Understanding the Atlas Reasoning Engine Workflow

Workflow - Step-by-Step

1. Agent Invoked / Topic Classification: The process begins when a user sends a message or query, or when an agent is invoked via an event, data change, or API call. The reasoning engine analyzes the user's message to classify it under the most relevant Topic, primarily looking at the topic name and classification description. If no appropriate topic matches, it defaults to an "off-topic" classification
2. Context Assembly / Scope, Instruction, & Action Injection: Once a topic is selected, its Scope, Instructions, and associated Actions are injected into the prompt, alongside the original user message and conversation history (typically the last six turns). This combined prompt is then sent to the LLM to determine the agent's next move
3. Agent Decision: The agent analyzes this combined input (user message, instructions, potential actions) and decides on the next step:
   1. Run an Action: If specific information is needed or a task must be performed. 
   2. Respond to User: If the agent has sufficient information or is seeking more details.
4. Action Execution & Observation: If an action is chosen, the agent invokes the selected Action, and the output or result from that action is captured
5. Reasoning Loop: The action's output (all the outputs that don't have "Filter from agent action" selected) is sent back to the LLM and combined again with the original context (instructions, actions, conversation history). The agent then re-evaluates the situation based on this new information and the ongoing conversation. It decides whether to run another action or formulate a response. This loop continues until the agent determines it has all the necessary information and is ready to respond directly to the user
6. Grounding Check: Before sending the final response, the agent performs a crucial grounding check to ensure the proposed response adheres to the provided Instructions for the Topic. This step verifies that the response is:
   1. Properly grounded in source information. 
   2. Adherent to the topic’s instructions and scope. 
   3. Free from hallucinated or unverified information. 
   4. Protected from potential prompt injection risks. If the grounding step fails, the agent will retry to produce a grounded response. If it cannot, it will send a standard message indicating it cannot help with the request
7. Send Response: The final, validated response is sent to the user

Example: Changing an Order's Delivery Address

Let's walk through a complete example of how these components interact when a customer wants to change a delivery address:

Customer Message: “I ordered a red sweater yesterday but I need to change the delivery address.”

Reasoning Process:

1. Topic Classification:

• The engine compares the message to all topic names and classification descriptions.
• It selects the Order Management because its classification description mentions "changing orders".

2. Context Assembly:

• The reasoning engine brings the "Order Management" topic into focus, including its scope, instructions, and available actions, along with the conversation history and the user message.
• The scope might indicate that the agent can modify orders but can't cancel orders after they’ve been processed.
• The instructions include guidance on handling order modifications, such as identifying the order first.

3. Action Selection:

• Based on the topic instructions and action descriptions, the engine determines it needs to first identify the order (using a "Find Order by Description" action) and then update the shipping address (using an "Update Order Shipping" action).
  

4. Execution and Response:

• The engine asks the customer for their email to find the order.
• Once the order is located, it collects the new shipping address.
• It runs the "Update Order Shipping" action with the order ID and the new address.
• Finally, it confirms the address change with the customer

Understanding the Atlas Reasoning Engine's workflow and how to effectively configure topics, instructions, and actions is paramount for building powerful and reliable Agentforce agents. By carefully designing these components, you can ensure your agents accurately understand user requests, make informed decisions, and deliver consistent, helpful responses, leading to successful agent adoption and better AI outcomes