Section 7 · Multi-Agent Orchestration · 5% of exam
7.1 Choosing a Multi-Agent Architecture
Exam objective
Determine when a multi-agent architecture is appropriate for a business use case that needs greater specialization, scalability, or control.
Begin with the work, not the agent count
A multi-agent architecture connects a coordinating agent with one or more specialized agents so that work can be divided by capability. Start by mapping the user’s request, the decisions required, the systems involved, and the expected result. A single agent is often the simpler choice when tasks share the same instructions, data access, actions, and user journey. It avoids extra routing, handoffs, and places where context can be lost. Multiple agents become useful when a broad agent would otherwise carry unrelated scopes, when specialist behavior changes independently, or when distinct business domains need clear ownership.
Look for boundaries that justify specialization
Good boundaries are based on meaningful differences, such as separate domains, audiences, processes, permissions, or knowledge sources. For example, a service assistant may need billing expertise for invoices and field-service expertise for repair scheduling. A coordinator can identify the request and connect it to the right specialist while keeping a consistent front door. Separate teams may also own different specialists and release them on different schedules. The architecture should make responsibilities clearer. Creating several agents that all answer the same requests with overlapping instructions can make routing less predictable and increase maintenance work.
Balance control with coordination
More agents do not automatically make a system more scalable. Each delegation can add latency, consume usage, introduce a failure point, or require the user to repeat context. Consider how work is routed, what information is passed, whether results must be combined, and what happens if a specialist is unavailable or cannot complete its task. Define a fallback or human escalation path. Keep each specialist’s scope narrow enough to be understood, but broad enough to complete a coherent task. Decide which agent owns the user-facing response and which agent has authority to take an action.
Apply the scenario test
For an exam scenario, identify whether the need is internal specialization or external interoperability. If one assistant needs to coordinate distinct agents, compare single-agent and multi-agent designs against control, scope, security, and user experience. If a business process is deterministic and must follow exact steps, a flow or other explicit automation may be more appropriate than adding another reasoning agent. Then select the smallest design that meets the requirement. A strong answer explains the benefit of delegation and the tradeoff it accepts, rather than assuming every complex request needs multiple agents.
List the specialist responsibilities, shared user experience, data boundaries, and failure path before deciding to split one agent into several.
An extra agent adds routing and operational overhead. Use multiple agents only when specialization, ownership, or control provides a clear benefit.
Extra concept: Scalability can mean independent ownership and focused scopes; it does not mean adding agents without limit.
Official references:
- Salesforce Certified Agentforce Specialist Exam Guide https://help.salesforce.com/s/articleView?id=005298924&language=en_US&type=1
- Multi-Agent Orchestration https://help.salesforce.com/s/articleView?id=ai.agent_multi_orch.htm&language=en_US&type=5
- Considerations and Limitations in Multi-Agent Solutions https://help.salesforce.com/s/articleView?id=ai.agent_multi_orch_consider.htm&language=en_US&type=5
Key takeaways
- Use multi-agent architecture when specialist responsibilities or control boundaries justify coordination.
- Keep a single agent when the same scope, instructions, actions, and access can handle the work clearly.
- Include routing, context transfer, latency, failures, security, and maintenance in the design decision.
