How does Salesforce's AgentForce work? So, we've heard Salesforce rant and rave about AgentForce since Dreamforce last month, finding every way to tell us how brilliant it is and how it has the power to revolutionize the AI space as we know it. But we've heard much less on how it actually works.
If you weren't one of the lucky 5200 customers invited to the AgentForce launchpad at Dreamforce and didn't get the chance to try out AgentForce for yourself, don't worry. We'll give you the lowdown on what this tool is for and how it really works. So, let's get stuck in. In this video, I'll walk you through the main components of what makes an agent function, taking insights from the Dreamforce 24 Keynote demo, where Sophie Sachs' service agent was put to the test.
Agent Description So, where do you build your agents? That would be the agent builder. which is part of the AgentForce studio. The configuration for the agent is the key starting point. This is done in natural language, the way you would have a conversation with another human.
And Salesforce says if you can dream it, AgentForce can do it. On screen now we can see the interface where the avatar, the agent name and the description are set. Channels are the ways that humans and agents communicate with each other, for example email, voice and WhatsApp.
Notice that there are four sections to the agent builder interface. The left hand sidebar to navigate between agent settings which we will walk through in this guide. The left panel showing the setting page you're currently on.
The right-hand panel enabling you to start a conversation to test drive your agent. And the middle panel where the test drive output will be displayed. Agent Topics Topics are the foundational building blocks for agents, determining the scope of what they can do. Another way to think about the topics is categories for information.
For example, an order management topic that could enable the agent to have access to order history data and change the product specs. Take a look at the agent's topics in the image on screen now. In the Dreamforce 24 keynote demo, the agent was able to produce answers and outcomes for some of the customer's queries.
However, some were not able to be resolved because the agent didn't have these topics assigned. This is part of the concept of guardrails which is what the agent cannot do. When inspecting a topic's configuration, you can see the description and scope of the topic as shown here.
Topic Actions. Alongside topics, actions are the building blocks for agents. Actions are tied to topics. Here we can see some of the actions that the agent already has assigned.
So what they can do? These are actually flows. View agent for actions from Salesforce setup and from the topic, see the assigned actions from the sub tab as shown here. Assigning additional actions that are available in your org is as simple as sticking the actions from the list that appears in the pop-up window. When it comes to creating a new action, a pop-up screen requests you to add a reference action type such as Apex, Flow, a Prompt or MuleSoft API.
Then you need to reference the library of processes and APIs that have already been created. In this example, the new agent force action will need to be called out via MuleSoft as this source of data is not integrated directly. Choose the MuleSoft API and the DataSource API. What's notable is that now this agent query can talk to these APIs just like human users can.
Upon adding the API, there's the chance to edit the inputs and outputs. Atlas Reasoning Engine Testing your agent. Back at the AgentForce Builder, you can utilize the center and right hand panel to test drive your agent. This takes the Atlas Reasoning Engine out of the shadows and into the user interface.
Upon giving a prompt, follow along with the agent's reasoning step by step. Here's what it will do. First, it identifies the relevant topic. Then, it will see the actions taking place, querying the CRM database via flow. From there, it returns the correct records.
Lastly, it reasons or confirms that the response is accurate, in other words, grounded. Upon deploying this agent, it will be able to function using the channels listed in the agent's description. You don't need to repeat this process for each channel.
Omni Supervisor You may already be familiar with Omni Supervisor, originally a feature tied to the service cloud for managers to oversee teams of customer service agents. Now, Omni Supervisor is being repurposed for agents. See overall trends and agents working in real-time. The previous few interactions are shown in the image here. You can even listen in on ongoing or recently closed conversations, labeled with a summary column in the table.
How agent force and data cloud work together. Now we get into the true nuts and bolts that power agent force. Data.
Data that can be used to train agents could be either structured, example, a Salesforce record, or unstructured, example, emails or voice memos. The vector database in Data Cloud makes processing unstructured data possible. Data Cloud has become the underpinning for the Salesforce platform.
In simple terms, it gets the data flowing between the various Salesforce apps, cloud products. AgentForce is a layer on top of these apps, catering to the use cases that these apps champion. So, does Data Cloud really have an impact when it comes to agent force? See the customer perspective here.
The agent can't answer the customer's query and instead answers with, sorry, I had trouble coming up with a response, which behind the scenes is due to a lack of data to reference that's relevant to the query. Retrieval Augmented Generation, RAG. It doesn't have to stop there though. You can bring in large language models to learn about your organization through prompts, which is a set of instructions sent to an LLM to teach it.
Salesforce uses a technique called Retrieval Augmented Generation, or RAG, that lives inside the Atlas Reasoning Engine and produces a feedback loop with Data Cloud. A request made by a user or agent, a prompt, becomes a more contextual and relevant augmented prompt once Data Cloud and RAG work together. By searching through all of the data that's been sorted by Data Cloud, the output from the prompt improves and in turn the AI, the LLM, learns more about your business.
New Data Streams As we saw with AgentForce demos, the data that is required for highly contextual responses may not be stored directly in Salesforce. There are a few options on how to bring this data into AgentForce to make it usable. Data Cloud Ingestion Set up a new data source to bring data into Data Cloud on a scheduled basis.
Zero copy. Set up a new data source to virtualize record data, meaning that you're not physically moving and storing the data from the source system onto the Salesforce platform. MuleSoft APIs, which we saw earlier in this video.
An example of this could be that you would like to connect your order management system data that resides inside of Snowflake, which happens to be an out-of-the-box connector offered by Salesforce. Data Graphs. Data graphs enable you to visualize the relationships between data model objects, DMOs, even going several layers deep.
This helps you to investigate whether the correct data is available at the time of writing a prompt or instructing an agent force agent, which will be required to generate an optimal and accurate output. Real-time data graphs perform faster identity resolutions, segmentations, and actions based on ingested data to ensure data is ready for agent force to work with. Here, you can see the data graph showing how every data touchpoint is connected.
Inside Prom Builder. Under the hood, an action, as we saw earlier, is sourced from a prompt. You can create a new or edit an existing prompt to make it more intelligent. Pick your choice of LLM model that will operate under the hood of prompts. You can see this in the right-hand panel.
Prompt engineering in Salesforce is low-code, thanks to the clean user interface that guides you through a preview and feedback toxicity ratings. Search Index Retrieval Augmented Generation is the reasoning part of the Atlas Reasoning Engine. In other words, the brain behind Agent Force.
To improve the results retrieved from the wealth of data in Data Cloud, you can set up a search retriever. This retrieves the correct data from your connected data streams and then grounds your prompt with this wealth of data. assign search parameters within the prompt. There are three types of setups you can choose from. Easy setup, advanced setup, or from a data kit, a package of metadata and data relationships that can be installed into Data Cloud.
In short, Data Cloud makes every agent better. To finish off, take a look at the handy diagram that was featured in the sales agent session. This is another visualization of how these components, some familiar, some unfamiliar, all stack together. While Salesforce are doing what they can to hype Agent Force up, there's no question that the stats spell promise for Agent Force going forward.