Quick Navigation
Atlas Reasoning Engine
- Atlas Reasoning Engine
- The AI brain of Agentforce that evaluates user utterances, selects the matching topic, plans the action sequence, and executes it to deliver context-aware responses.
- Utterance Processing Pipeline
- User utterance → Atlas evaluates intent → topic selected → action sequence planned → actions execute → response returned. Atlas selects topics; it does NOT execute actions directly.
- Topic Selection
- Atlas matches the user utterance to the most relevant configured topic using the Classification Description; only topics with a matching classification will be invoked.
- Action Planning
- After selecting a topic, Atlas determines which actions to run and in what order to satisfy the user's request, using the topic's instructions as guardrails.
- Classification Description
- A text field on each topic that tells Atlas when to route to that topic; poor classification descriptions lead to wrong topic selection.
- Deterministic Behavior Controls
- Filters and variables added to a topic that constrain agent behavior to produce predictable responses for specific scenarios rather than pure LLM reasoning.
Agent Types
- Service Agent
- External-facing agent that handles customer service inquiries across channels (web, Messaging, email) without preprogrammed scenarios; uses Atlas to reason about service issues.
- Sales SDR Agent
- Engages prospects 24/7, answers questions, manages objections, and schedules meetings based on CRM data to scale the top of the sales funnel.
- Sales Coach Agent
- Provides personalized role-play sessions to sales reps using Salesforce deal data, helping them practice pitches and handle objections for specific opportunities.
- Employee Agent
- Internal-facing agent that acts as a role-based peer/collaborator, accessing the same data and org settings as each employee for internal workflows.
- Service vs Employee Security
- Service Agents are restricted to customer-appropriate data; Employee Agents can access internal data matching the employee's own permissions — security models differ.
- Standard vs Custom Topics/Actions
- Standard topics and actions are pre-built for each agent type; custom topics and actions are user-created for unique requirements not covered by standards.
- Agent User
- A special Salesforce system user that the agent runs as; permission sets on the Agent User control what data and actions the agent can access — separate from the end user.
Topics, Instructions, and Actions (TIA)
- Topic
- Defines a job the agent can perform; contains a Classification Description (when to use), Instructions (how to behave), and one or more Actions (what to do).
- Instructions
- Text rules inside a topic that set decision-making boundaries for the agent; guide Atlas on how to handle edge cases and constrain response scope.
- Action — Invocable Apex
- Action type backed by an @InvocableMethod Apex class; requires 75% Apex test coverage just like any Salesforce Apex code.
- Action — Autolaunched Flow
- Action type backed by an autolaunched Flow; allows no-code or low-code action building using the Flow Builder. Most common action type for admins.
- Action — Prompt Template
- Action type that executes a Prompt Builder template; allows agents to generate AI content or summarize data as part of an action sequence.
- Action — MuleSoft API
- Action type that calls a MuleSoft-managed API; used when agents must interact with external systems through MuleSoft's integration layer.
- Action — REST Apex
- Action type backed by a REST-annotated Apex class (@RestResource); used for exposing custom REST endpoints as agent actions.
Agent Builder and Channel Configuration
- Agent Builder
- The no-code workspace in Salesforce Setup for creating, configuring, and managing AI agents; used to configure topics, actions, channels, and security settings.
- Channel — Web Chat
- Deploy an agent to a website or Experience Cloud site using the embedded chat widget; the most common channel for Service Agents.
- Channel — Messaging (SMS/WhatsApp)
- Connect agents to Salesforce Messaging for SMS and WhatsApp interactions; requires a Messaging channel configuration in Setup.
- Channel — Slack
- Connect agents to Slack workspaces so employees or customers can interact with agents inside Slack; common for Employee Agents.
- Agent User Permission Sets
- Must be explicitly configured for each channel context; connecting an agent to a channel does NOT automatically grant the Agent User the required permissions.
- Apex Action Test Coverage
- Custom actions built on Apex classes require 75% code coverage in unit tests before deployment — the same Salesforce platform requirement as any Apex.
Prompt Builder and Template Types
- Prompt Builder
- Salesforce tool for creating, testing, activating, and managing reusable AI prompt templates; templates can be used as agent actions or invoked directly by users.
- Sales Email Template
- Generates a personalized email draft using record data; appears in Draft Email actions on record pages once activated. Tied to a specific email use case.
- Field Generation Template
- Auto-fills a specific object field using AI; tied to a single object and field and outputs structured content saved directly to the record.
- Flex Template
- General-purpose template not tied to a specific field; supports up to five custom inputs, suitable for any text-generation use case not covered by other types.
- Template Activation
- After creating a template it must be explicitly activated before it can be used by agents or invoked from record pages; inactive templates exist but cannot be executed.
- Field Generation vs Flex
- Field Generation writes output to a specific record field; Flex produces output for any purpose and is not bound to a field — do not confuse them.
- Prompt Template Lifecycle
- Ideation → Build → Test → Activate → Deploy → Observe; activation is a discrete step required before any template can execute in an agent or user workflow.
Grounding Techniques
- Merge Fields
- Direct references to Salesforce record field values (e.g., {!Account.Name}) inserted at runtime into the prompt; pulls raw data from the current record context.
- Flow Merge Fields
- References to data that has been processed or calculated by a Flow before insertion into the prompt; use when data needs aggregation, transformation, or cross-object lookups.
- RAG Grounding (Data Cloud)
- Retrieval-Augmented Generation using semantic search against Data Cloud vector indexes; best for knowledge bases and large unstructured content libraries.
- Merge Fields vs Flow Merge Fields
- Merge fields pull raw field values directly; Flow merge fields pull values that have been transformed by a Flow — use Flow merge fields when data needs calculation first.
- Merge Fields vs RAG
- Merge fields are for known, structured record data; RAG is for searching large unstructured content to find the most relevant context dynamically.
- Dynamic Grounding
- Automatically injects relevant Salesforce data into the prompt at runtime; a Trust Layer capability that ensures grounding data stays within the secure boundary before it reaches the LLM.
- Reducing Hallucinations
- Ground prompts in relevant Salesforce data, set a specific output format, include guardrails in instructions, and test iteratively — vague prompts increase hallucination risk.
Einstein Trust Layer
- Data Masking (PII Detection)
- Identifies and masks sensitive data (names, emails, SSNs) in prompts before they reach the LLM using pattern-based detection and ML models; prevents PII leakage to external models.
- Toxicity Detection
- ML models score both prompts and LLM outputs across five categories (violence, sexual, profanity, hate, self-harm); harmful content is flagged or blocked before reaching users.
- Prompt Defense
- Secures prompts against injection attacks before they reach the LLM; works alongside data masking and toxicity detection as part of the Trust Layer pipeline.
- Zero Data Retention
- Contractual guarantee with LLM partners (e.g., OpenAI) ensuring customer data is never stored or used to train external models after processing.
- Audit Trail
- Timestamped log of every AI interaction including the original prompt, safety scores from toxicity detection, LLM output, and any end-user action taken; available for compliance review.
- Trust Layer and MCP
- MCP actions execute within the Einstein Trust Layer boundary; PII is masked before data reaches an LLM even through external MCP connections.
- Data Masking vs Toxicity Detection
- Data masking protects SENSITIVE DATA from reaching the LLM; toxicity detection prevents HARMFUL CONTENT from being generated or shared — they address different risks.
Data Cloud for Agentforce (Data 360)
- Data Cloud / Data 360
- Salesforce's data platform rebranded from Data Cloud to Data 360 in late 2025; the same technology. Provides real-time data unification and grounding context for agents.
- Data Library
- The storage layer in Agentforce managing knowledge sources for agent grounding; sources include Salesforce Knowledge articles, uploaded files (text, HTML, PDFs), and Data Cloud objects.
- Chunking
- Breaking unstructured documents into semantically appropriate segments for indexing; chunk size and method must be configured — poor chunking degrades retrieval relevance.
- Indexing (Vector Embeddings)
- Converts chunked text into numerical vector representations; field selection during indexing directly impacts which content the retriever can find.
- Keyword Search Index
- Exact lexical matching that finds results based on specific words and terms; handles product codes and domain-specific terms well but does not understand semantic meaning.
- Vector Search Index
- Semantic similarity search using vector embeddings; understands contextual meaning and finds conceptually related content, but struggles with exact numbers and specialized terms.
- Hybrid Search Index
- Combines keyword and vector search for balanced results; the best choice when content includes both exact terms (model numbers) and conceptual explanations.
- Enriched Search Index
- Enhances vector or hybrid indexes by automatically extracting metadata, entities, and question chunks from documents to improve RAG retrieval accuracy.
Retrievers
- Retriever
- The component that searches a Data Library index and returns the most relevant data chunks to ground an agent's response or prompt template.
- Individual Retriever
- Queries a single search index; simpler to configure, lower latency — use when all grounding data lives in one source.
- Ensemble Retriever
- Combines and ranks results from multiple search indexes; use when relevant grounding data is spread across multiple Data Library sources.
- Individual vs Ensemble
- Individual retrievers are simpler and faster; ensemble retrievers are more comprehensive but add latency and configuration complexity — use ensemble only when data spans multiple sources.
- Data Library vs Data Cloud Data Streams
- Data Library is specifically for agent grounding knowledge sources; Data Cloud data streams are for general data ingestion and unification — they serve different purposes.
Development Lifecycle and Testing
- Agent Lifecycle Stages
- Ideation → Build → Test → Deploy → Observe; the cycle is iterative — Utterance Analysis observations should feed back into ideation and building improvements.
- Agentforce Testing Center
- Batch testing tool that validates agent configurations at scale; runs automated test cases to measure topic selection accuracy and action execution correctness before production.
- Testing Center vs Utterance Analysis
- Testing Center is PRE-DEPLOYMENT validation of configured test cases; Utterance Analysis is POST-DEPLOYMENT analysis of real production conversations — they are not interchangeable.
- Sandbox Support for Agentforce
- Data Cloud and Agentforce support sandbox environments for safe development and UAT; Einstein Trust Layer audit trail is active in sandboxes; configurations do not sync automatically to production.
- Deployment Artifacts
- Agent configurations must be explicitly migrated from sandbox to production; they do not automatically sync — treat agent configs as deployable metadata.
- Agentforce Analytics
- Post-deployment monitoring built natively on Data Cloud; provides adoption metrics, accuracy insights, usage patterns, and performance dashboards.
- Utterance Analysis
- Post-deployment tool that analyzes real production conversations to identify where the agent failed to understand intent, revealing new topic opportunities and improvement areas.
- Digital Wallet
- Salesforce's consumption metering system for Data Cloud and Agentforce usage; tracks AI credit consumption across development and production — measures costs, not performance.
Multi-Agent Interoperability
- Model Context Protocol (MCP)
- An open standard originally from Anthropic that enables AI agents to securely connect with external tools, systems, and data through a standardized client-server architecture.
- MCP is Open, Not Proprietary
- MCP was developed by Anthropic and adopted by Salesforce — it is an open standard, not a Salesforce-proprietary protocol; this is a common exam trap.
- MCP in Agentforce
- Agentforce acts as an MCP client to consume external tool APIs; also exposes Salesforce data through MCP so external agents can access it with enterprise-grade security enforcement.
- AgentExchange
- Curated marketplace of vetted MCP servers and pre-built agent components deployable into Agentforce through Agent Builder with no code required; evolved from AppExchange for the agent era.
- AgentExchange — Vetted Servers
- AgentExchange lists curated, security-reviewed MCP servers — it is NOT an open marketplace where anyone can publish; entries are reviewed for reliability and security.
- Agent API
- Programmatic REST interface for interacting with Agentforce agents from external applications and systems; used for external integrations, not internal agent-to-agent communication.
- MCP Trust Layer Boundary
- All MCP actions execute within the Einstein Trust Layer; PII masking and toxicity detection apply to MCP-sourced data before it reaches any LLM.