Classifiers Overview

A classifier runs before each turn and decides which agent should handle it. The Orchestrator calls classify(_:history:agents:), gets back a ClassifierResult, and dispatches to selectedAgent — or falls back to the default agent when selectedAgent is nil.

Core types

public struct ClassifierResult: Sendable {
    public let selectedAgent: (any AgentProtocol)?
    public let confidence: Double

    public init(selectedAgent: (any AgentProtocol)?, confidence: Double)
}

public protocol Classifier: Sendable {
    func classify(
        _ input: String,
        history: [ConversationMessage],
        agents: [any AgentProtocol]
    ) async throws -> ClassifierResult
}

ClassifierResult fields

Field	Type	Notes
selectedAgent	(any AgentProtocol)?	The chosen agent, or nil when nothing matched. Always one of the agents passed to classify.
confidence	Double	Nominally 0...1. Captured for tracing only.

Note

confidence is observability data, not a routing threshold. The Orchestrator only inspects selectedAgent; it falls back to the default agent when that is nil. If you want confidence-gated routing, implement the threshold inside your Classifier.

How routing works

1. The Orchestrator fetches merged conversation history from storage.
2. It calls classifier.classify(_:history:agents:) with the user input, that history, and all registered agents.
3. selectedAgent from the returned ClassifierResult is dispatched directly. No second lookup is performed — the agent instance must come from the agents parameter.
4. When selectedAgent is nil, the Orchestrator routes to the first registered agent (the default).

Skipping the classifier (nil path)

Pass nil as the classifier when constructing the Orchestrator. Every turn goes straight to the first registered agent with no LLM call for routing.

let orchestrator = Orchestrator(classifier: nil, storage: storage)

Note

nil classifier is the right default for single-agent setups. Adding LLMClassifier to a single-agent orchestrator costs a full LLM round-trip per turn with no benefit.

Built-in classifiers

Classifier	Description
LLMClassifier	Uses an LLMClient and a select_agent tool call to pick the best agent. The default when a classifier is needed.

Next steps

• LLMClassifier — built-in LLM-based routing, init options, and custom instructions.
• Custom classifiers — implement the Classifier protocol to write your own routing logic.