Salesforce Agentforce 360 for Public Sector AP study plan

Define the agentic enterprise in your own words and contrast it with classic chatbots and copilots, listing three criteria that make an agent 'agentic'

Diagram the Agentforce reasoning loop (intent detection, topic selection, action planning, action execution, response synthesis) and label where the Atlas reasoning engine sits

Differentiate the main agent types (service, sales, employee, and custom/agentforce-built agents) and match each to one public-sector scenario

Identify the components of an agent in Agentforce Builder (agent metadata, topics, actions, instructions, and channel settings) and sketch their hierarchy

Install or access a Developer/trial org and open Agentforce Builder to confirm the toolchain works

Write classification descriptions for four topics so each scope is mutually exclusive, and list the utterances/phrases each should match

Configure at least three actions on a topic, using at least two different action types (Flow, Apex, prompt template, or API/External Services)

Design input and output schemas for each action, including required vs optional inputs and the confirmation step before write operations

Explain in writing how the reasoning engine chooses among competing topics and actions, including what makes a topic description weak or ambiguous

Build and publish a working agent with two topics and three actions and run it in the preview/simulator

Author deterministic flows in Agent Script for a topic where free-form reasoning is risky (for example, a benefits eligibility or permit status lookup)

Compare a reasoning-driven topic with a scripted topic and document three decision rules for when to script instead of reason

Describe the grounding pipeline end to end: data ingestion, harmonization to a data model, retrieval, and injection into the prompt at runtime

Map three public-sector knowledge sources (policy manuals, statutes, case history) to the Data 360 capability that would ingest each

Draft the retrieval step for one agent action, specifying what is retrieved, how much is retrieved, and what the agent does when nothing is returned

Differentiate data lake objects, data model objects, and unstructured/search-indexed data and state which one each of your agent's actions should query

Configure a data stream and map source fields into a harmonized data model for a constituent or case data set

Explain how retrieval-augmented generation is implemented in Agentforce and where chunking, embeddings, and search indexes affect answer quality

Identify three data-quality failure modes (stale records, duplicate constituents, unmapped fields) and write the check that catches each

Trace one user question from prompt to retrieved records to final answer, documenting the exact objects and permissions involved at each hop

List the core Public Sector Solutions modules (case management, licensing and permitting, benefit management, inspections, grants) and state one agent use case for each

Compare Government Cloud and Government Cloud Plus on compliance posture, residency, and what they restrict, and record which one your target deployment implies

Identify the licensing and permission model that lets a constituent-facing agent act on behalf of a user without over-privileging it

Design an end-to-end use case for either permit status or benefit eligibility, naming the objects, topics, actions, and data sources required

Map the public-sector data handling constraints (PII, CJIS-style sensitive data, retention rules) that the use case must satisfy

Harden the week-5 use case by adding a second grounded data source and confirming the agent resolves conflicts between sources correctly

Write test utterances that should fall outside every topic and verify the agent escalates or refuses rather than guessing

Document how the agent hands off to a human case owner, including what context is passed and where it is stored

Evaluate retrieval relevance on ten representative constituent questions and record which answers were grounded vs ungrounded

Refine topic descriptions and action instructions based on the failure cases found in that evaluation

Explain the Einstein Trust Layer components (zero retention, dynamic grounding, data masking, toxicity detection, audit trail) and state the risk each one mitigates

Configure masking or field-level restrictions so the agent cannot surface sensitive constituent data it should not return

Build a test suite in the testing tool covering happy path, ambiguous intent, out-of-scope, and adversarial prompts

Define the guardrails that prevent an action from performing an unauthorized write, including confirmation and permission checks

Document the audit and monitoring signals you would review after go-live and the threshold that triggers rollback

Score 85% or higher on a full-length timed practice exam and log every missed item by domain

Re-teach each weak area by rebuilding the smallest artifact that demonstrates it (a topic, an action, a grounding step, or a guardrail)

Reconcile your week-1 architecture diagram with everything you now know and correct it

Write a two-minute verbal explanation of the agentic enterprise, grounding, and trust layer as if briefing a government stakeholder

Finalize a one-page exam-day reference sheet limited to concepts you personally got wrong