What Is Agentforce? 30 Key Terms Every Salesforce Buyer Should Know 

1. Agentforce

Agentforce is Salesforce’s platform layer for building and running autonomous or semi-autonomous AI agents inside the Salesforce ecosystem. In buyer terms, Agentforce meaning is not simply “chatbot.” It is a way to combine Salesforce AI agents, data access, business rules, actions, channels, and governance so work can move from request to outcome.

2. Agentforce 360

Agentforce 360 is Salesforce’s broader platform direction that connects humans, applications, agents, and data into one operating model. Salesforce’s Agentforce 360 announcements describe features such as Agentforce Builder, Agent Script, Agentforce Voice, and Intelligent Context. Buyers should read Agentforce 360 as the umbrella story, not as one single feature.

3. Salesforce AI agents

Salesforce AI agents are software workers configured to handle defined jobs inside sales, service, marketing, commerce, IT, and operations. They can answer questions, summarize records, draft content, retrieve context, and trigger actions. The important point is scope. Salesforce AI agents should have clear jobs, data boundaries, and escalation rules before they meet real users.

4. Agent

An agent is the user-facing or system-facing AI worker. It receives a request, interprets intent, retrieves context, and either responds or takes action. A buyer should ask what the agent is allowed to do without human review. Reading Agentforce terminology this way keeps the agent’s job separate from the product label.

5. Subagent

A subagent is the work-specific unit that handles a defined type of task. In older language, many buyers still hear this described as a topic. Subagents matter because routing quality depends on boundaries. A refund subagent, a warranty subagent, and a pricing subagent may need different sources, instructions, and actions.

6. Agentforce topics and actions

Agentforce topics and actions is a useful phrase because it captures the practical operating model. A topic or subagent defines the kind of work. An action defines what the agent can do. Together, they decide whether the agent only talks or can actually execute. This is one of the most important Agentforce glossary ideas for buyers.

7. Action

An action is a callable task the agent can use to get something done, such as creating a case, updating a field, drafting an email, launching a Flow, or calling an API. Actions deserve scrutiny because they create operational consequences. The more actions an agent has, the more the team needs permissions, testing, rollback logic, and monitoring.

8. Instructions

Instructions tell the agent how to behave inside a subagent or task. They can define tone, sequence, limitations, data use, and decision rules. Weak instructions create vague behavior. Overloaded instructions create brittle behavior. The buying question is whether instructions are precise enough to govern common cases without hiding exceptions.

9. Reasoning instructions

Reasoning instructions guide how the agent decides between possible paths. They matter when a request could fit more than one subagent or action. In a service workflow, for example, the agent may need to classify whether a customer wants order status, refund help, technical troubleshooting, or escalation. Poor reasoning instructions cause misrouting.

10. Atlas Reasoning Engine

The Atlas Reasoning Engine is the reasoning system behind Agentforce. Salesforce describes it as the brain of Agentforce, responsible for interpreting requests, planning steps, and orchestrating action. For buyers, Atlas Reasoning Engine is the term to inspect when a vendor explains how the agent chooses what to do next.

11. Agentic loop

An agentic loop is the recurring pattern where the agent interprets the request, retrieves context, selects a path, takes an action, evaluates the result, and continues if the job is not finished. This term matters because multi-step work creates more failure points than one answer. Buyer governance should follow the loop, not only the final response.

12. Agentforce Builder

Agentforce Builder is the workspace used to define, test, and manage agents. It helps teams assemble subagents, instructions, actions, and testing flows. A focused Agentforce Quickstart usually starts here with a narrow, measurable use case.

13. Agent Script

Agent Script gives builders explicit control over how an agent behaves, step by step, rather than relying on the reasoning engine to choose a path. It is useful when a team needs determinism, not only probabilistic reasoning. Buyers should ask where free-form reasoning is acceptable and where Agent Script or structured logic is needed. That distinction matters in regulated, financial, service, and revenue workflows.

14. Hybrid reasoning

Hybrid reasoning combines flexible AI reasoning with more deterministic control. It matters because enterprise teams rarely want pure autonomy everywhere. A policy answer may need a grounded response. A refund approval may need a hard workflow. Hybrid reasoning is the language for deciding which parts can flex and which parts must be scripted.

15. Data 360 for Agentforce

Data 360 for Agentforce is the data foundation that gives agents trusted business context. It can support unified profiles, metadata context, unstructured information, and governed retrieval. It has to be enabled for Agentforce to work at all, though a full implementation is only needed for specific capabilities. Our Data 360 services exist because the data layer, more than the model, determines whether the agent answers with company-specific truth or with generic fluency.

16. Data Cloud

Data Cloud is the former name for Data 360. Salesforce renamed it on October 14, 2025, and the functionality did not change. In practice, buyers will hear both names for some time yet, because documentation, the product interface and internal runbooks are all still catching up. The safer phrase in a mixed room is “Data Cloud, now Data 360”.

17. Grounding

Grounding means supplying the agent with trusted context before it generates or acts. That context can come from CRM records, Knowledge, files, Data 360, a retriever, or an approved external source. Grounding is central to Salesforce Agentforce meaning because it separates an enterprise agent from a general-purpose language model answering from memory.

18. RAG Salesforce

RAG Salesforce refers to retrieval augmented generation patterns inside the Salesforce ecosystem. IBM defines retrieval augmented generation as connecting an AI model to external knowledge bases so responses can be based on relevant outside information. In Salesforce, RAG includes data access, permissions, retrievers, prompt context, and response generation.

19. Retriever

A retriever decides which pieces of content or data should be returned for a prompt. Good retrievers improve relevance. Poor retrievers feed the agent the wrong evidence, and they also cost more, because every retrieval consumes credits. In an Agentforce glossary, retriever belongs next to grounding and RAG Salesforce because it is one of the mechanisms that controls what the model actually sees.

20. Vector index

A vector index stores content in a form that supports semantic retrieval. Instead of matching only exact keywords, it helps find passages that are similar in meaning. Buyers do not need to become vector database specialists, but they should understand that document quality, chunking, metadata, and source control all affect vector search quality.

21. Agentforce Data Library

An Agentforce Data Library is a defined source layer for grounding agents in approved business content. It can help a narrow agent answer from a controlled knowledge base or file set. It should not be mistaken for a full enterprise data strategy. A library can prove usefulness, while Data 360 for Agentforce can support a broader production context.

22. Intelligent Context

Intelligent Context refers to extracting usable meaning from complex business content, including unstructured material. It matters when policies, PDFs, product sheets, diagrams, and support documents contain the knowledge an agent needs. Buyers should ask which content types are truly supported, how they are indexed, and how retired documents are excluded.

23. Salesforce Trust Layer

The Salesforce Trust Layer is the security and trust architecture that helps protect data and users when generative AI and Agentforce features operate. It covers secure data retrieval, grounding, masking, policies, audit, and feedback. The Salesforce Trust Layer does not remove the need for business governance, but it gives the architecture a safer foundation. It also requires Data 360 to be enabled, which is one reason the data requirement is not optional.

24. Guardrails

Guardrails are the controls that keep the agent inside approved boundaries. They can include instructions, permission rules, human review, content restrictions, topic boundaries, policy checks, and escalation paths. The Agentforce governance framework is where guardrails become operating design rather than a paragraph in a policy document.

25. Human in the loop

Human in the loop means a person reviews, approves, or takes over at specific points. This is not a sign that Agentforce failed. It is often the correct design for sensitive, high-value, or ambiguous work. The glossary term helps buyers separate automation from autonomy. Some decisions should remain supervised.

26. Prompt template

A prompt template is a reusable instruction pattern that can include placeholders filled with business data. It helps standardize generated responses, summaries, or recommendations. Prompt templates matter because they connect language generation to approved context. Buyers should ask who owns templates, how they are tested, and how changes are approved.

27. Digital labor

Digital labor is the business phrase for AI agents performing work that previously required human effort. The term can be useful, but it can also become too broad. A buyer should translate digital labor into specific tasks, such as qualify a lead, summarize a case, draft a reply, update a record, or escalate a workflow.

28. Flex Credits

Flex Credits are Salesforce’s consumption-based pricing unit for Agentforce. They are sold at $500 per 100,000 credits, which works out to about $0.005 per credit. A standard Agentforce action uses 20 credits, roughly $0.10, and an Agentforce Voice action uses 30 credits, roughly $0.15. Buyers should treat Flex Credits as both a finance term and a design term, because action volume is what drives spend. Our Agentforce pricing guide works through the full cost math.

29. Conversation pricing

Conversation pricing charges around an interaction unit rather than every individual seat or action, at $2 per conversation. It can be simpler to model for some customer-facing agents, but it still requires volume assumptions. A buyer should model normal, peak, and failure scenarios. A third option now exists: pay-per-resolution, where the Help Agent is billed $2 only when it actually resolves the inquiry.

30. Agentic Enterprise

Agentic Enterprise is Salesforce’s strategic phrase for a company where humans and AI agents work together across data, apps, and workflows. It is useful as a vision term, but buyers should connect it back to implementation. Agentic Enterprise only becomes real when data, governance, users, processes, and measurable outcomes are ready.