Skip to main content

​
Multi-Agent Workflows

Multi-agent workflows allow a single user-facing agent (the orchestrator) to delegate tasks to multiple specialized subagents. This is useful when your organization has distinct knowledge domains or toolsets that benefit from dedicated agents.

​
How it works

  1. The user sends a message to the orchestrator
  2. The orchestrator analyzes the request and determines which subagent can best handle it
  3. The request is delegated to the appropriate subagent
  4. The subagent processes the request using its own tools and knowledge bases
  5. The orchestrator receives the subagent’s response and returns it to the user
The user only interacts with the orchestrator — they don’t need to know about or select individual subagents. Delegations run as part of the orchestrator’s own agent task, so the whole tree counts as one task against a user’s parallel-task limit no matter how many subagents it calls. Each delegation is checkpointed individually: if a worker restarts mid-task, subagent work that already finished is not repeated.

​
Setting up a multi-agent workflow

1

Create subagents

First, create the individual specialized agents that will serve as subagents. Each should have:
  • A focused knowledge base for its domain
  • Tools relevant to its area
  • A clear system prompt defining its expertise
2

Create the orchestrator

Create a new agent and configure it as an orchestrator. In the agent configuration:
  • Select Multi-Agent as the agent type
  • Add the subagents you created in the previous step
  • Write a system prompt that describes when to delegate to each subagent
3

Configure delegation rules

The orchestrator’s system prompt should clearly define routing logic:
4

Test the workflow

Start a chat with the orchestrator and test questions that should route to different subagents. Verify that:
  • Questions are routed to the correct subagent
  • Responses include citations from the right knowledge bases
  • Cross-domain questions are handled gracefully

​
Best practices

  • Keep subagents focused — Each subagent should be an expert in one area
  • Write clear routing instructions — The orchestrator’s system prompt should leave no ambiguity about which subagent handles what
  • Limit the number of subagents — 3–5 subagents is typically optimal; too many can confuse the routing. The platform also caps how many subagents run concurrently within a single task, so extra breadth queues rather than speeding things up
  • Watch the time budget — The whole delegation tree shares the parent task’s limit of roughly an hour of active work. Orchestrators are the likeliest agents to reach it; prefer several focused requests over one sweeping one
  • Test edge cases — Try ambiguous questions that could go to multiple subagents