Agentic Layer for Developers

I have the pleasure today to dive deep on the agentic layer we introduced in yesterday's Vision Keynote. Now, the speed of generative AI technology has moved at a pace we've never seen before in history.

We feel it. Tom mentioned his nervous excitement. I feel it every single day as I'm coding. There's a new agentic client every single quarter. ChatGPT, Codex, Claude, Claude Cowork, Gemini, Copilot, Copilot Cowork, a new model, a new model provider, a new protocol.

We made a bet early on not to tie you to a single agentic client, model, or protocol. The agentic layer in agents your way means you pick the agentic client of your choice. You pick the tools that best serve your business, or you pick the native agents purpose -built for the office of the CFO.

We deliver on this bet with three tightly integrated pillars that make up our agentic layer. Number one, the Agentic Gateway. The mandatory entry point for all agentic traffic into the OneStream platform.

Number two, the agentic toolkit. The new composable surface for agents over the OneStream platform. And number three, our native agents. Purpose-built for the office of the CFO. The Agentic Gateway is a foundational piece of infrastructure we have built and will continue to build.

Every agentic request, no matter who or what it is, is going to pass through the Agentic Gateway. It's what gives the rest of this infrastructure enterprise-grade security and performance. Now, it sits on top of the OneStream platform.

And we built it with four key properties in mind. Number one, it's universal. It's the mandatory entry point for all traffic into OneTream. And it's the routing layer to send agents to the right back-end service.

Number two, it's secure. And we get asked this a lot. Identity and access resolution happen at the edge. What does this mean? Every request is validated. Every context is resolved. Authorization policies are enforced before any back-end request is made.

So, if you're a user interacting with the OneTream platform through our Windows application or our web application or through an agent, you get the same security and access policies applied. Number three, it's reliable.

Rate-limited and budget enforcement were built into the gateway. This means as you scale it into production, you do so with confidence. And finally, number four, it's adaptive. Today, the industry is converging on MCP as the primary protocol.

But we already see different protocols coming. We see agent-to-agent interactions being the next thing that will happen. And that's fundamentally going to change how agents interact with platforms. No longer will they call tools.

They'll talk to each other to get work done. The gateway will absorb all of these changes for you. This is what will allow you to continue to bring the OneTream platform everywhere. This will result in every request being secured, governed, and cost-controlled.

Now, this is, I think, one of the most exciting things that we're introducing here at Splash and the agent-to-agent. It's the agent-to-agent toolkit. It's a completely new composable surface over all of OneStream for agents.

It represents the repackaging of all the IP we've been building for many, many years. The platform capabilities, the financial intelligence, the semantic layer, all exposed as primitives or building blocks for agents to interact with.

It's how we scale OneTream in the era of agents. Let's dive into it. There's five core primitives that make up the agent-to-toolkit. MCP tools, MCP apps, MCP, or the semantic layer, skills, and plugins.

And before we dive in, people ask me, what is MCP? MCP, or model context protocol, it's how agents communicate to systems. The same way that our websites communicate to systems through APIs, think about that for agents.

Now, MCP tools, they're the verbs of the platform. They're the purpose-built functions we expose through MCP, like searching dimensions, getting application context, executing data queries. We purpose-built these from the ground up to how agents and LLMs want to interact with data and do so in the best way possible.

Number two, MCP apps. You heard Drew talk about this yesterday with the forecast agent. How do we bring the dashboards of forecast into ChatGPT, into Cloud, into Gemini? MCP apps. They're how we surface these dashboards you built in OneStream and your third-party agent to clients.

They allow us to expose the UIs that you've invested in as well. It's what gives the audibility and transparency of OneStream, but in Copilot, in Cloud, in the agent to client of your choosing. The semantic layer, I think, is where we've built most of our IP.

It's what keeps every agent grounded in your OneStream application. It's how they know what user-defined dimension one versus user-defined dimension two is. When you ask for North America, are you asking for the entity or are you asking for the region? It's how our agents learn the nuances that our internal delivery teams have learned over years of working in OneStream.

Without the semantic layer, our agents don't know. But with a semantic layer, our agents, your agents, they understand the custom OneTream implementation and metadata you've built. If not, it's like talking to a confident stranger.

They'll tell you something. They won't be right. Skills. What are skills? We used to talk about prompt engineering in large language models about a year or two ago. What does that mean? We're teaching LLMs how to behave, how to act.

Skills are how we take those prompts, the words we've written, the instruction sets, and we package them. The same instruction sets that we've given our native agents, we package them and we can give them to you.

So your agents can understand OneTream just as well as our native agents. And then there's plugins. It's how we bundle everything. It's how we take the tools, the apps, the semantic layer, the skills.

We package them all together so you can download them into your agents. What brings all this together, our agentic toolkit, is we build, we test, we maintain every layer against internal data sets. Every primitive is version controlled.

Regression tested. We manage all the layers of intelligence and you inherit all of the engineering IP we've been building over the years. Regression tested. I'm going to bring back Charlie Nummer, who you met yesterday.

I want to show you what the agentic toolkit looks like today in action. Take it away, Charlie. Awesome. Thank you, Fed. Now we're going to hop in and take a look at how our third-party agents are now able to leverage our OneStream agentic toolkit.

So we're going to start here by clicking in to see our OneStream CubeData MCP directly within Claude. And as you can see here, we now have seven new tools that create two core capabilities for our third-party agents.

They now have the ability to both search and execute existing OneStreamCubeView reports, as well as the ability to pull any slice of OneStreamCube data. These were purpose-built to ensure that you get the highest level of accuracy when interacting with your OneStream data via third-party agents.

We can also take a look at the OneStream fundamental skill that comes shipped alongside these MCP tools. This skill provides all the necessary instructions and context around core OneStream concepts like cubes, scenarios, and dimensions, as well as ensuring that these tools, or I guess that we provide the instructions on how to best leverage these specific set of tools.

If we pop out, we can ask Claude a question here. We're going to ask to please run a trend analysis of our product lines over the past six months. Awesome. And as we ask this, Claude might come back with a couple follow-up questions, so we'll give it a second here.

And for the sake of the demo, we're going to do the gross margin account. For the time period, let's focus on the last six closed months. For the deliverable, we're going to do a chat summary with an included chart.

And we're going to have the data poll from the OneStreamCube. Now, the agent's going to go through and do some work on our behalf here, so let's pop over to another question, an example, and we'll come back to this when it completes.

So we now find ourselves directly within PowerPoint. And here we have our Claude add-in enabled with those same set of OneStream tools. And we started off here with a completely blank PowerPoint deck and wanted to ask Claude to create us a monthly product review showing the key metrics for each of our product lines, the top five SKUs within each of those lines, and then how they've been trending over the past six months.

And now Claude having access to our OneStream semantic layer, it's now able to go through and search for all the specific application context that we've defined, all of the relevant cube views, cubes, dimensions, everything that's needed for the agent to go execute accurate data pools.

Using that information alongside the new tools that we've given the agent, it's now able to go through and search your dimension hierarchies, traverse your metadata, and finally execute a number of data pools all on your behalf.

Using that information, the agent was then able to build us a complete board-level deck, including all of the key information that we requested in the upfront prompt. Along the way, each of those individual data pools flows through our OneStream agentic layer, ensuring it follows your existing security and access model.

This should start to show you the power of Claude, especially here within PowerPoint. However, it's not just limited to PowerPoint. Claude also has add-ins with the OneStream tools in Excel and Word and a vast range of other features.

And if we click back, we can see if our previous question is completed. It looks like it just finished up. And we now can see that Claude was able to generate us this highly rich visualization, showing how our product lines have trended again over the last six months.

From here, we could continue iterating, asking follow-ups, drilling deeper. We could even take this information, share it up with our management team. But really, this should start to show you what the future of work can start to look like for your teams, when you guys can start to leverage the OneStream cube data directly from third-party agents to get your work done fast and efficiently.

And I'll now pass it back to Fed to talk about our native agents. Thanks, Charlie. Give a round of applause. And again, what you just saw there was the agentic toolkit and the power of our new agentic layer.

And what's awesome about that, what's kind of fascinating, what's going to happen is as the industry expands, as the models get more intelligent, more capable, we're going to be in lockstep. You're going to be in lockstep.

You're going to be able to bring the power of the OneStream platform anywhere. Now, moving on to pillar three. The gateway in the toolkit represents what happens when we expose all the IP we've been building externally.

But what happens when we take that same IP, the tools, the skills, the semantic layer, and we build the system from the ground up, and we tweak everything? You get our purpose-built agents for the office of the CFO, our native agents.

Now, I want to revisit our next generation of native agents that we introduced in yesterday's Vision Keynote, capable of solving full workflows before we dive in. Number one, you have finance analyst, enabling every user not just to build reports, but to automate analysis workflows.

Number two, search, helping your users navigate the platform and the custom processes you guys have set up. Number three, deep analysis, automating manual document analysis that we do every single day in finance.

And number four, which we introduced for the first time yesterday, our forecast agent, exposing all of sensitive way I forecast through an agentic interface. But how do we do all this? It's my pleasure today at Splash to introduce a completely new agentic runtime and harness we built from the ground up.

And before I dive in, what is an agentic harness? It's what sits around a large language model in an agent. It's how an agent is able to communicate to systems, get data, output a report. It's what gives us its arms and its legs.

And we focus on four core key components when we rebuilt this harness from the ground up. Number one, memory. Ensuring our agents over long-running tasks, when it's hour two, three, four, five, they remember what you said four hours ago.

They remember the goal of what they're doing. But not just that. They'll begin to learn your user preferences over time. Meaning, our agents will grow with your team as you scale your usage. Number two, sub-agents.

Allowing our main agents to spin up sub -agents to do smaller tasks of work for them. It's how we ensure our main agents don't get what we call context rot. They don't get confused with every single little detail that they're doing for you.

They delegate so they can stay on task and do the work on your behalf. Number three, open protocols. And you just saw what happens when we expose our own IP through open protocols, through our agentic toolkit.

But we're also allowing our own native agents to leverage open protocols like MCP to interact with other systems. So if you have a compatible ERP, HR system, IT system, our native agents will be able to start communicating to those systems as well to do work on your behalf.

And finally, checkpoints. And really, deterministic checkpoints. Every agent requires tracing their actions. Replayability from certain spots so we can tune their behavior. And resumability so that when an agent pauses, comes to a human and asks for a question, more context, it can pick up where it left off.

We didn't just stop there, though. We took a really strong stance on auditability and transparency with our native agents. A CFO today cannot sign off on a variance commentary if we cannot see how we got to that conclusion.

Every agent, every action, every tool call, every output is traceable just like core OneStream is today. And finally, we built an evaluation framework from the ground up. Before you roll out agents, you have to test them like any other AI.

And when you roll them out, you have to see how they're doing in production. The same lessons we learned when scaling sensible way of forecast, we brought into the agentic era. We know what this means.

And not just that. We allow our customers now to bring their own data sets, their own tasks, their own questions, so we can test our agents before we deploy them for you. Now, I want to show you what this new agentic harness and runtime can do with our longest-running agent, Finance Analyst.

The new generation of Finance Analyst allows you to schedule hundreds of agents. They can conduct variance analysis every single morning for you before you walk into the office. They can do flux analysis on day one and close.

Or they can run driver investigations across hundreds of entities. I'm going to hand it back over to Charlie to show you our new Finance Analyst agent in action. Take it away, Charlie. Awesome. Thank you, Fed.

Looks like we have the wrong one up here. Give me one second. Sorry. Looks like we loaded up the wrong demo. There we go. Awesome. Thanks, Fed. Now we're going to hop in and take a look at the next generation of Finance Analyst.

Our Finance Analyst agent is designed to allow users to query and analyze their one-stream financial data. And throughout the private preview process, we learned that although users get a lot of value out of the manual, ad hoc question and answering you might typically have with a chat assistant, a lot of times those questions actually came as a part of larger flows like variance or flux analysis.

So with the next generation of Finance Analyst, we have now extended the capabilities to now allow businesses to automate hundreds of recurring flows to execute in the background on their behalf. So let's hop in and take a look.

We start here on the Analysis Plans page. This is where users can create and manage each of their individual analysis plans. Let's click in and go through the process of creating a plan. We start here on our Planning Agent.

This is where users work iteratively with the agent to help create an analysis flow that perfectly captures exactly what they're trying to go through. And we can start here by giving an outline of the analysis, stating we want to build a monthly variance analysis of our North America Equipment Division, starting with our standard income statement versus budget.

And as we kick this off, similar to how a new hire would operate, the agent's going to start by going through and searching your OneStream application for relevant information like your cube views, your cubes, your specific dimensions, and then using that information to start to generate a highly contextualized analysis flow.

Along the way, we might come back with some follow-up questions, which we'll go through here. We're going to focus on the last closed month for the time period, so we'll click that. We're going to go ahead with the default materiality.

Awesome. And as we submit that, the agent's going to use the information we just answered in those follow-ups, as well as the context it gathered about the application, to start to generate us this highly detailed analysis flow.

This is an extremely important part of the process, as it ensures the agent has all the proper context up front, to ensure that when it goes and executes on a recurring basis, it does so the same way each time.

From here, we can scroll up and take a look at what's included in the plan. And we're going to see we have the high-level objective, going down again into some of the key assumptions. And as we go deeper, you're going to see it get really detailed, going into the structural setup, the specific investigation approach and flow we want the agent to take, all the way down to the specific output requirements.

And along the way, as you're reviewing this plan, if there's anything you want to change, you can come over, make any of those adjustments directly in natural language, and the agent will fix it up for you.

From here, users have the ability to set up their analysis plans to run on a schedule. Currently, we support the ability for time-based schedules, things like the first of each month, but we'll also be introducing workflow-based triggers, like having these execute at the end of a closed process or consolidation.

And if we head over to activities here, we can click in to take a look at a report that we had generated. So we'll open this up. And at the top here, you can see a summary showing each of the underlying cube-level drivers we surfaced for this specific analysis.

As we go down, again, we can see the process the agent took, starting with that top-level income statement, as the agent analyzes it and finds variances, going down into some additional drill-downs. We just saw the accounts there, the operating expense drill-down.

And along the way, not only including the tables, but again, also including some rich commentary that we can review. And at any point, we can click on any of these blue citations to take us directly to the one-stream data pool that was used to generate that insight.

This is a critical part of the process as it ensures users have that traceability and auditability into exactly where their numbers are coming from within one-stream, a critical thing in the office of the CFO.

And if we then come over to the left-hand side, we can see users also have the ability to ask any additional follow-up questions or drill deeper into any of the insights found in the report, making for a very interactive experience and process when reviewing these reports.

From here, this should start to show you where OneStream believes the future of finance is going, where your team can set up hundreds of recurring analysis flows to execute in the background on your behalf, allowing your teams to focus on steering the business and interpreting data instead of just pulling numbers.

Thank you, and now I'll pass it back to Fed. And I just want to pause for a second because it makes me very excited to see, you know, all the work we've been building, the foundations come together. What you saw there is one iteration.

One iteration of finance analysts set up to solve a monthly variance analysis on income statement. You can set up hundreds of scheduled versions of our native agents to run on behalf of your business, on behalf of your users.

I'm super excited to give you just a sneak peek of what we're building. But with our new agentic layer, you have a completely new, secure, governed entry point into OneStream for your agents, a whole new toolkit to leverage the OneStream platform in the agentic era, and our latest purpose-built agents for the office of CFO to help automate all your workflows.

Come to the innovation hall. Let us show you not just in Claude, but in Copilot, Cowork, and other providers everything that we can do on behalf of your people.