AI Agent Router

The AI Agent Router node uses an LLM to intelligently decide which workflow path to follow. It presents each sub-node as a possible tool call to the LLM, and the LLM selects the most appropriate one based on the input and instructions.

• Node kind: extensions.com.cloud-apim.llm-extension.router

How it works

1. The router receives an input and a set of sub-nodes (paths)
2. Each sub-node is presented to the LLM as a tool with its id and description
3. The LLM chooses which tool (path) to call based on the instructions and input
4. The selected sub-node is executed and its result is returned

This is similar to a switch node, but the routing decision is made by an LLM rather than by deterministic predicates.

Configuration

Parameter	Type	Required	Description
provider	string	yes	Id of the LLM provider to use for routing
input	string	no	The input to route (defaults to the input variable in workflow memory)
instructions	array	yes	System instructions that guide the routing decision
model	string	no	Override the model used by the provider
model_options	object	no	Override model options
paths	array	yes	The paths to choose from. Each one is a node with an id and a description, or { id, description, node }, the way the workflow designer writes it

Example

{
  "kind": "extensions.com.cloud-apim.llm-extension.router",
  "provider": "provider_xxxxx",
  "input": "${input.question}",
  "instructions": [
    "You determine which agent to use based on the user's homework question"
  ],
  "paths": [
    {
      "id": "math_tutor",
      "description": "Specialist agent for math questions",
      "kind": "call",
      "function": "extensions.com.cloud-apim.llm-extension.llm_call",
      "args": {
        "provider": "provider_xxxxx",
        "payload": {
          "messages": [
            {
              "role": "system",
              "content": "You provide help with math problems. Explain your reasoning at each step."
            },
            {
              "role": "user",
              "content": "${input.question}"
            }
          ]
        }
      },
      "result": "call_res"
    },
    {
      "id": "history_tutor",
      "description": "Specialist agent for historical questions",
      "kind": "call",
      "function": "extensions.com.cloud-apim.llm-extension.llm_call",
      "args": {
        "provider": "provider_xxxxx",
        "payload": {
          "messages": [
            {
              "role": "system",
              "content": "You provide assistance with historical queries. Explain important events and context clearly."
            },
            {
              "role": "user",
              "content": "${input.question}"
            }
          ]
        }
      },
      "result": "call_res"
    }
  ],
  "result": "router_result"
}

Paths in the workflow designer

In the designer, the router shows one handle per path: plug the node, or the chain of nodes, the path leads to, and give the path its name and description in the form of the router. The designer saves each path as { id, description, node }. A path left without a name takes the id and the description of its node.

Example with AI Agent nodes as paths

The router paths can be any workflow node, including AI Agent nodes:

{
  "kind": "extensions.com.cloud-apim.llm-extension.router",
  "provider": "provider_xxxxx",
  "input": "${input.question}",
  "instructions": [
    "Route the question to the most appropriate specialist."
  ],
  "paths": [
    {
      "id": "math_tutor",
      "description": "Specialist agent for math questions",
      "kind": "extensions.com.cloud-apim.llm-extension.ai_agent",
      "name": "math_tutor",
      "provider": "provider_xxxxx",
      "instructions": [
        "You provide help with math problems. Explain your reasoning at each step and include examples."
      ],
      "input": "${input.question}"
    },
    {
      "id": "history_tutor",
      "description": "Specialist agent for historical questions",
      "kind": "extensions.com.cloud-apim.llm-extension.ai_agent",
      "name": "history_tutor",
      "provider": "provider_xxxxx",
      "instructions": [
        "You provide assistance with historical queries. Explain important events and context clearly."
      ],
      "input": "${input.question}"
    }
  ],
  "result": "router_result"
}

Router vs Handoffs

Both the router and handoffs allow dynamic routing, but they serve different purposes:

Feature	Router	Handoffs
Decision point	Before any agent runs	During agent execution
Use case	Route to different workflow paths	Transfer between agents mid-conversation
Context	Only the input and instructions	Full conversation context
Flexibility	Can route to any node type	Routes to other agents only

Agent workflow with routing