Skip to content

strands.models.routing.classifier_strategy

Route among configured candidates by asking a model or decision model one closed question.

class ClassifierStrategy()

Defined in: src/strands/models/routing/classifier_strategy.py:106

Choose a candidate by applying a configurable routing policy with a classifier model.

The classifier is asked one closed question that carries the SDK’s fixed routing rules followed by the routing policy. A DecisionModel, such as a System One provider, can be passed in place of a model.

Each candidate is offered under a positional key (c0, c1, …); names, descriptions, and metadata travel only as evidence in the routing context, so optional or arbitrary names never become answer values. The routing context is a bounded projection of the latest request, the parent agent’s textual instructions, and candidate evidence; it crosses the classifier provider boundary and must not contain secrets.

Selection runs only before the first model call. A single candidate is returned without asking. The classifier must commit to a key. Classification failures, unknown keys, and timeouts warn and decline selection, so ModelRouter serves candidate zero. After an attempt this strategy declines, so a failure of the selected candidate surfaces without switching. Nested routers are treated as opaque candidates using only their wrapper evidence.

Example:

router = ModelRouter(
models=[
RoutingCandidate(fast, name="routine", description="Short factual requests"),
RoutingCandidate(strong, name="complex", description="Multi-step design work"),
],
strategy=ClassifierStrategy(classifier_model),
)
agent = Agent(model=router)
def __init__(
model: Model | DecisionModel,
*,
routing_policy: str | None = None,
system_prompt: str | None = None,
timeout: float = 30.0,
max_message_chars: int = _DEFAULT_MESSAGE_CHARACTER_LIMIT,
max_agent_instructions_chars:
int = _DEFAULT_AGENT_INSTRUCTIONS_CHARACTER_LIMIT,
max_candidate_chars: int = _DEFAULT_CANDIDATE_CHARACTER_LIMIT) -> None

Defined in: src/strands/models/routing/classifier_strategy.py:136

Initialize the strategy.

Arguments:

  • model - Model or DecisionModel used for classification. A Model must honor forced tool selection, since a provider that ignores it fails classification and every selection falls back to candidate zero.
  • routing_policy - Routing policy, sent verbatim and never truncated. The SDK adds mandatory isolation, candidate-key, and structured-output rules that the policy cannot override. Defaults to the SDK input-complexity policy.
  • system_prompt - Deprecated alias for routing_policy.
  • timeout - Maximum seconds to wait for the decision. On timeout the decision call is cancelled.
  • max_message_chars - Maximum characters copied from the latest request into the routing context.
  • max_agent_instructions_chars - Maximum characters copied from the parent agent’s system prompt text into the routing context.
  • max_candidate_chars - Maximum aggregate characters for the serialized evidence (names, descriptions, and metadata) of all candidates. Evidence is never truncated; selection raises ValueError when the budget is exceeded.

Raises:

  • TypeError - If model is not a Model or DecisionModel, or another argument has the wrong type.
  • ValueError - If both routing_policy and system_prompt are set, timeout is not finite and greater than zero, or a character limit is not positive.
async def select(context: RoutingContext,
**kwargs: Any) -> RoutingCandidate | None

Defined in: src/strands/models/routing/classifier_strategy.py:186

Select one opening candidate, declining on a decision failure.

Raises:

  • ValueError - If the candidates’ serialized evidence exceeds max_candidate_chars. This misconfiguration is permanent, so it propagates instead of declining.