Orchestrator Agent Pattern

What Makes an Agent an Orchestrator?

Any Octopus agent can act as an orchestrator. What distinguishes it is:

• Its system prompt focuses on intent classification, not domain answering
• Its tool registry includes the route_to_agent and handoff_to_agent tools (registered by the multi-agent plugin)
• It has a TeamConfig that lists the specialist agents it can route to
• Its LLM may be a smaller/cheaper model since it only routes — not a heavy reasoning model

AgentRouter

The AgentRouter is the component that makes the routing decision. It is configured with a routing strategy:

public class AgentRouter
{
    public async Task<AgentRoutingResult> RouteAsync(
        string userMessage,
        AgentTeam team,
        CancellationToken ct = default)
    {
        return _config.Strategy switch
        {
            RoutingStrategy.Keyword   => await KeywordRouteAsync(userMessage, team),
            RoutingStrategy.Embedding => await EmbeddingRouteAsync(userMessage, team),
            RoutingStrategy.LLM       => await LLMRouteAsync(userMessage, team),
            _ => AgentRoutingResult.Default(team.DefaultAgentId)
        };
    }

    private async Task<AgentRoutingResult> EmbeddingRouteAsync(
        string message, AgentTeam team)
    {
        var queryEmbedding = await _embeddingProvider.EmbedAsync(message);

        var scores = team.Members.Select(member => new
        {
            AgentId = member.AgentId,
            Score = CosineSimilarity(queryEmbedding, member.IntentEmbedding)
        });

        var best = scores.MaxBy(s => s.Score);
        return new AgentRoutingResult(best!.AgentId, best.Score, RoutingStrategy.Embedding);
    }
}

Orchestrator System Prompt Template

The orchestrator's system prompt is focused entirely on routing:

You are a routing assistant for the BizFirstGO enterprise support team.
Your job is to determine which specialist agent should handle the user's request.

Available specialists:
- HR Agent: leave requests, payroll questions, benefits, onboarding
- Finance Agent: expense reports, budget queries, invoice approval
- IT Support Agent: software issues, access requests, hardware problems
- Legal Agent: contract review, compliance questions, policy interpretation

Rules:
1. Route to exactly one agent per user message.
2. If the request spans multiple domains, route to the primary one.
3. If you cannot determine the domain, route to the default agent.
4. Do NOT answer the user's question yourself — only route.

Use the route_to_agent tool to route.

Routing Decision with LLM Strategy

// The orchestrator LLM is given the route_to_agent tool:
{
  "name": "route_to_agent",
  "description": "Route the user's request to the appropriate specialist agent.",
  "inputSchema": {
    "type": "object",
    "properties": {
      "agentId": { "type": "string", "description": "The ID of the target specialist agent" },
      "reason": { "type": "string", "description": "Why this agent was selected" },
      "confidence": { "type": "number", "description": "0.0 to 1.0 routing confidence" }
    },
    "required": ["agentId", "reason"]
  }
}

Routing Configuration in agents-app

In the agents-app, the team's routing policy is configured in the Team Config panel: