Skip to main content

Agent Router

Routers: CRUD, member management, stats, decisions and feedback.

Access it as client.agent_router on a Strongly client, or the same path on AsyncStrongly with await. All methods exist on both with identical signatures.

Quick start​

from strongly import Strongly

client = Strongly()

# List (auto-paginates as you iterate)
for agent in client.agent_router.list():
    print(agent.id)

Methods​

Core​

list​

list(*, limit: int | None = None) -> SyncPaginator[AgentRouter]

List the routers you can use, a page at a time as you iterate.

create​

create(*, name: str, strategy: str, description: str | None = None, members: Sequence[Mapping[str, Any]] | None = None, fallback_behavior: str | None = None, sticky_routing: bool | None = None, health_check: Mapping[str, Any] | None = None, bandit_config: Mapping[str, Any] | None = None, canary_config: Mapping[str, Any] | None = None, semantic_config: Mapping[str, Any] | None = None, complexity_config: Mapping[str, Any] | None = None, content_type_config: Mapping[str, Any] | None = None, token_budget_config: Mapping[str, Any] | None = None, routellm_config: Mapping[str, Any] | None = None, custom_strategy: Mapping[str, Any] | None = None) -> AgentRouter

Create a router and return it.

Parameters

  • name (str): Router name (required).
  • strategy (str): Routing strategy (required): priority, weighted, round-robin, latency-optimized, thompson_sampling, ucb1, epsilon_greedy, linucb, complexity, content-type, token-budget, semantic, routellm, canary or custom.
  • description (str | None, optional): Router description.
  • members (Sequence[Mapping[str, Any]] | None, optional): {"model_id", "weight", "priority" (lower = preferred), "enabled", "capabilities"} dicts.
  • fallback_behavior (str | None, optional): next (fail over by priority) or error.
  • sticky_routing (bool | None, optional): weighted: the same caller always gets the same member.
  • health_check (Mapping[str, Any] | None, optional): {"enabled", "interval_seconds", "failure_threshold", "recovery_threshold"}.
  • bandit_config (Mapping[str, Any] | None, optional): learning strategies: {"reward_metric": success|latency|feedback, "epsilon" (epsilon_greedy), "exploration_bonus" (ucb1 c, linucb alpha), "max_latency_ms", "linucb_features" (linucb)}.
  • canary_config (Mapping[str, Any] | None, optional): {"control_model_id", "canary_model_id", "canary_percent"}.
  • semantic_config (Mapping[str, Any] | None, optional): {"embedding_model_id", "confidence_threshold", "default_model_id", "routes": [{"name", "model_id", "examples"}]}.
  • complexity_config (Mapping[str, Any] | None, optional): {"simple_model_id", "complex_model_id"?}.
  • content_type_config (Mapping[str, Any] | None, optional): {"type_model_map", "default_model_id"}.
  • token_budget_config (Mapping[str, Any] | None, optional): {"long_response_tokens", "short_model_id"}.
  • routellm_config (Mapping[str, Any] | None, optional): {"embedding_model_id" (text-embedding-3-small), "threshold"}.
  • custom_strategy (Mapping[str, Any] | None, optional): {"code"}: Python that sets result to a member id.

retrieve​

retrieve(router_id: str) -> AgentRouter

Retrieve a single router by id.

update​

update(router_id: str, *, name: str | None = None, strategy: str | None = None, description: str | None = None, members: Sequence[Mapping[str, Any]] | None = None, fallback_behavior: str | None = None, sticky_routing: bool | None = None, health_check: Mapping[str, Any] | None = None, bandit_config: Mapping[str, Any] | None = None, canary_config: Mapping[str, Any] | None = None, semantic_config: Mapping[str, Any] | None = None, complexity_config: Mapping[str, Any] | None = None, content_type_config: Mapping[str, Any] | None = None, token_budget_config: Mapping[str, Any] | None = None, routellm_config: Mapping[str, Any] | None = None, custom_strategy: Mapping[str, Any] | None = None) -> AgentRouter

Update a router and return it (arguments as in create).

What a learning router has learned is kept.

delete​

delete(router_id: str) -> None

Delete a router. Raises on failure; returns nothing.

Other​

add_member​

add_member(router_id: str, *, model_id: str, weight: float | None = None, priority: int | None = None, enabled: bool | None = None, capabilities: Sequence[str] | None = None) -> AgentRouter

Add a member model to a router and return the updated router.

Parameters

  • router_id (str): Router id.
  • model_id (str): Model id to add (required).
  • weight (float | None, optional): Relative share of traffic (for the weighted strategy).
  • priority (int | None, optional): Priority, lower = preferred.
  • enabled (bool | None, optional): Whether the member is enabled (default true server-side).
  • capabilities (Sequence[str] | None, optional): Content types it handles (content-type strategy).

decisions​

decisions(router_id: str, *, limit: int | None = None) -> SyncPaginator[RouterDecision]

List the router's decisions, newest first, a page at a time as you iterate.

Each has the strategy's reason, the outcome and the reward.

feedback​

feedback(router_id: str, *, decision_id: str, reward: float, label: str | None = None) -> RouterFeedbackResult

Score a routed response (routers whose reward is feedback).

Parameters

  • router_id (str): Router id.
  • decision_id (str): response.router_decision.id of the routed response.
  • reward (float): From 0 (bad) to 1 (good). One reward per decision.
  • label (str | None, optional): Optional note kept with the feedback.

remove_member​

remove_member(router_id: str, member_id: str) -> None

Remove a member model from a router. Raises on failure; returns nothing.

stats​

stats(router_id: str, *, period_hours: int = 24) -> RouterStats

Return what the router did in the last period_hours and what it has learned.

Per member: decisions, share, answered / failed, latency, cost, rewards; pulls and average reward over its whole life.