Skip to main content
Handle complex questions by routing to the right specialist (billing, support, research) and returning a single, well-structured answer.

​
What you’ll build

  • Multiple CrewAI agents (specialists) plus a coordinator agent.
  • A workflow that sequences or hierarchically calls those agents.
  • NDJSON streaming so CometChat shows progress as the coordinator works.

​
Prerequisites

  • CrewAI project (crew-ai.mdx)
  • Agent configs for each specialist (billing/support/research/etc.)

​
Steps

1

Define specialists

Add agents like billing, support, and research with narrow goals and tools.
2

Create a coordinator task

The coordinator inspects the user question, decides which specialists to call, and merges their outputs.
3

Choose a process

Use Process.sequential for simple fan-out or Process.hierarchical if you want dynamic branching.
4

Stream results

Keep the existing /kickoff NDJSON stream; tool calls from sub-agents will appear in CometChat.

​
Example configuration

src/crew_demo/config/agents.yaml 
billing:
  role: Billing Specialist
  goal: Handle invoices and refunds
  backstory: >
    Use billing tools only. Escalate if payment methods need human approval.

support:
  role: Support Specialist
  goal: Troubleshoot product issues
  backstory: >
    Ask clarifying questions and suggest next steps using available tools.

coordinator:
  role: Relay Coordinator
  goal: Route the request to the right specialist and return a concise summary
  backstory: >
    Decide which specialist to invoke. Combine their findings into one clear response.
src/crew_demo/crew.py (excerpt) 
from crew_demo.tools.get_deals import get_recent_deals

@agent
def billing(self) -> Agent:
    return Agent(config=self.agents_config["billing"], tools=[get_recent_deals], verbose=False)

@agent
def support(self) -> Agent:
    return Agent(config=self.agents_config["support"], tools=[], verbose=False)

@agent
def coordinator(self) -> Agent:
    return Agent(config=self.agents_config["coordinator"], tools=[], verbose=False)

@crew
def crew(self) -> Crew:
    return Crew(
        agents=[self.coordinator(), self.billing(), self.support()],
        tasks=self.tasks,
        process=Process.sequential,  # swap to Process.hierarchical for dynamic routing
        verbose=False,
        stream=True,
    )
src/crew_demo/config/tasks.yaml 
coordination_task:
  description: >
    Decide which specialist should handle: {user_message}
    Share a merged answer back to the user.
  expected_output: >
    A concise response summarizing each specialist's findings.
  agent: coordinator

​
CometChat setup

  • Provider: CrewAI
  • Agent ID: coordinator
  • Deployment URL: /kickoff
Use suggested prompts like “Ask billing if invoice #123 is paid and summarize it” to exercise multi-agent flows.