Section 7 · Multi-Agent Orchestration · 5% of exam
7.4 Model Context Protocol (MCP)
Exam objective
Explain how MCP standardizes an agent’s access to tools and data in external systems and how it differs from agent orchestration.
MCP connects agents with external capabilities
Model Context Protocol is an open standard for connecting AI applications with external systems and their capabilities. In a typical MCP arrangement, a server makes capabilities available, a client connects to the server, and a host coordinates the client experience. MCP servers can define tools, prompts, and resources at the protocol level. In Salesforce’s current Agentforce MCP integration, agents use MCP tools: executable functions exposed by a server and added to an agent as actions. These actions let an agent access an external capability through a standardized connection rather than requiring a separate custom integration for every tool.
Map the pattern to a business request
Suppose an agent needs a risk score from an external service. The service exposes an MCP tool, Agentforce connects to the MCP server, and the agent can invoke the tool when the request and instructions call for it. The tool performs the defined operation and returns structured data for the agent to use. Salesforce’s setup flow depends on the server source: third-party servers use the Agentforce Registry, while Salesforce-hosted and MuleSoft servers use API Catalog workflows. Administrators allowlist the individual tools that agents may use, then add the desired tool actions to the agent.
Distinguish MCP from multi-agent design
MCP is primarily a standardized way for an agent to use tools or data provided by a server. It does not, by itself, decide which specialist agent should own a conversation or how an orchestrator should combine delegated results. That is an architecture and routing decision. An agent can use MCP tools while remaining a single agent; a multi-agent design can delegate among specialists without MCP. Some solutions may use both: the coordinator chooses a specialist, and that specialist calls an external tool through MCP. On an exam, identify whether the requirement is tool interoperability or agent-to-agent collaboration.
Include governance and deployment checks
A protocol does not remove the need for security review. Confirm the server identity, authentication, tool allowlist, data returned, and action permissions. Test the tool in Agentforce Builder and Testing Center, and inspect traces when an action is missing or out of sync. Salesforce notes that MCP server registration metadata is not packaged with deployed agents, so target environments need their server registrations and allowlisted tools configured before the agent can use them. Validate the connection and tool behavior in each environment, and review changes to server definitions before activation.
If the requirement is “call an external tool or service,” think MCP; if it is “delegate work to another agent,” think A2A or an orchestration pattern.
MCP can standardize tool access, but the connected server, authentication, available tools, and data access still require governance.
Extra concept: MCP tools can be used as agent actions. Their presence does not turn the agent architecture into a multi-agent system.
Official references:
- MCP for Agentforce https://help.salesforce.com/s/articleView?id=ai.agent_mcp.htm&language=en_US&type=5
- Connect Agentforce Solutions to MCP Servers https://help.salesforce.com/s/articleView?id=ai.agent_mcp_connect.htm&language=en_US&type=5
- Considerations for MCP for Agentforce https://help.salesforce.com/s/articleView?id=ai.agent_mcp_considerations.htm&language=en_US&type=5
Key takeaways
- MCP is an open protocol for connecting agents to server-provided capabilities; Agentforce uses MCP tools as actions.
- Allowlist only the required tools and verify server access, authentication, data, and action behavior.
- MCP tool interoperability and multi-agent orchestration solve different design problems.
