Salesforce Agentforce Specialist cheat sheet

Atlas Reasoning Engine runs the agent loop: interpret the request, retrieve and ground data, select a topic and plan, execute actions, then compose the reply.

An agent is three declarative parts: topics (what it can handle), actions inside those topics (what it can do), and instructions (how it should behave).

Atlas routes each utterance to one best-matching topic using the topic's classification description, so descriptions must be distinct from one another.

Action types available to agents are Flow, Apex, Prompt Template, and Retriever; each declares inputs and outputs that Atlas uses when planning.

Prebuilt agents include Agentforce SDR, Sales Coach, Service Agent, and the Employee Agent; all can be inspected, extended, or used as design references.

Employee Agent serves internal users inside Salesforce and can surface CRM records; Service Agent handles external conversations on messaging and voice channels.

Delivery channels include enhanced web and in-app messaging, voice via Service Cloud Voice, Slack, and the Agent API for custom applications.

Consumption is metered in Flex Credits: each conversation and each action execution draws credits, so agent design is also a cost decision.

Agents are metadata configured in Agent Builder, not hand-written code, and only take effect on a channel after the agent is activated.

Grounding injects retrieved, authoritative business data into the prompt so the model answers from your records instead of its training data.

Structured grounding uses Prompt Template merge fields or Flow-retrieved record data; unstructured grounding uses Retrievers over Data Cloud search indexes.

RAG flow: ingest content, chunk it, embed chunks into a vector index, retrieve top-matching chunks at runtime, and pass them to the model with the question.

Data Cloud layers data as DLOs (raw ingested), DMOs (harmonized model objects), and unified profiles produced by identity resolution rules.

A Retriever is an action type, so a topic can only ground on a search index if a Retriever action referencing that index is attached to the topic.

Knowledge articles ground answers through Knowledge or Retriever actions; stale or thin articles directly bound answer accuracy.

Calculated Insights precompute aggregate metrics into a DMO so agents answer quantitative questions consistently instead of estimating.

Retrieval quality depends on chunk size, embedding choice, and re-indexing after source content changes; unreindexed content is invisible to the agent.

Missing or unmapped fields in a prompt template create hallucination risk, so instructions must define behavior when grounding data is absent.

Build order: create the agent, define topics with distinct classification descriptions, attach actions, write agent-level and topic-level instructions, preview, activate, then attach a channel.

Use one topic per distinct user intent; overlapping classification descriptions cause misrouting and are the most common routing defect.

Action descriptions tell Atlas when to select an action, so each input and output description should state exactly what data is passed and returned.

Autolaunched Flow actions require flow variables marked as available for input and output; Apex actions require an @InvocableMethod with request and response wrappers.

Prompt Template actions return generative text and can be grounded with record fields or retrievers, and the same template can be reused across agents.

Agent-level instructions set persona, tone, and global guardrails; topic-level instructions narrow behavior to that intent and must not contradict global rules.

Conversation preview lets you chat with the draft agent to confirm topic routing and action execution before anything reaches a customer.

Expose only the actions a topic actually needs; broad action sets increase misrouting and consume more Flex Credits per conversation.

Define an explicit escalation or fallback topic so the agent has a deterministic path when it cannot resolve the request.

Testing Center holds test suites built from utterances paired with an expected topic and expected actions; each run reports pass or fail per case.

Rerun the regression suite after every change to topics, classification descriptions, instructions, actions, or the underlying model.

Pre-activation checks: routing accuracy on the suite, action success rates, working escalation path, Trust Layer settings, and permissions on the agent user.

Deployment steps: activate the agent, attach it to the channel and messaging deployment, and confirm the agent user's object and field access.

Core metrics are containment (resolution without handoff), escalation rate, average handling time, CSAT, and per-topic misroute rate.

Session tracing records each turn's selected topic, invoked actions, and grounding data, which is how a wrong answer is traced back to its cause.

Optimization loop: identify the highest-volume failing utterances in traces, tighten classification descriptions or grounding, then retest the suite.

Credit optimization: replace multi-step chatty plans with one action that returns consolidated data in a single call.

Roll out to a pilot channel or segment first, compare containment and escalation against baseline, then expand.

Einstein Trust Layer sits between the agent and the LLM and provides dynamic grounding, data masking, zero retention, toxicity detection, prompt injection defense, and audit logging.

Zero retention means the model provider does not store prompts or completions, and Salesforce does not train foundation models on customer data.

Data masking replaces sensitive field values with tokens before the prompt leaves Salesforce, then restores real values in the response returned to the user.

Prompt injection defense blocks instructions hidden in user input or retrieved content that attempt to override agent rules; action scope is the second line of defense.

The audit trail stores each generation's prompt, response, masked fields, and model used, enabling review of what the agent said and why.

Agents act as a dedicated user, so every action runs inside that user's object and field permissions; least privilege on that user caps what the agent can expose.

For a Service Agent on a public channel, verify guest user access and record sharing so a grounded answer cannot surface another customer's data.

Toxic content detection scores model output and blocks or replaces responses that exceed the configured threshold.

Governance: document grounding sources and action permissions, review flagged conversations, and re-review when topics or data sources are added.