Why Multi-Agent Workflows?
A customer call can require several specialized skills: verifying identity, checking an order, making a payment, and escalating to a manager. Instead of one agent trying to do everything, a multi-agent system assigns each subtask to a specialized agent, orchestrated by a supervisor.
What Is LangGraph?
LangGraph is a framework for building stateful, multi-actor applications with LLMs. It lets you define nodes (agents or tools) and edges (transitions), creating cyclic or branching workflows. It is especially useful for long-running, interactive processes like voice calls.
What Is MCP?
MCP (Model Context Protocol) is the bridge between agents and external tools. An agent that needs to check Salesforce, send a Slack message, or create a Zendesk ticket uses an MCP server to do so securely.
How They Work Together
- LangGraph orchestrates: Routes the conversation between agents based on state and intent.
- MCP executes: Each agent calls the right tools through standardized MCP servers.
- State persists: The workflow remembers context across tools and turns.
- Human-in-the-loop: Critical steps can pause for human approval before continuing.
Example: Refund Workflow
A customer calls for a refund. The voice agent verifies the order, checks the return policy, initiates the refund via Stripe, sends a confirmation email, and updates the CRM. If the refund amount is above a threshold, it asks a manager for approval before processing.
Why Crawlii Uses This Stack
Crawlii’s platform combines LangGraph-style state machines with a managed MCP registry. This lets businesses build complex, multi-step voice agents that span multiple apps and departments — without writing custom orchestration code.
Frequently Asked Questions
Is LangGraph required for multi-agent systems?
No, but it makes building stateful, cyclical workflows much easier than raw prompt chaining.
Can non-technical users build these workflows?
With a platform like Crawlii, yes. The orchestration and MCP connections are managed behind a visual or prompt-based interface.
How many agents can work in one call?
As many as needed. In practice, 3–7 specialized agents plus a supervisor cover most enterprise use cases.