The Atlas Reasoning Engine Architecture
An architectural blueprint of how Salesforce Agentforce utilizes the Atlas execution loop to plan, refine, and orchestrate complex autonomous workflows.
Deep Dive: How Atlas Orchestrates Decisions
At the core of Salesforce Agentforce is the Atlas Reasoning Engine. Atlas is responsible for parsing user queries, selecting action tools, validating state changes, and delivering grounded responses.
Unlike typical LLM agents that run a simple “re-act” loop, Atlas separates planning, context retrieval, tool selection, execution, and verification into modular phases:
graph TD
A[Query Received] --> B[Phase 1: Intent Parsing & Semantic Grounding]
B --> C[Phase 2: Execution Planning & Tool Selection]
C --> D[Phase 3: Security & Guardrail Filter]
D --> E[Phase 4: Tool Execution Flow/Apex/MCP]
E --> F[Phase 5: Output Validation & Self-Correction]
F --> G{Goal Satisfied?}
G -- No --> C
G -- Yes --> H[Phase 6: Final Customer Response Synthesis]
Phase 1: Intent & Semantic Grounding
Atlas uses semantic search against your Data Cloud indexes and metadata descriptions to locate contextual records.
Phase 2: Action Selection
Atlas searches your Topic and Action libraries to find suitable tools. Descriptions matter:
- Poor Action Description: “Runs script”
- Excellent Action Description: “Retrieves customer shipping address, tracking numbers, and delivery date by Order ID”
[!IMPORTANT] The reasoning engine matches user query embeddings against tool metadata descriptions. Up to 80% of agent routing errors are resolved by improving invocable method and variable descriptions.
Frequently Asked Questions
How does Atlas differ from RAG?
Retrieval-Augmented Generation (RAG) simply retrieves facts to ground LLM outputs. Atlas goes beyond by running actions, validating outputs, checking safety boundaries, and re-planning dynamically.