Multi-Agent Orchestration (Beta)

Multi-Agent Solutions

Multi-Agent Orchestration for Agentforce is an interoperability framework that you can use to extend your agent’s capabilities. In this framework, the orchestrator agent communicates with the user. It uses subagent classification to route the request (not the conversation) to a connected subagent, based on the connected subagent's description and any instructions in the Agent Router. It retains context throughout the handoff and synthesizes the connected subagent's response before it reaches the user. With multi-agent orchestration, businesses can deploy a single agent across their channels with shared context and without customer friction.

Maybe you've built a great customer service agent and now your team wants to add a field service use case. Or you're already running three agents in different business units and the business requires that they talk to each other. Or you're staring at an Instructions block that's grown so long that it covers six different domains and you're not sure why the routing keeps breaking. Whatever the trigger, you've arrived at the same problem. At some point, one agent is not enough.

Orchestrator Agents and Connected Subagents

If your agent manages and delegates tasks to other connected subagents, it’s known as an orchestrator agent. An orchestrator agent can route a conversation to any combination of subagents and connected subagents to get the job done. You can think of an orchestrator agent as the single “front door” that interacts with the user. It also coordinates efforts behind the scenes, delegating work to the various “back door” connected subagents, and communicating the results back to the user.

A connected subagent is an entirely separate agent in your Salesforce org that your orchestrator agent can use like a subagent. It’s important to understand how connected subagents differ from standard subagents added from the Asset Library:

• Subagents are localized components that simply contain the instructions (which control logic and reasoning) and actions (the tools) available for the primary agent to use.
• Connected subagents represent a complete, independent agent with its own distinct expertise and identity. A connected subagent contains one or more subagents of its own. Connected subagents are modular, represent their own domain, and can be well suited for reuse across many multi-agent solutions.

Collectively, the mix of internal subagents and external connected subagents assigned to your orchestrator defines the full scale of capabilities your agent can handle.

In the context of an orchestrator agent, a reference agent is the agent upon which your connected subagent is based. When you add a connected subagent, it contains settings from the reference agent (such as the description) that you can override to customize its behavior.

Plan Your Agent-to-Agent Solution

Here are a few tips to get you started designing an agent for use in Multi-Agent Orchestration for Agentforce.

• For determinism and consistent behavior, base connected subagents on agents that are unique and distinct. For example, authorizing a return, sending a shipping label, and refunding a purchase represent three discrete actions that different connected subagents can perform.
• Think modularity and reuse. Specialized agents can represent different business domains that must interact across organizational boundaries (for example, Sales, Human Resources, and Finance) to achieve a common goal. When designed for reuse, specialized agents can do the work across many multi-agent solutions.
• Craft the descriptions (instructions) used for connected subagents. The orchestrator agent and Agent Router rely on the description to determine when to route to a connected subagent. Descriptions that are vague, or that overlap with other subagents or connected subagents, can result in inconsistent behavior.
• Test your agent individually before connecting to it from an orchestrator agent. Verify its instructions and reasoning actions, and any agent actions included with the connected subagent, including the underlying flows, prompt templates, or other reference actions. Unit testing helps you find and resolve issues in isolation so that you’re more confident when testing it in combination with other solution components.

Considerations and Limitations (Beta)

Review these considerations and limitations for Multi-Agent Orchestration for Agentforce (beta).

• You can connect agents as subagents within a single Salesforce org. Multi-agent orchestration isn't currently supported for Agentforce agents across multiple Salesforce orgs or external, non-Agentforce agents.
• You can use the following agent type combinations for orchestrator agents and connected subagents:
  • ASA orchestrator agent with ASA connected subagents
  • AEA orchestrator agent with AEA connected subagents
  • AEA orchestrator agent with ASA connected subagents
• These agent types and combinations aren't supported in the beta release:
  • ASA orchestrator agent with AEA connected subagents
  • File-based agents (for example, SDR or Analytics)
• Connect only active agents to your orchestrator agent. Don’t select an inactive agent.
• Only an orchestrator agent can escalate to a human. Connected subagents can’t escalate to a human.
• An orchestrator agent supports one level of delegation. A connected subagent can’t delegate to other agents.
• An orchestrator agent can support multiple connected agents. We recommend no more than 7 connected subagents per orchestrator agent. Connected agents can have up to 5 subagents.
• Latency in multi-agent solutions is generally higher than in single-agent solutions. Intrinsic factors include coordination overhead and a higher frequency of LLM calls (to synthesize responses, for example).
• The timeout limit for orchestrator agents is 120 seconds. The timeout includes the elapsed time spent by the orchestrator agent doing specialized work, such as delegating to connected subagents and synthesizing the results. The connected subagent timeout is 30 seconds. The orchestrator agent synthesizes timeout errors.
• User-level authentication is supported. Orchestration among AEA agents is handled on behalf of a logged-in user, so the user must be logged in beforehand. For ASA agents, unauthenticated guest users can access them as well.
• For the beta release, we support only context variables (not custom variables), which must be mapped. Mapping is supported in one direction only: from an orchestrator agent to connected subagents.
• You can configure variable mappings only in the Script view, not in the Canvas view.
• The Transition option isn’t available for adding determinism to your Agent Router. You can use other options, such as If/Else (Conditional). In this case, specify invoke under the If condition, expand Actions, then scroll to select the connected subagent you want to invoke.
• For customers who didn't participate in the Multi-Agent Orchestration for Agentforce (Pilot) program, before using an agent with Multi-Agent Orchestration for Agentforce (Beta):
  • If the agent was built using the new Agent Builder, create and activate a new version of it.
  • If the agent was built using the legacy Agentforce Builder, upgrade it to the new Agent Builder. See Upgrade an Agent from the Legacy Builder to the New Builder.