The short version
Use this sheet to review, not to learn from scratch. It lists the facts, the term pairs people mix up, and rules of thumb for scenario questions. It does not cover everything on the exam, and it contains no real exam questions. The Salesforce exam guide (aligned to Spring ’26) is the authority.
Exam facts at a glance
| Item | Detail |
|---|---|
| Scored questions | 60 multiple-choice, plus up to 5 unscored |
| Time | 105 minutes |
| Passing score | 72% |
| Languages | English, French, Japanese |
| Delivery | Proctored, testing center or online |
| References | None allowed |
| Maintenance | Trailhead module once a year |
Weightings: AI Agents 35%, Prompt Engineering 20%, Data 360 Fundamentals 20%, Testing/Deployment/Maintenance 10%, Governance and Observability 10%, Multi-Agent Orchestration 5%.
AI Agents (35%)
Key concepts
- Agent Script mixes fixed rules (conditions and transitions) with language-model reasoning. This is called hybrid reasoning.
- Filters, variables, and template expressions let you force behavior where variation is not acceptable.
- Topics group what an agent can handle; actions are what it can do.
- Channels include digital experience, email, voice, and Slack.
Easily confused
- Employee agent is for signed-in internal users and runs in their context. Service agent is for customers and does not depend on a signed-in Salesforce user.
- Topic is the subject area. Action is the task performed inside it.
- Canvas view is the visual editor. Script view shows the Agent Script itself.
Decision rules
- If the scenario says “always” or “never”, look for a deterministic option, not more prompt text.
- If the users are not signed in, think Service agent.
- If the conversation happens in website or mobile chat, think messaging chat. If it is email, think the email channel. If it is voice, think telephony.
- If an outside system needs to call the agent, think the Agent API.
Trap: choosing the most advanced feature. The exam rewards the simplest one that meets the constraint.
Prompt Engineering (20%)
Key concepts
- Prompt Builder creates templates. Types include field generation and flex.
- Access to a template is controlled with access settings, not by editing every profile.
- The Trust Layer is a set of protections. Learn each protection separately.
Easily confused
- Record merge field delivers a stored value as is. Retriever returns passages judged relevant to the question.
- Template type decides where the template can be used. Grounding decides what data it sees.
Decision rules
- Exact number or field value that must not be paraphrased: merge field.
- Answer lives in long, unstructured documents: retriever.
- Grounding methods can be combined in one template.
Trap: assuming a protection from one area (for example masking in templates) works the same way everywhere. Read what the scenario says is in use.
Data 360 Fundamentals (20%)
Key concepts
- Content is split into chunks, indexed so it can be searched, and returned by a retriever.
- The Agentforce Data Library builds that chain for you from files.
Easily confused
- Search index stores searchable data. Retriever fetches from it.
- Library (generated for you) versus a custom retriever (you configure it, useful when the content already lives in Data 360 objects).
Decision rules
- Library first, custom retriever when the data is already in Data 360 or needs manual control.
- A library needs Data 360 set up and a supported edition with a qualifying add-on. Check what exists before choosing a feature.
Trap: indexing everything. Only index what the agent will be asked about.
Testing, Deployment, and Maintenance (10%)
- Testing Center runs evaluations against test cases. Know what each evaluation checks and how to find which step failed.
- Moving from sandbox to production: ask what moves with the agent, such as a prompt template it depends on, and what must already exist in the target org.
- Channel and routing setup tends to be environment specific, so a deployed agent can be invisible to customers until it is configured.
Trap: assuming “deployed” means “live”.
Governance and Observability (10%)
| Need | Think |
|---|---|
| Alert within minutes when errors start | Real-time health monitoring |
| Performance across many interactions over time | Agent analytics |
| Why one conversation went the way it did | Tracing and optimization tools |
Observability can depend on other platform pieces, such as Data 360, being provisioned first.
Multi-Agent Orchestration (5%)
- Use several agents when scale or control makes one agent hard to keep reliable. One well-organized agent is often enough.
- MCP and A2A are open protocols. Know each one’s purpose in one sentence.
Last-day checklist
- Reread the six section names and weightings.
- Say aloud one decision rule per section.
- Review your misses from practice sets, not new material.
- Plan about 1 minute 45 seconds per question; flag and move on.
- Read the constraint in each stem first.
- Sleep. Check the exam day guide for logistics.
More guides for AGENTFORCE-SPECIALIST
- Agentforce Specialist Study Plan
- Agentforce Specialist Sample Questions
- Is the Agentforce Specialist Exam Hard?
- Agentforce Specialist Prerequisites
- Agentforce Specialist Renewal and Retakes
- Is Agentforce Specialist Worth It?
- Agentforce Specialist vs AB-620: Which First?
Next step
Go deeper with domains explained and the exam format page. To test your recall under time pressure, try our Agentforce Specialist practice tests.