Navigator Finds Any App — No Agent Needed

Toilet. Show my screen. So, as Elliot mentioned, we've worked on.. . adding more capabilities to quick asks.

So a lot of navigator is in the context of quick actions, quick sections and panels. And we saw an opportunity to make it more aware of those. So how this started was that we wanted to pick up on questions that cannot be straight answered, but could be related to certain quick actions that the user has access to. and we would display a list of available actions for that.

So that got us into building an MFE MCP over the open component registry that we have. So this is the registry in pre-prod, where all the MFEs get. published to. They have metadata related to what application ID, they refer to what organisation IDs that are available, to so on and so forth.

We've also seen that there were lots of agents that were built specifically for rendering MFEs. So that was adding into the Quick Asks experience, you would need to do 2 things as a soft core. You would need to build the MFE. as well as build the agent.

So we wanted to make that simpler by potentially... being able to skip the second part. Um... So yeah, if we build this MCP and it's supposed to be generic, we can reuse it in other components as well in the future if you wanted to.

So if I move to the second tab, I'm using MCP inspector here to highlight what it essentially looks like. So I'm connecting here with my bearer token, so everything is authenticated and. The MFEs that I see here are mapped as tools that an AI agent could use to display this information.

All these tools are filtered for the organisation ID I'm in and for the products that I have as a user. So just to pick on a few examples here, with the name of the MFE becomes the name of the tool, the description of the MFE becomes the description of the tool. Now we've added support for system instructions, so this is a relatively new thing.

But if I go into this one, for example, so this MFE defines a system instruction property, and we can use that to enhance the description of the tool to have the AI better understand how it should use this tool specifically. And we also have, we also use the Let me see if I can find this. We also use the parameters from the MFE and we turn them into parameters for the tools themselves.

So for example, in this case, let me see if I can close this. The add new customer, I will demo specifically this one later, has Is a simple MFE for adding a new customer record, and it includes name, business address, and main contact details. So, we have customer name, business address, main contact.

All of these parameters are exposed to the AI, and it can it can populate those just to. move over a bit to this specific MFE. So if I just pull this over, this is from a technical side, how it looks like. So this is the EVO experiment MFE.

It has a parameter section and we're using the parameter schema from the documentation and we're turning these into AI parameters. So to essentially go through the demo of this, I have a few saved prompts. So I can do things like, what can I do in Workday from here?

Now, you'll notice that in the MCP inspector, I have lots of tools. In pre-prod, my account has around 1000 of them, and there are too many tools for an AI to process. So what we've done here, we're essentially capturing all the available tools, but instead of making them available to the agent directly, we essentially collect them and add a tool search tool so that it can look into the bag of available MFEs to figure out which are relevant for the user's question.

So if I say something very exploratory here, like what can I do in workday from this chat interface, right? Because it's sometimes... can not be clear. If I click on send here.

It will look for available apps and then it will list a few of them, which I can use to request absence or do my time sheet or view my benefits, so on and so forth. From here, I can open them and interact with them. interact directly with quick actions. So this is using a special tool for suggesting a list of MFEs.

Now, the agent can invoke a tool directly, just as they would any other tool. So if I do some... If I say something like, show me my benefits from workday.

There is specifically an MFE that has this, and it's not registered for an agent. But because the navigator chat has direct access to all the MFEs, it can see that, okay, this MFE is related to the user's question. I'm just going to open it in line here.

So this is when one specific MFE from the search result matches a lot on what the user is asking. And for the third, demo here. I'll say, I'll use this prompt, which I have favorited.

So I'll say, help me register this new customer. I provide a name, I provide a business address, I provide a main contact name, and I am specifying here that this is not a financial customer because I have. A few financial agents as well that some of them can register new customers.

So I just want to highlight the fact that I can get to using an MFE with parameters in this case that is not linked to an agent. So in this case. It just opened my add new customer MFE with data pre-filled in on the all corresponding slots where it had the data, right?

I didn't provide any main contact e-mail, so it left that blank, and it even specifies at the end that I'll need to add e-mail address. if I have it to submit the complete registration. This is all just because we have the correct context for the AI to do its job. So this is part of the first demo.

I'll quickly go into the second one, but please pose questions in the chat if there are any. For the second part, we've received feedback that The suggestions we can configure currently in Agents Builder. are not granular enough. So if I go into, so what we've done this week is being able to configure suggestions per application role.

So if I go into this Alex demo agent, if I click on edit. I'll notice that I've linked this agent to collaborate. And this then allows me to configure suggestions based on user roles.

So I have a few here just to sort of highlight the specific text here. But when I add a new prompt, I need to define a label, which is the short text to be displayed in the UI. The question, that is the actual question that is sent to the LLM.

And I have a dropdown here that essentially goes to the available roles for this specific application. In this case, I have user and administrator, but if your application has different roles, they will be available. They will all be available here in the dropdown.