Скачать или смотреть Salesforce Atlas Reasoning Engine of Agentforce

Salesforce Atlas Reasoning Engine of Agentforce

The Atlas Reasoning Engine is the “brain” behind Agentforce, the agentic layer of the Salesforce Platform. Built on a proprietary system designed to simulate how humans think and plan, it uses multiple techniques, including advanced retrieval augmentation generation (RAG), to analyze information and determine the next-best actions for completing simple or complex‌ tasks. Like other reasoning engines, it ensures that AI agents can not only follow instructions, but also think proactively about the best way to translate those instructions into accomplishable goals.

How Does It Work?
To deliver reliable assistance, Atlas relies on a few key concepts.

Reason–act–observe Loop (ReAct): Atlas uses a cycle of reasoning, acting, and observing. It thinks through a step, takes action, checks the outcome, and loops until it completes the user’s goal. If needed, it even asks for clarification mid-task. This keeps conversations natural and accurate.
Topic classification: Every request is mapped to a topic, which serves as the AI’s job description for that moment. Topics define what actions are valid and which policies apply. This helps Atlas scale safely across thousands of scenarios.
Richer conversations: Unlike earlier bots, Atlas allows Agentforce to respond in natural language, handle follow-up questions, and pull from past conversational context. The result? A smoother, more human-like interaction.
Visible reasoning: Atlas shares its "thoughts"–the reasoning behind why it acted a certain way–so that admins can review decisions, prevent errors, and build trust. You can access those insights in Agent Builder.