Salesforce Agentforce Specialist study plan

Distinguish Agentforce agents from copilots and classic chatbots by listing what makes a session autonomous, and name the business problems each pattern fits

Decompose an agent into its building blocks - topics, instructions, actions, and the reasoning engine - and explain how each one shapes a single session

Trace the end-to-end agent lifecycle from design through build, test, deploy, and optimize, and identify the artifact produced at each stage

Assemble a disposable sandbox agent with one topic and one Flow action, then run a session against it to observe the metadata in action

Author a topic with tightly scoped instructions, a clear description, and explicit out-of-scope language

Attach at least three action types (Flow, prompt template, and Apex or API) to topics and define their inputs, outputs, and confirmation requirements

Configure agent-level settings - company and brand context, tone, escalation path, and target channel

Explain how the reasoning engine chooses among topics and actions, and predict selection behavior for five sample utterances before testing

Compare grounding options - Data Cloud retrieval, Salesforce object and Flow lookups, document libraries, and external retrieval services - and match each to a question type

Design a retrieval setup for unstructured knowledge, specifying chunking, filters, and the search index or vector search configuration

Map the exact source fields and objects each topic needs, and define refresh cadence and data-quality validation rules

Write prompt templates that constrain responses to retrieved content and define a fallback when retrieval returns nothing relevant

Wire the grounding design into the agent and validate retrieval accuracy against ten real questions drawn from the source data

Tune retrieval parameters, filters, and topic instructions when answers are wrong, stale, or fabricated

Add a guardrail that declines out-of-scope questions and redirects to an escalation path

Instrument the agent so each answer records which source grounded it

Describe the trust layer capabilities - zero data retention, dynamic masking, toxicity detection, and audit trail - and state which risks each one mitigates

Configure user access so the agent operates with least privilege, using permission sets and a dedicated agent user rather than an admin context

Apply privacy controls for PII, including field-level restriction, masking rules, consent handling, and data residency considerations

Define a human-in-the-loop and escalation policy for sensitive or irreversible actions the agent can take

Build a test plan with utterance sets that cover happy path, ambiguous phrasing, adversarial input, and out-of-scope requests

Run batch tests with the platform testing tools and read session traces to pinpoint where routing or grounding failed

Deploy the agent to a messaging or chat channel and verify behavior end-to-end outside the builder preview

Set up monitoring on session analytics - containment, escalation rate, fallback frequency, and error signals

Red-team the agent with prompt-injection, data-exfiltration, and toxic-input attempts, and confirm the guardrails hold

Optimize the weakest topics using trace analysis and the platform optimization tooling, then re-run the regression suite

Document a release and rollback procedure with a change-review checklist covering configuration, data, and permissions

Re-audit access, masking, and audit logging after deployment and record any drift

Rebuild a minimal agent from a blank sandbox covering topic design, one grounded action, and one guardrail, without consulting notes

Complete two timed mixed practice sets spanning all five domains and log every miss by domain and root cause

Re-teach the three weakest concepts to yourself in writing and verify each with a fresh practice item

Review the topic, action, and grounding decision points most likely to be tested as scenario trade-offs