jeevesagent.architecture.router

Router: classify input → dispatch to one specialist Agent.

OpenAI Agents SDK March 2026 “Handoff” pattern, plus the classify-and-route shape every framework reinvents (CrewAI sequential, LangGraph conditional edges, …). The simplest multi-agent pattern.

Pattern

  1. Classify. A small fast LLM call decides which route handles the input best.

  2. Dispatch. The chosen specialist Agent runs to completion with its own model / memory / tools / architecture.

  3. Return. The specialist’s output becomes the Router’s output. No cross-specialist synthesis.

Compared to Supervisor:

  • Cheaper (1 classification + 1 specialist; no synthesis pass).

  • Deterministic (single specialist owns the task).

  • Less flexible (no multi-domain tasks; routing errors cascade).

When to use

  • Customer support (route to billing / tech / sales / general).

  • Helpdesks where each query has one right specialist.

  • API-gateway-style intent routing.

Replay correctness

The specialist’s Agent.run() is invoked with a deterministic session_id derived from the parent session and the route name: {parent_session_id}__route_{route_name}. On replay, the same specialist session_id reproduces, and the specialist’s own journal (under its own session) takes over from there. The parent’s journal caches the classification step — replay flows cleanly through both layers.

Specialists are full Agents

Each route is a fully-constructed Agent instance. They are NOT shared dependencies of the parent. If you want shared budget / memory / telemetry, pass the same instances when building the specialists.

Attributes

DEFAULT_CLASSIFIER_PROMPT

Classes

Router

Classify input → dispatch to ONE specialist Agent.

RouterRoute

One specialist + classification metadata.

Module Contents

class jeevesagent.architecture.router.Router(*, routes: list[RouterRoute], fallback_route: str | None = None, require_confidence_above: float = 0.0, classifier_prompt: str | None = None)[source]

Classify input → dispatch to ONE specialist Agent.

declared_workers() → dict[str, jeevesagent.agent.api.Agent][source]
async run(session: jeevesagent.architecture.base.AgentSession, deps: jeevesagent.architecture.base.Dependencies, prompt: str) → collections.abc.AsyncIterator[jeevesagent.core.types.Event][source]
name = 'router'
class jeevesagent.architecture.router.RouterRoute[source]

One specialist + classification metadata.

name is what the classifier emits in its route: line and must be unique within a Router. description is shown to the classifier alongside the name — keep it specific and distinguishing so the classifier picks reliably.

agent: jeevesagent.agent.api.Agent
description: str = ''
name: str
jeevesagent.architecture.router.DEFAULT_CLASSIFIER_PROMPT = Multiline-String
Show Value
"""You are a routing classifier. Given the user's request, decide which
specialist handles it best.

Available routes:
{route_descriptions}

Output exactly two lines, in this order:
route: <one of the route names above>
confidence: <number between 0 and 1>

Then optionally one line of brief reasoning. The first two lines
must match the format exactly so they can be parsed.
"""