The company brain for shared skills and ontologies

0:00Austin KronzIf you heard Prukalpa share, you know, this actors, artifacts, activity, and, you know, sometimes I even have to remind myself. I was giving a presentation a couple weeks ago and, and someone actually stopped me and said, "You know, this is nice. We love this idea. This is a great framework and mental model, but does it exist?" And I have to say, "Oh, that, that's like the whole idea of what we're doing here at Atlan." And so we feel like we figured out a lot of these nuances to putting together an enterprise context layer and what that means for different use cases. And so hopefully, this next session will answer just how do you actually get there.

0:33It seems like every organization I talk to is somewhere in this maze. You know, they see AI value on the right-hand side. Yes, we want to get certified answers through agents on our enterprise data. We would love a support ticket to just answer and resolve itself. We want everyone in our company to be an AI superhero. We want agents that build things automatically and have autonomous software factories. We have this massive landscape somewhere on the left that is an input to all of that. And in the middle, we kind of get caught up and lost in a lot of big questions. And the questions start from very practical, Uh, you know, "Hey, should I have a local context layer or one global context layer?" Fair question.

1:11Should-- You know, what happens if we have thousands of skills? How will the agent actually pick them up? Uh, we're gonna need an enterprise ontology, right? And the questions start to get more and more theoretical and what if, and, "Hey, I heard RAG is dead. That means we need Graph Rag." Uh, and I'm starting to get to the point in the conversation where I'm asking like, "What are you actually building?" You know, these are getting harder to answer because, yes, you might need them, but no, you might not need them right now. Uh, you know, if the models get smarter, won't we just have to start over? And, you know, sometimes I jokingly say like, "Hey, if it's Sunday and it's raining and my car just got hit in the parking lot, will our agents still work?" And so this is actually really challenging and gets very theoretical.

1:55And so the way we like to think about it is just as every technology implementation starts, uh, start with an actual use case and unit of work that you want to accomplish. And for us, each workload needs a different type of context, and this is where the yes, but or yes, but not right now might come into play. The workloads that we see, and there's four of them, a lot of these agents in LLM started in the coding space. It's where skills and tools were created, these instruction files. There were core context artifacts for a particular workload around, um, software development and coding. Knowledge work became the next thing. How do we take, h-honestly, it was some of these same harnesses, and you-- then you eventually saw, you know, OpenAI, ChatGPT come up with their cowork types of harnesses, but it was still largely skills, tools, instructions that powered many of these.

2:47In the enterprise, though, we want to talk to our data. And so if I want to build even a personal productivity, a knowledge work kind of agent, I'm gonna need real metrics and data from my organization. And so this idea of AI-ready data, certified metrics, um, an ontology that can actually map the business become key context artifacts for that workload. And then we get into operational agents. It's probably a mix of all of them. These are the most autonomous and kind of closed loop agents that you might be building. And so, yes, you will need all of these things, and retrieval will change for, uh, the-- depending on the workload. It could be simple retrieval filling the context window.

3:24It could be graph retrieval, a traversal. It could be, you know, a vector search, all of the above. And we'll get into this more. But enterprise context comes in because you start to realize there's a lot of overlap in these use cases. Yes, my knowledge work might involve metrics, and that same metric is used in an analytics use case, but I also need it for operational. Ontology is the map of the business, so naturally, it's gonna span multiple workloads, so this context artifact is very important across different streams and teams, uh, and different skills. You know, internally, we build a lot of, you know, brand-related content. Uh, and so whether that's an enterprise agent or claw that we have built internally or something that's externally facing, we need the same type of brand skill, uh, across multiple workloads and multiple people.

4:12Ultimately, these all fall into the same type of loop where we wanna register these context artifacts, manage them like code. We'll eventually utilize them, and that utilization is what will actually tell us how we can improve this via traces, uh, and see what context is missing, needs improved, or needs, um, uh, pushed further along the life cycle. Just to anchor on everything Prukalpa talked about this morning before we show you what this actually looks like, you know, context, uh, artifacts need to be built for agent retrieval. I think our first iterations were really focused on humans and agents, but agents need very different things than humans do. Manage context like code. We have a long software history about how to version, release manage, um, own, you know, manage dependencies in code.

4:56We take that to the context world. Context compounds over time, and it needs to be open and portable. So if you need to switch platforms, like Steven was saying, every platform is gonna have their own kind of context assembly agent type studio. But if we wanna switch tools or we just wanna reuse context across these four workloads that I just mentioned, we need to be open and importable by default, and that's the idea of an enterprise context platform. So with that, uh, I'm really excited to first introduce Rohan and Ashwin, who will walk us through what this looks like in Atlan Trident, and then we'll very quickly get into a deep dive on the analytics workload itself, uh, with Zach from our team, who's been working with a lot of our customers building these talk to data agents.

5:36And then you'll see me pop on for a short second to introduce the rest of the, the panelists and have a live discussion. So with that, I will stop sharing my screen and hand it to the people who are actually in the weeds with this to show you what we're doing with the enterprise context platform. Rohan, Ashwin.

5:54Rohan GoelHey. Thanks, Austin. Uh, super excited to share how we are using Trident at Atlan, uh, since last few months and show you all a demo of, uh, what we have been building. Uh, I'm Rohan, part of the product team at Atlan, working on a lot of initiatives. Super excited to share it with you. So before I go deep into the product demo, I'll show you our AI stack. So we are a pretty multi-harness organization. You can see we use Claude, we use ChatGPT, we use Conductor, DragBot, anything you name it, uh, we got our eyes on it. Uh, and this will be true for a lot of organizations out there. Nobody's going to put a bet on a single harness or a model provider.

6:35So we have built Atlan as a desktop-first experience, so we can integrate very closely with these harnesses where all of the work is happening. It's very easy to get started. You just download the app, sign up, set up your organization, and you can get going. Uh, it can solve a lot of different problems for you, like Prukalpa and Austin shared, uh, in their keynotes. Uh, so one, you can manage your skills where you can create shared knowledge for all your agents that are running in your enterprise, all the harnesses and all your teams. You can also manage your fleet of agents. You can govern their cost. You can look at all the agent behavior, all the activity that these agents are performing and take action on them at scale.

7:17Uh, and you can also run your analytics workload, uh, and build these analytics agents grounded in enterprise context. So we are going to show you a demo of all of these three, uh, things and how, uh, Trident can solve it for you. Uh, one other thing before we go in. You can see I just signed up, and it says there are seventy-two workspaces that are already present in Atlan. So Workspaces is a shared space for you to organize all your artifacts, all your agents, and all the activity in one place. Generally, uh, what we have seen, uh, with our preview customers, they set up these workspaces the way their business is structured, so different departments, teams, sub-teams.

7:55And it gives decentralized ownership to all these different teams to manage all their artifacts. So I'm going to go in. You can see as soon as-- The first thing that the app does for you is it looks at what all agents have you already been using in your laptop and brings everything together in one place. So if you see, it has curated a personal workspace for me. I'm just going to go in. So this is my personal workspace. You can see there are skills that we have collected from your different harnesses. So I'm a Claude Code and both ChatGPT user, so I have all my skills in one place. I have all the artifacts like presentations or images or any files that I create, uh, across these harnesses also clubbed into one place.

8:36All the activity that I've been performing is, uh, mapped to all the artifacts that are present here, and I'll show and talk more about this. Uh, I, as a person, actually hated switching harnesses. Uh, so we actually started with Claude Code, uh, and then we also got in ChatGPT. I was actually not willing to switch, uh, because all of my daily workflows, all of the context was already present in Cla-Claude Code. With Trident, that problem is now gone, uh, because it just brings everything in one place. And if, let's say, I have a skill to build product demo videos, uh, that I created in ChatGPT Uh, but, uh, Claude launched a new model, Opus 5.5.

9:14I want to try that. Uh, I can just use the same skill in Claude, right? Or do the other way around. Uh, so this, this has been a great, uh, experience for me as a user who's us-using multiple harnesses. Now, let's maybe go into one of these skills that I built a month back for, just for myself to build, uh, demo videos. If I'm super happy with it, I know my whole team can leverage it, I can just click Share, and I can move it to a shared workspace, right? Uh, if I feel the whole team can leverage it, I can just pop it in, and everyone in engineering will have access to this. So this is what will move, uh, AI from a single-player game to a multiplayer game, where if you have best practices that your one team or one person in the team is devel-developing, uh, they can just, uh, sort of make, upskill the whole team in one shot, right?

10:03Uh, so for now, I'm going to skip that. Uh, I actually have a customer call, uh, right after this conference coming up, and I have to prepare for it. Our revenue team is actually pretty good, uh, when it comes to interacting and talking to customers, so I'm going to go check out that workspace and see if someone in the revenue team has actually created a skill that I can leverage for doing my calls. Uh, so if you see on your left, uh, all the artifacts, uh, are actually present here, like skills, projects, metrics, ontologies. You can also bring in all your agents. We're going to deep into, uh, all of these, uh, concepts and activity, uh, whi-which is, which these agents are performing, right?

10:43For now, I'm going to go check out if there's a skill that I can use. Uh, so I do see a call prep skill. I'm going to go in and see if it's actually useful for me. So before we built this product at Atlan, we were using GitHub to manage all of our agents and skills, and it created a lot of chaos and problems. And this is a product, uh, we, we sort of built, uh, from all the learnings we had internally and talking to all our customers. So for example, if there's a skill that is present on GitHub, it was very hard for me as a non-technical user to actually understand what it is doing, right?

11:18So here, if you can see, it shows how it works. It also shows me prerequisites. So if I have to do this call prep, uh, it actually needs connectivity to Google Calendar and Confluence. Otherwise, it's not going to work. So it tells you what you can do with it, when it will get triggered, and how you can use it. For skill maintainers, there were bigger problems. So let's say if you created a skill, put it on GitHub, or people used to share these skills over Slack and email. Now, the problem it creates is: How do I know if my teammates have actually used it, right? So if you see, I told you we, uh, integrate very closely with all these harnesses.

11:54You can see this skill has actually been used by four people in my team, and it has, has vari-various runs. So this is how we map all the activity from all your agents and bring it back to all these artifacts. So skill owners and maintainers will actually know that, hey, people are actually using it. And I, as a consumer, will be able to build trust that this is not a skill that will be stale. It's actually something that my whole team is leveraging, so I should probably just use this as well. Um, another thing that we realized was, uh, you create a skill, you put it out there, and then nobody looks after it, right?

12:29Like, if it's actually getting used, you should probably keep improving it over time. So because we have all this activity already present in, uh, Atlan, we can easily map that to the artifact and come up with suggestions. And we can also tell you what exactly is it going to improve. So if you see, uh, in different parts of the skill, it's trying to suggest these are the things you should change. Uh, I'm not going to do that. I'll leave that to Ashwin. But you can just in one shot, uh, click Apply, and it'll upgrade your skill. Uh, I'll just quickly show you a few more things. So we believe context should be managed as code, how you manage code on GitHub, and every repository has a profile.

13:09Similarly, we create a profile for your skills and other artifacts and bring in the most important information together. You can see the source code here, obviously. But if let's say I wanted to do a actual comparison on how the skill has been changing over time, I can do that. Uh, we automatically version any upgrade that happens to any artifact by default, so it helps in a few ways. Uh, one, to skill maintainers, they'll actually know if, let's say, twenty people in my team are still on the old version, right? Or there are a few agents that are leveraging this skill, and they're using the old version, right? Or if you are building critical agents, let's say a support agent or any operational agent, and you're leveraging skills from Atlan, you can say, "Hey, don't automatically upgrade it when someone pushes a new skill.

13:54Just pin it to v1. Until I test it or run evals on it, I don't want you to move it to v2." So you can see right now it's set on auto-release. You can say, "I want to do manual releases," and you'll be able to do that. Uh, one other problem, uh, and again, related to context being managed as code, how in code you have code dependencies. If you are installing a library, you also need to install all the dependend-dependencies from that library, right? Same thing will happen with all your context. Uh, so if you see in this skill to do call prep, it's actually dependent on a lot of other skills. If I just use this Markdown file in my Claude code, our ChatGPT is not going to work because I need these five other skills to work.

14:36Uh, for example, there's a skill for customer references, uh, which will actually me, help me do a better call prep. Right. Uh, we can also automatically detect duplicates, and this, this happens a lot. It happened with us as well, that people were just creating their own stuff before, uh, checking out if there's anything existing out there. So it can automatically detect. And we also have agent, which you can just click and say, "Hey, can you find conflicts or figure out what is different about these two skills and just create one version for me?" So we create all of these relations. Again, as soon as a new upgrade lands in for any of these skills, uh, we'll make sure that it's correctly mapped, uh, in these different relationships.

15:18Uh, one last thing I would like to talk about, it's the new hot topic, uh, since last week, is cybersecurity. Uh, so skills and any important context artifact will create a new attack vector for your enterprise, right? Uh, you're going to use these skills in harnesses. You're going to put these skills in agents, and that agent is going to be available to all your teammates, to maybe external, uh, people. Uh, so it-- someone can do, let's say, a prompt injection or ask it to sort of install a malware in someone's laptop. So we make sure every skill is scanned as soon as it lands in Atlan. And we also ensure we have policies where you can say, "Hey, if there's any high critical vulnerability in any skill, you won't be able to-- we won't return that skill in APIs, MCPs, and through our SDKs." So, uh, you can set those kind of policies.

16:09Uh, maybe one other thing I will show, uh, before I, uh, recap is, uh, you would ask, "Okay, this is great, uh, but my users are going to be in ChatGPT and Claude in their harnesses because that's where they're doing all the work. So how do I take all of these skills, all of this context in the harnesses which people are using every day?" So I'm just going to open my ChatGPT. I just prompted it, "Hey, can you prepare me for my next customer call using the same call prep skill that you saw?" Right? It's the exact same skill. If, let's say, Ashwin or anyone in the team pushes a new copy of the skill, I'm going to get the latest version of it, right?

16:46So just imagine five hundred people in your team, uh, using the same prep skill. Uh, anytime someone upgrades the skill, all five hundred of them get the latest copy, right? And it's very seamless integration. You don't have to literally do anything. Uh, it just works out of the box. Uh, and it works, uh, with almost all the harnesses. So just to recap what you saw, we have created a workspace, uh, product which brings in all the context artifacts, all the agents, and all the activity, uh, these agents are performing. It takes you from a single player to a multiplayer game, uh, and how you can manage context trust code. So that's a quick recap. Uh, I want to show one more thing, uh, and this is a question that, uh, our customers ask us a lot.

17:31Uh, so the first one is, "Hey, I have a bunch of agents running in my company. Uh, first, I don't even know what all agents do I have. Uh, and second, I, I cannot track how much am I spending for these agents. Is it even worth it?" Right? Because initially, people were trying to do more with AI. Now, now they want to get more value out of AI. And the second question is, uh, "I'm buying a lot of subscriptions for all of these new harnesses that are coming out. Uh, investing in new harnesses, are people even using them day to day?" Right? So I'm going to quickly show you, uh, how we can solve that problem.

18:05Because Atlan, uh, has all the context, all the ac- artifacts, all the activity in one place, we'll, we'll be able to visualize all of these, uh, question-- We'll be able to answer all of these questions for you. So for, for example, if you see in our case, we have about, uh, one thirty agents that are running in production, but only fifty-three of them have been active in the last thirty days, right? So maybe our engineering team should sit and figure out if there are agents that I should just deprecate. Right? You can also see the cost over time, uh, how much it's spending week over week. Uh, and obviously, you can break it down by different workspaces or different agents.

18:39There are a lot of, uh, uh, useful, uh, questions we can help you answer. So our engineering team actually loves this. You can see which agents are most active, how, how much are they spending, and you can also do some more technical and nuanced cuts. Uh, so for example, which agent is actually spending most per run, right? Uh, which agent is slowest? Uh, so this will matter when you actually have hundreds of agents running in production and you want one place, one control plane to manage all of them together. You can see adoption, uh, who are the people who are actually using these agents, because there's one thing on creating internal agents or external agents, and there's another thing on people are actually using them to get, get something useful, right?

19:20Similarly, uh, we also can do, uh, help you figure out how people are using the harnesses that you've bought. So it shows some token spends. This, this is a pretty interesting metric that we track internally, where we have created a lot of skills, but are people actually using them? Because if they're not, uh, we should just get rid of them. It's very easy to create new skills. If you see, we have about eleven hundred skills, but only three hundred of them have been used in the last thirty days. So you can just put, uh, it in front of your Claude or ChatGPT or MS Copilot and ask it, "Hey, find out all the skills that have never been used." Uh, so to recap, uh, we help you understand what your agents are doing, what-- how your humans are using these AI tools.

20:05Uh, I'll pass it on to Ashwin, who will go deep and help everyone understand how to actually put operational agents in production.

20:15Ashwin MuraliAll right. Thanks, Rohan. Hi, everyone. My name is Ashwin Murali, Lead Solutions Consulting at Atlan. I spend most of my time with customers helping take AI agents into production where real users depend on them. And what we keep seeing is that this last step is where most agents break, usually because the agent doesn't really know how your business works. Rohan showed you how our own teams work with AI, so I wanna show you the kind of agent your customers actually talk to, and one customer in particular. This is Maya. Maya is a customer of Halcyon Mobile. She's out $10, and she's called three times about it. So this time she opens the Halcyon app and asks the customer care AI agent instead.

20:59If you think about what a really good support person does with a call like this, there are basically three steps. First, they observe, so they figure out who she is and what she has already been through. Then they troubleshoot, working out what actually happened between the promise that was made to her and the bill that she finally received. And then they respond in a way that makes it right and makes the customer feel valued. That's a hard call even for a support rep who's been in the company for many years, and the agent now has to get it right every single time or there's a good chance Maya discontinues the service. Let's see how an agent does exactly that.

21:42This is Halcyon's care agent. Rohan showed you the profile every skill gets. Every agent gets one, too, describing what it does, what are the system instructions powering it, what are the artifacts it's connected to, and every activity or session that is registered. You can see that this agent is connected to Salesforce, so it can look Maya up, see her cases, but nothing in Salesforce really tells you why she got $40 instead of 50. So the agent needs two more things. The first is a map of the business, which is an ontology, and the second is a playbook for making things right, which is a skill. Those are the two artifacts this agent runs on. Let's start with the map.

22:28Every company already has relationships sitting inside its systems. So Salesforce knows a customer and billing knows that bill has lines. But you need a single system that connects all of it together. An ontology sits one level above those systems and writes that down in the words that business actually uses. So for Halcyon, a customer has an account and a loyalty membership, and an account has support cases. Each case holds every call and interaction with the customer and a promise that was made to the customer, and that leads to a credit. The credit follows a tax rule, and then it lands on the bill that the customer receives. This is just how the business works. There's no customer data in here.

23:11The case still lives, lives in Salesforce, and the credit still lives in the billing system, and the map is what connects the two. And if I click on Credit, you can see the agent has a way to track what was approved, how much tax came off, and what is actually posted in the bill. So the agent follows this map straight to the right record. Now the playbook. You saw skills in Don's demo, and all these skills belong to Halcyon Support and Care Team. Here is a care tone skill. It says three things: explain what happened in plain words, add $15 if you made the customer contact us more than once, and talk to them like the valued customer they are.

23:50The care team wrote this once, and any agent that loads it follows those same tools. Now let's look at activity. These are the real sessions or traces captured as the agent's activity. This is the actual run of the agent answering Maya. The first thing it does is observe. It looks Maya up in Salesforce and, um, gets her records, gets the case details as well, where you can see three calls in three weeks and a note saying that she was promised $50. Next, it troubleshoots. Salesforce has no idea what actually reached her bill, so the agent follows the map from her case to the credit, and this is where it finds the answer. $50 was approved, yes, but $10 was Ohio state tax that came off.

24:39So $40 was, was what was posted to her bill. The credit was posted after tax, but nobody told her. But the agent found this by traversing through these context artifacts. And finally, it loads the care skill, and it responds to Maya. It writes to her apologizing for the-- for making her call three times, explaining the tax adjustment, adding $15 as tax-- as credit to her, and thanks her for six years with Halcyon. So with the ontology, the tools, and the skill, the agent got Maya the right answer in its interaction with her. All this was just Monday, and this was one agent on one good day. Halcyon has eight agents, and they add pretty much one new agent every month.

25:24On Tuesday, a storm takes over, takes the whole network down. Loyalty decides that affected customers should get $25 instead of 15. They put that in a slide deck. Finance changes the refund policy, puts that in a PDF. And the skill our agent just used still says $15. The business keeps changing. And all these eight agents, your actors, need the rules as soon as they get approved, which is why you need to manage all of these context artifacts like code, the way that Rohan just showed you with skills. And you do this along with the actors, the agents, and the people who write these playbooks, and their activity, which is every place in one place. We'll now launch a poll.

26:10If you want Atlan Trident for your team, tell us there, and the Atlan team will reach out. While you do that, Halcyon's leaders ask an analytics agent or a talk-to-data agent a question like, "How many customers had a poor CSAT this month?" And the agent needs additional context artifacts, and that's where Zach picks up. Over to you, Zach.