AgentForce's Atlas Reasoning engine classifies what you're asking for, then runs actions like retrieving calendar data or generating schedules using your CRM information. It checks guardrails before responding and shows you how it reasoned through each step in the builder UI.
In this demo, we'll explore how the Atlas Reasoning engine works. Because every time you ask Agent Force a question like what should I focus on Today? It's using the Atlas Reasoning engine to produce this outcome. So to trust this response, let's take a look inside. This is the Atlas Reasoning engine, and this is what enables AgentForce to respond agentically by reasoning through situations and autonomously deciding on the best course of action. Let's break it down step by step. It starts when AgentForce is invoked either manually by receiving a user's message or automatically from a different processor system. Then it will use the Large Language model or LLM to truly understand the intent of the question and classify the topic. Let's take a look at an example. Here we are in agentforce Builder and and here are the four topics this agent supports. So when I type a question, the reasoning engine will pass my question with the name and description of each topic to the large language model. And we can see it correctly identified that our question is about the daily schedule topic. But if we try a question that is off topic, the agent will no longer engage. Instead, it will provide a response to let the user know this is outside of its guardrails. So let's take a look at what happens once we've defined our topic. In the next step, the reasoning engine will inject a prompt to send the model the topic's scope, instructions and available actions. They're all configured within a topic in order to decide which actions, if any, are required to satisfy the request. Let's go back to AgentForce Builder to see what this looks like. Within each agent's topic, you can use free text to define the agent's scope and instructions. In the Daily schedule topic, one of our instructions asks the agent to confirm with the user before proceeding any further. So when I ask a question, the agent chooses not to execute any action and instead generates this initial response. Let's review what we just saw. Since our agent chose not to execute an action, we skip down to step seven and eight where it sends the response. After we reply, we go back to the beginning of the flow, confirming we're still on the same topic from the conversation history. And this time, the agent decides to take action. Actions define what an agent can do. They can be flows to create or update records, prompts to provide answers for massive amounts of data, or even API actions to trigger external systems. Let's continue with our example to see how this works. Here I built a flow for our agent to access calendar meeting data, which is outside of salesforce using an apex action. That way, when the agent is coming up with the schedule, it knows what is already planned for the day and back. In agentforce Builder, once the flow runs, we can see the flow action output providing our meeting schedule, but that response won't go directly to the user yet. This is because the agent isn't limited to just one action. It's able to loop through multiple actions and reassess based on the new action output. In our example, the agent used the flow to get our existing meetings, but now it needs to go one step further to generate a recommended schedule that incorporates those meetings. To do that, we'll use a prompt action. Let's dive in to see how it works here. In Prompt Builder, I've built a prompt to give instructions on how to create a daily schedule. It's using some of our native CRM data from this Get Prioritized Opportunities flow, and it's also using this retriever to pull in external unstructured data. The prompt can access this unstructured Data through the AgentForce data library, where we've uploaded an example of a sales rep's quarterly goals. Within this PDF, we use a method called Retrieval Augmented Generation, or rag, which is looking through a vector database of small segments or chunks of the quarterly goals and identifying which ones are most relevant to our prompt. So if I preview the prompt, we can see it combining all that data in order to draft the daily schedule as a response. So what we just saw was a second iteration over a reasoning loop. First it used the flow action to get our external calendar meetings, and then it used those as an input to the second action where the prompt generated our daily schedule. Once the agent is satisfied with the action output, it does a grounding check to ensure the response is following all instructions and is within the agent's guardrails. And then it will send the response. And if we navigate back to AgentForce Builder, you can track all this live on your screen as the agent reasons, providing the transparency on the reasoning specific to your agent. Here we can see the agent correctly identified the prompt action and displayed that same output we just previewed. And so finally, the agent will take that output, identify that no more actions need to run, and then communicate that response back to the user. So now you understand how the Atlas reasoning engine works, enabling AgentForce to create truly autonomous agents. And when we type that same question into AgentForce, we know it's performing the reasoning loop to produce this deeply personalized answer that we can trust. Now you've seen how the Atlas Reasoning engine works. To learn more, visit agentforce.com thanks for watching.
