LearnedAgentRouter<T>

Models & Types · AiDotNet.Agentic.SelfImproving

A router that learns, from graded AgentTrajectory history, which candidate agent performs best and routes new requests accordingly — a contextual reward-weighted bandit. As more rewarded runs accumulate, the routing improves, so the orchestration gets better at its task over time without a hand-tuned router.

Implements: IAgent<T>

For Beginners

Imagine a dispatcher who remembers which worker did best on each kind of job. Given a new job it sends it to the worker with the best track record (occasionally trying others to keep learning). Feed it the graded history and it gets smarter about who to pick.

How It Works

Because it is an IAgent, the router drops in anywhere an agent is expected and composes with tracing (so its own routed runs feed back into the trajectory store, closing the self-improvement loop). It exploits the highest mean-reward agent for the request’s context, explores unseen agents first (optimistic), and takes a random agent with probability explorationRate to keep discovering. An optional context key lets it learn different best-agents for different kinds of task.

Constructors

Constructor Summary
LearnedAgentRouter(IReadOnlyList<IAgent<T>> candidates, double explorationRate, Func<IReadOnlyList<ChatMessage>, string> contextKey, int? seed) Initializes a new learned router.

Properties

Property Summary
string Description { get; }
string Name { get; }

Methods

Method Summary
Task<AgentRunResult> RunAsync(IReadOnlyList<ChatMessage> messages, CancellationToken cancellationToken)
string SelectAgentName(IReadOnlyList<ChatMessage> messages) Returns the name of the agent the router would choose for the given request, without running it (deterministic when explorationRate is 0).
void LearnFrom(IEnumerable<AgentTrajectory> trajectories) Updates the router’s policy from graded trajectories.