An agent is a model with tools and a loop: it can run code, read files, call APIs and keep going until a task is done. That makes it the right target for requests that need doing - fix the failing test, book the appointment, reconcile the statement - and the wrong, expensive target for requests that need a sentence. Agent routing is the decision between those, made per request.
The decision works like model routing with a wider catalogue. Agents are declared as targets with capabilities, a complexity band, a cost and a latency, the same way models are; MCP servers, A2A agent cards and SKILL.md packages are imported as targets without being rewritten. The request's signals - an action verb, a complexity above a threshold, a domain - raise an agent's score when the request needs doing and leave the plain models ahead when it does not. Policy applies to agents as to everything else, and tool-parameter provenance and resource limits bound what an agent may do with an untrusted argument.
Sometimes the right target is a sequence: a persona to set the tone, a skill to structure the work, a model to write it. A plan is a routing target like any other; the router picks it, executes its steps under the same policy and budgets, and the trace records each step. Human queues are targets too, so a request that needs judgement is routed to a person by the same decision that would have picked an agent.
