institutional memory. Institutional memory is like is much more than like no graph of skills or you know, an ontology over records like that. There's there's something more to the way that people make decisions explicitly and implicitly in large organizations. I think like as agents start owning more and more work, like real work and real outcomes, they are going to have this natural, you know, advantage to actually capture and reconcile, represent the way those implicit and explicit decisions are made in very complex organizations.

And then then, yeah. So we're excited for that amount

Allie Howe: because today the OS agents are just doing static task completion.

Animesh Koratana: Exactly. Yeah, yeah. And they're like building their understanding of the world from scratch on average trajectory there. There has to be their way to do this.

Allie Howe: It will waste of tokens too.

Animesh Koratana: Exactly. Yeah, yeah, she should look at our bills. Our entrance bills are crazy.

Allie Howe: Yeah. Oh my God, it someone wants to get through with contact scraps today but revise you how to.

Animesh Koratana: Be honest, you just have to really deeply understand how decisions are made in the problem space that you're trying to trying to operate it right. If you're like, if you're building something in, in our space, like we have to think a lot about, you know, how are decisions actually made? Like not about how they say it or made, but how are they actually made right for incidents, right? When an incident is being responded to, like when we're thinking about a rollback where we're actually executing A rollback, we're making a code change, we're reviewing that code change. Like every single tiny decision from, you know, something happened to an outcome has to be modelled and thought about in order to actually capture the decision trace.

I think that is probably the best advice I can get. So just think deeply about the problem and how does these are made and if you can capture those things truly, I think a lot of the the context graph pieces end up kind of falling out underneath it. You know, beyond that we're over engineered a technology or anything like our front capture incision and the rest of the code happens.

Allie Howe: That makes sense. And that deep understanding of the problem, that's something you can't find code. Yeah, I think is so party. Yeah, because in such a crowded space. So I'll talk about my product just as a real example. I sit in the agent like SAS space. There's so many different players in that space, but I hear like the difference between someone that's going to make it long term and not is someone that able to take advantage of that context graph and be able to pull an understanding from different parts of the SDLC, understand how decisions are being made. Really close our loop. That loop between developers and the security team.

They can have a lot faster so.

Animesh Koratana: Yeah, to the local account that for a second the like done really well. The context graph ends up creating a bunch of feedback loops, right from one agent to the next agent's trajectory from, you know, one agent to the outcome that's actually going to be driving to its efficiency. Like all these different things, You know, in the same way that agents today are agents, I guess, humans, the way we hire humans is like some combination of like, you know, expertise and experience. Like that's how we value employees. Basically. The same thing is going to happen with agents so long as, you know, they're capturing labor budgets, right?

The expertise comes from, you know, Sonnet 4.6 or whatever, Opus 4.6, right, like BME Opus 4.7 or whatever, that's expertise like queen malls. Great intelligence experience comes from this context graph. It comes from some sort of externalized learning. And that's how I think these these types of companies PlayerZero are basically going to stay relevant, right? Or how we're going to remain sticky in in enterprises over time. And I think that's the other reason why context graphs actually went fire, because we actually kind of like, you know, put out a plan that said like, you know, This is why PlayerZero is going to get better every single day that you use it through every incident that we respond to and all those kinds of things.

Whereas most other agents that you're that you might buy from, you know, potential competitors or that you might try to build internally don't actually have the same positive feedback loops.

Allie Howe: Yes, that makes sense. And I agree that like this is quite a decade of agents like you Will said in 25 take some time and every year, I think year over year we'll see more and more people get left behind. Like the the diff gets so much wider between like who is using this technology the fullest and who's not and so many of those startups that got started today and it's just doing aesthetic. Here's a task that we did for you and it's not like a context graph. It's not going to improve dramatically over time. That's not going to help enterprise predict outcomes and pretty old behaviors.

It's just going to be wider and wider. And for anyone that wants to take advantage of such a revenue generating opportunities to actually unlock the ability to have enterprise AI to predict outcomes, they're going to need to think about things like AI security and building contacts for Ashram day one and making a big part of what they're focusing on. Absolutely. Amazing Super Bowl. Is there anything else you would like to share those before I wrap up?

Animesh Koratana: No, I mean, this is such an exciting time to be building. Like this is, as you said, like every couple weeks, you know, not even every couple weeks, every every week, every day, right? Like there's something new there. It's really important in this, in these times of moments to like think about, you know, it's like the Jeff Bezos thing, right? It's like in 10 times of change, you have to think about what doesn't change, right? And again, like the way that we evaluate and underwrite, you know, employees, the way that we, you know, think about their experience and their expertise, like all these kinds of things.

I think I've, I've like long stayed the same. And the way that we evaluate agents are probably going to be similar as well. And so like these, these primitives that we build for how our agents operate, how they get better, I think it has stayed true no matter how smart these models actually get. Excuse me. And so that's kind of like the, the, the space that, you know, start-ups have great opportunities actually now capture they're really, really big loads though big businesses on on the docket those builds as well.

Allie Howe: And then he's like super Bolt. If people want to continue the conversation about context graphs or learn more about what PlayerZero where can they find you

Animesh Koratana: Well. You'd find me on Twitter, I'll make Koratana notes where you'll also find the original context graphs articles that though we'll be talking about. And then beyond that, you just should be an e-mail Koratana.

Allie Howe: Amazing. Thank. You so much for taking the time today. Appreciate it.

Animesh Koratana: Thank you.