Salesforce Agentforce Specialist flashcards
Salesforce Agentforce Specialist · Build recall with a focused review session.
32 cards. Read the term and the definition here. The deck at the top of this page is optional.
- Atlas Reasoning Engine
- Salesforce reasoning engine inside Agentforce that interprets a request, grounds it in business data, plans which topic and actions apply, executes them, and composes the reply.
- Topic
- Configured area of an agent that handles one user intent; it carries a classification description, scoped instructions, and the set of actions Atlas may call for that intent.
- Action
- Discrete capability an agent can invoke, built from a Flow, an Apex method, a Prompt Template, or a Data Cloud Retriever, with declared inputs and outputs.
- Employee Agent
- Agentforce agent that serves internal Salesforce users inside the app, handling CRM tasks such as record lookups, summaries, and updates rather than external customer conversations.
- Service Agent
- Customer-facing Agentforce agent that handles conversations on messaging and voice channels, answering questions and performing service actions such as case updates or order lookups.
- Flex Credits
- Consumption currency for Agentforce; each conversation and each action execution draws credits, so agent design decisions directly affect operating cost.
- Channel
- Surface where an agent conversation occurs, such as enhanced web messaging, in-app messaging, voice, Slack, or a custom application connected through the Agent API.
- Grounding
- Injecting retrieved, authoritative business data into the model prompt so responses reflect your records and current values instead of the model's general training data.
- Retrieval-Augmented Generation
- Pattern where relevant content chunks are retrieved from a search index at runtime and passed to the language model along with the question to produce a grounded answer.
- Retriever
- Agent action type that runs semantic search over a Data Cloud search index and returns the best-matching chunks for the agent to ground its answer on.
- Data Lake Object
- Raw ingested data in Data Cloud, stored close to how it arrived from the source before harmonization into the customer data model.
- Data Model Object
- Harmonized Data Cloud object mapped to a standard schema, used for identity resolution, calculated insights, segmentation, and agent grounding.
- Chunking
- Splitting unstructured content into passages before embedding; chunk size and overlap determine how precisely retrieval can target the relevant passage.
- Search Index
- Data Cloud index storing embeddings of chunked content, queried at runtime to return passages that are semantically similar to the user's question.
- Classification Description
- Plain-language statement of what a topic covers; Atlas matches user utterances against it, so overlapping descriptions between topics cause misrouting.
- Topic Instructions
- Rules scoped to a single topic that tell the agent how to behave within that intent, including required steps, constraints, and what to do when data is missing.
- Agent-Level Instructions
- Global guidance applied to every conversation, such as persona, tone, and safety rules, which all topics inherit and must not contradict.
- Invocable Apex Action
- Apex method annotated as invocable, with request and response wrapper classes, that Agentforce exposes to an agent as a callable action.
- Autolaunched Flow Action
- Screenless Flow exposed as an agent action, where variables must be marked available for input and output so the agent can pass and receive data.
- Prompt Template Action
- Reusable generative template with merge fields and grounding resources, exposed to an agent as an action that returns generated text.
- Conversation Preview
- Builder panel for chatting with the draft agent to verify topic routing and action execution before the agent is activated on any channel.
- Testing Center
- Agent Builder area where you assemble test suites of utterances with expected topics and actions, then run them to check routing and behavior.
- Test Suite
- Named collection of agent test cases run together, with pass or fail results per utterance that expose regressions after topics, actions, or instructions change.
- Containment Rate
- Share of conversations the agent resolves without escalating to a human; a primary measure of service agent effectiveness and self-service value.
- Escalation Rate
- Percentage of agent conversations routed to a human or queue; elevated values usually indicate missing topics, weak grounding, or unclear instructions.
- Session Tracing
- Per-turn diagnostic record showing the topic Atlas selected, the actions invoked, and the grounding data used, enabling a wrong answer to be traced to its cause.
- Einstein Trust Layer
- Security layer between Agentforce and the language model that handles dynamic grounding, data masking, zero retention, toxicity scoring, prompt injection defense, and audit logging.
- Zero Data Retention
- Arrangement where the model provider stores neither prompts nor completions; Salesforce additionally does not train foundation models on customer data.
- Data Masking
- Replacing sensitive field values with tokens before a prompt leaves Salesforce, then restoring the real values in the response shown to the user.
- Prompt Injection Defense
- Trust Layer check that detects and blocks instructions hidden in user input or retrieved content that attempt to override the agent's configured rules.
- Toxic Content Detection
- Trust Layer scoring of model output for harmful or abusive content, blocking or replacing responses that exceed the configured threshold.
- Agent User Permissions
- Agents execute actions as a dedicated user, so object and field access of that user caps what the agent can retrieve or expose in a reply.