Study guide
10.1 When a Multi-Agent Architecture Is Appropriate
Salesforce Certified Agentforce Specialist Exam
Track your progress as you read
Finished sections get a check mark.
1 Chapter 1: Exam Overview & How Agentforce Agents Work
1.1 Agentforce Specialist (AI-201) Exam Facts & Spring '26 Blueprint 1.2 How an Agent Works & the Building Blocks of Agent Script 1.3 Hybrid Reasoning: Agent Script in Canvas View and Script View
2 Chapter 2: Deterministic Control, Topics (Subagents) & Actions
2.1 Deterministic Behavior: Filters, Variables & Template Expressions 2.2 Standard and Custom Topics (Subagents) 2.3 Standard and Custom Agent Actions
3 Chapter 3: Agent Types, Security Context, Channels & Agent API
3.1 When to Use an Employee Agent or a Service Agent 3.2 Security Context: Who the Agent Runs As and How It Affects Actions 3.3 Connecting Agents to Channels: Digital Experiences, Email, Voice & Slack 3.4 When to Use the Agent API
4 Chapter 4: Prompt Builder Fundamentals
4.1 When to Use Prompt Builder 4.2 Access Controls Governing Prompt Templates 4.3 Prompt Template Types: Field Generation, Flex & More 4.4 Creating, Activating & Executing Prompt Templates
5 Chapter 5: Grounding, Prompt Best Practices, Trust Layer & Model Access
5.1 Choosing the Right Grounding Technique 5.2 Best Practices for Writing Effective Prompts 5.3 Einstein Trust Layer Security and Privacy Features 5.4 Managing and Preventing Access to Specific Models
6 Chapter 6: Data 360 Fundamentals for Agentforce
6.1 Agentforce Data Library: Types and Concepts 6.2 Chunking: Breaking Content Into Retrievable Units 6.3 Indexing and Search Types: Vector, Hybrid & Keyword 6.4 Retrievers: Individual, Ensemble & Filtered Retrieval
7 Chapter 7: Testing Agents with Testing Center
7.1 Testing an Agent with Testing Center 7.2 How Testing Center Evaluations Work
8 Chapter 8: Deploying Agents and Prompt Templates
8.1 Deploying an Agent from Sandbox to Production 8.2 Deploying Prompt Templates from Sandbox to Production
9 Chapter 9: Governance and Observability
9.1 Managing and Monitoring Agents 9.2 Agent Analytics and Agent Optimization
10 Chapter 10: Multi-Agent Orchestration, MCP & A2A
10.1 When a Multi-Agent Architecture Is Appropriate 10.2 Open Multi-Agent Protocols: MCP and A2A
10.1 When a Multi-Agent Architecture Is Appropriate
Key Takeaways Explain This Topic AI
In Multi-Agent Orchestration, an orchestrator agent handles the conversation and delegates to connected subagents, which are separate Agentforce agents in the same org.
Handoff mode (the default) lets the connected subagent respond directly with lower latency, while supervisor mode has the orchestrator synthesize responses and can call multiple agents per turn.
Salesforce describes multi-agent orchestration as appropriate when a single agent's instructions span many domains, routing keeps breaking, or separate business-unit agents must work together.
Supported combinations are Service-to-Service, Employee-to-Employee, and Employee orchestrator with Service connected subagents, with one level of delegation.
Orchestrator agents have a 120-second timeout and connected subagents a 30-second timeout, and cross-org or external non-Agentforce agents aren't supported by this feature.
Last updated: September 2026
10.1 When a Multi-Agent Architecture Is Appropriate
Quick Answer: Use a multi-agent architecture when one agent isn't enough . Signs include instructions that span many domains, routing that keeps breaking, or separately owned agents that must collaborate while customers use one front door . In Salesforce's Multi-Agent Orchestration for Agentforce , an orchestrator agent routes and delegates to connected subagents , which are complete, separate agents in the same org . Choose handoff mode when a specialist should talk to the user directly, and supervisor mode when the orchestrator must combine results from several agents.
Signs You've Outgrown a Single Agent
Salesforce describes scenarios that point to multiple agents:
A great customer service agent now needs a new domain , such as field service.
Several business units already run agents that now need to talk to each other.
An instructions block covers many domains , and routing keeps breaking .
Without orchestration, users on a single chat channel must pick among a growing list of agents and repeat themselves when switching. Engineers must untangle complex routing when handoffs fail. Orchestration gives users a single front door with shared context.
When a single agent is still the better choice
- Situation Better design
- A few related jobs in one domain, owned by one team One agent with several subagents
The agent only needs data or operations from an external system Actions (flow, Apex, API) or MCP tools (section 10.2)
- An external app needs to talk to one agent Agent API
- Latency is critical and the scope is simple Single agent. Multi-agent latency is intrinsically higher
- Orchestrator Agents and Connected Subagents
Role What it is
Orchestrator agent Handles the initial conversation, then routes and delegates tasks. It can route to any mix of its own subagents and connected subagents
Connected subagent An entirely separate agent in the org that the orchestrator uses like a subagent. It has its own expertise, identity, and subagents, and can be built from the Employee or Service agent template
Standard subagent (for contrast) A localized job inside one agent's definition, containing instructions and actions
Reference agent The agent a connected subagent is based on. Settings such as its description can be overridden in Canvas
Handoff Mode vs. Supervisor Mode
Handoff mode (default) Supervisor mode
Who responds to the user The connected subagent responds directly . The orchestrator sees the response but doesn't synthesize it The orchestrator synthesizes the connected subagent's response and replies
Routing Transition: @utils.transition to @connected_subagent.X . HyperClassifier preferred (required for handoff to work correctly) Tool call: @connected_subagent.X in reasoning actions, routed by the connected subagent's description
- Latency Lower Higher , because of an extra LLM call to synthesize
- Multiple agents per turn No , one handoff per turn Yes
Post-response chaining after_response block on the connected subagent (transition, if/then, set) Not applicable; the orchestrator stays in control
Variable mapping inputs: block, plus set in after_response inputs: block on the connected subagent
Best for Specialists that run end-to-end conversations, such as a billing or HR agent Jobs where results from multiple agents must be combined before replying
Agent Script for Connected Subagents
When you connect an agent as a subagent , the builder adds a connected_subagent block. Context moves between agents only through the inputs mapping : connected_subagent CRM_Agent:
- label: "CRM_Agent"
- description: "Use this tool for any request about CRM information"
- loading_text: "Fetching CRM information..."
inputs:
- EndUserLanguage: string = @variables.EndUserLanguage
- currentRecordId: string = @variables.currentRecordId
start_agent agent_router:
model_config:
model: "model://sfdc_ai__DefaultEinsteinHyperClassifier"
reasoning:
actions:
go_to_crm: @utils.transition to @connected_subagent.CRM_Agent
Handoff-mode chaining with after_response : after_response:
if @variables.email_needed:
transition to @connected_subagent.EmailAgent
Patterns include always chain , conditional chain , multi-branch , set the state then chain , and pipeline . Anything written after a transition to is silently skipped , so set variables before the transition. Missing inputs mapping is the usual dataflow error , because context doesn't flow automatically.
Escalation in Multi-Agent Solutions
A connected subagent's delegate_escalation switch decides whose escalation flow runs:
Scenario What happens
User asks for a human at the top level, or the orchestrator hits a global escalation rule The orchestrator's escalation flow runs
User asks for a human deep inside a specialist, or the specialist hits its own policy, and delegate_escalation: True The connected subagent's escalation flow runs
The connected subagent has no escalation routing or subagent The request is surfaced to the orchestrator's routing (cascading handoff)
- All queues are offline, or the escalation flow fails or isn't defined An AI fallback message is generated
- Authentication and Supported Combinations
Orchestrator Connected subagent Runs as
Service agent Service agent The orchestrator's session user , not the subagent's own agent user. Guest users are supported
- Employee agent Employee agent The logged-in user (who must be logged in first)
- Employee agent Service agent The Service agent's own agent user
- Service agent Employee agent Not supported in this release
- Considerations and Limits
Same org only. Orchestration isn't supported across multiple Salesforce orgs or with external non-Agentforce agents.
One level of delegation from orchestrator to connected subagents.
Timeouts: orchestrator 120 seconds (including delegation and synthesis); connected subagents 30 seconds .
No enforced limit on connected subagents. Salesforce's testing found no latency impact with one orchestrator and 8 connected subagents, each with 8 subagents.
Agentforce Voice isn't a supported configuration for this release.
Variable mapping is one-directional , from orchestrator to connected subagents.
Connect only active agents. New agent versions must be activated to be available, and a connected subagent can't be deactivated.
Legacy-builder agents must be upgraded to the new builder first.
File-based agents (for example SDR or Analytics) aren't supported in this release.
Design and Testing Tips
Pick the pattern first: handoff or supervisor.
Make connected subagents distinct. Overlapping descriptions cause inconsistent routing, and descriptions work like function signatures with explicit scope.
Design for reuse across business domains such as Sales, HR, and Finance.
Minimize latency: prefer handoff mode and streaming (both defaults), use the HyperClassifier for routing, and avoid redundant groundedness checks.
Test each agent in isolation , then run end-to-end integration tests . Testing Center's Agent Handoff Evaluation column shows whether the expected handoff occurred.