Agentforce Specialist multi-agent orchestration
Multi-Agent Orchestration is worth 5% of the Salesforce Agentforce Specialist exam — around three of the 60 scored questions and the smallest section in the outline.
Three questions does not justify a week of study. It does justify an hour, because the concepts are self-contained and the marks are straightforward.
Salesforce names two objectives.
When a multi-agent architecture is appropriate
Given a scenario, determine whether a Multi Agent architecture is appropriate for scalability and control.
The case for splitting work across several agents rather than building one large one:
- Scope clarity. An agent with a focused purpose behaves more predictably than one asked to do everything. Fewer topics means less chance of selecting the wrong one.
- Control and permissions. Different agents can run in different security contexts, so an agent handling sensitive operations can be governed separately from a general enquiry agent.
- Ownership. Different teams can own different agents without editing each other’s configuration.
- Scalability. Adding a capability means adding an agent rather than enlarging and destabilising an existing one.
The case against is equally examinable: multi-agent adds coordination complexity. A single well-scoped agent is the right answer for a simple, contained use case, and a scenario describing one modest job does not need an architecture.
The judgement: does this genuinely comprise distinct responsibilities, ideally with different data, permissions or owners? If yes, multi-agent. If it is one job described at length, one agent.
MCP and A2A
Explain the purpose of existing open standard multi-agent protocols such as MCP and A2A.
Note the wording — explain the purpose. You are not being asked to implement them.
MCP — Model Context Protocol. An open standard for exposing tools, data and context to models in a consistent way. Rather than wiring each integration bespoke per agent, capabilities are exposed once through a common protocol and consumed by any model or agent that speaks it.
A2A — Agent-to-Agent. An open standard for agents communicating with other agents, including agents built by other teams on other platforms. It is what lets an agent delegate to a specialist agent it does not own.
The distinction worth holding:
| Purpose | |
|---|---|
| MCP | An agent reaching tools and context |
| A2A | An agent reaching another agent |
Why these appear at all
Worth a moment’s reflection, because it explains the section’s existence.
Both MCP and A2A are open standards, not Salesforce technology. Microsoft’s AB-620 exam covers both as well. Two competing vendors building the same protocols into their certification outlines is a strong signal that agent interoperability is becoming standard infrastructure rather than a proprietary advantage.
For the exam, that means: when a scenario involves reaching something outside Salesforce — another vendor’s agent, a shared tool layer — the answer is an open protocol rather than a Salesforce-specific feature.
Sample questions
Question 1. An organisation wants its Salesforce agent to hand a specialised underwriting question to an agent maintained by a partner on a different platform. Which applies?
- A. MCP
- B. A2A
- C. The Trust Layer
- D. The Agent API
Show answer
Answer: B
A2A is the open standard for agent-to-agent communication across implementations, which is what delegating to a partner’s agent on another platform requires. MCP exposes tools and context to models, the Trust Layer governs security and privacy, and the Agent API exposes your own agent programmatically.
Want more questions like this? Full Agentforce Specialist practice tests →
Question 2. A company needs one agent to answer general product questions for customers. The use case is well defined and involves a single set of data and permissions. Is a multi-agent architecture appropriate?
- A. Yes — multi-agent is always more scalable
- B. Yes — it allows a different model per agent
- C. No — a single well-scoped agent fits; multi-agent adds complexity without benefit here
- D. No — multi-agent is unsupported on Salesforce
Show answer
Answer: C
A single well-scoped responsibility with one data set and one permission context does not benefit from multi-agent coordination, which adds complexity without addressing a real constraint. Multi-agent is justified by distinct responsibilities, differing security contexts or separate ownership, none of which is present.
Want more questions like this? Full Agentforce Specialist practice tests →
Question 3. What is the purpose of the Model Context Protocol?
- A. An open standard for exposing tools and context to models in a consistent way
- B. An open standard for one agent to delegate work to another agent
- C. A Salesforce feature for masking sensitive data before it reaches a model
- D. A mechanism for deploying agents from sandbox to production
Show answer
Answer: A
MCP is an open standard for exposing tools, data and context to models consistently, so capabilities are defined once rather than wired bespoke per integration. It does not govern agent-to-agent delegation, which is A2A, nor does it handle data masking or deployment between orgs.
Want more questions like this? Full Agentforce Specialist practice tests →
What to practise
Write two sentences: one saying what MCP is for, one saying what A2A is for. Then write one sentence on when multi-agent beats a single agent.
That is genuinely proportionate for three questions. The time you save belongs in AI Agents, which is worth seven times as much.