LLMClassifier

LLMClassifier is the default classifier. It builds a system prompt from all registered agents’ id, name, and description, then calls your LLM client with a select_agent tool. The model calls that tool with the chosen agent’s id and a confidence value; an absent or hallucinated id resolves to nil, triggering the Orchestrator’s default-agent fallback.

See the Classifiers overview for how ClassifierResult and the routing dispatch work.

Initializer

public struct LLMClassifier: Classifier {
    public init(model: any LLMClient, instructions: String = LLMClassifier.defaultInstructions)
}

Parameter	Type	Default	Notes
model	any LLMClient	—	The LLM backend used for routing calls.
instructions	String	LLMClassifier.defaultInstructions	System prompt governing routing behaviour. See Custom instructions below.

Basic usage

Pass any LLMClient and you’re done:

let classifier = LLMClassifier(
    model: BedrockLLMClient(modelId: "us.anthropic.claude-3-5-haiku-20241022-v1:0")
)

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

A lightweight model (Haiku, Flash, etc.) is usually sufficient — the task is structured tool-call routing, not open-ended generation.

Conversation continuity

The history passed to classify is the merged, [agentId]-prefixed conversation fetched from chat storage. The default prompt instructs the model to use it for follow-up detection: short affirmations like “yes” or “ok” stay with the previously selected agent rather than triggering a fresh match.

Custom instructions

LLMClassifier.defaultInstructions contains an {{AGENT_DESCRIPTIONS}} placeholder that is replaced at classify time with a roster built from all registered agents. When you supply your own instructions string:

• With {{AGENT_DESCRIPTIONS}} — the placeholder is replaced as usual.
• Without {{AGENT_DESCRIPTIONS}} — the agent roster is appended to the end of your instructions.

let classifier = LLMClassifier(
    model: myClient,
    instructions: """
    You are a strict router. Only route to agents explicitly listed below.
    If no agent matches, do not call select_agent.
    {{AGENT_DESCRIPTIONS}}
    """
)

Note

Custom instructions must still guide the model to call the select_agent tool — the tool definition is injected automatically, but the model needs to be told to use it. The default prompt already covers this.

Default instructions reference

LLMClassifier.defaultInstructions is public so you can inspect or extend it:

public static let defaultInstructions: String

It encodes the AgentMatcher prompt ported from the Python agent-squad, including:

• Categorisation against the {{AGENT_DESCRIPTIONS}} roster.
• Confidence levels: High ≈ 0.9+, Medium ≈ 0.6–0.8, Low < 0.6.
• Follow-up continuation — short replies like “yes”, “ok”, or numeric answers route to the previously selected agent.
• Four worked examples covering initial queries, context switches, follow-ups, and multi-turn switches.

Related pages

• Classifiers overview — the Classifier protocol, ClassifierResult, and the nil-classifier path.
• Custom classifiers — implement your own routing logic.
• LLM clients — the LLMClient protocol and built-in implementations.