Skip to content
    Agent Arcade

    Model routing

    Beat the Router

    It's Black Friday and Penny the support bot's AI models keep crashing. Pick a working model for every customer message for 60 seconds. Can you score more than our router?

    1 min

    Penny's busiest day

    Penny

    It's Black Friday and customers are messaging me nonstop. I need an AI model to write every answer, and you get to pick which one.

    60 seconds · tap or use keys 1, 2, 3 and X · no sign-up

    How each trick worksA full explanation and a checklist for your own agent. Best read after you play.

    The router in the game follows the rules of the model router in Algen Agent Runtime, our open-source runtime for production AI agents, cut down to three models and one minute. The full rules, with code you can run, are in When the Model Provider Goes Down.

    What happened in your 60 seconds

    Everyone plays the same minute, so the scores are comparable.

    • 0–14 seconds: calm. Every model works. Big gives the best answers, so it's the right choice for every message.
    • 14–26 seconds: Big crashes. Nobody tells you. The first message you send there comes back, and the next one does too unless you change course.
    • 20–23 seconds: Medium crashes too. Short messages can still go to Small. Long and very long ones have nowhere to go.
    • 42–50 seconds: Big crashes again.
    • Three broken messages arrive along the way. No model can answer them, so the right move is to reject them.

    In the runtime's terms, Big is the primary model, Medium the fallback and Small a local model. A crash is a provider outage or a rate limit (HTTP 503 or 429), and a broken message is a request the provider rejects as invalid (HTTP 400).

    What the router did

    The router answers each message the moment it arrives, so no customer ever waits. But it can't see an outage either. It finds out the same way you do: by calling.

    It tried models in order. Big first, then Medium, then Small. That's routing_preference: ordered in the runtime: the fallbacks are there for outages, not to save money on a good day.

    It never sent a message to a model too small to read it. A very long message (24k tokens) can't fit Medium's 16k or Small's 12k window, so the router never sent one there. In the runtime, each candidate is packed separately from the original request, and one whose window can't hold the protected part of the conversation is skipped before anything is spent.

    It stopped calling Big after two crashes. At 15.4 and 17.0 seconds Big failed. Its circuit opened, and for the next few seconds the router went straight to the fallback without wasting a call. This is a circuit breaker. The runtime's default is five failures; the game uses two to fit into a minute.

    It tried again after a pause, and paid for it once. At 23.0 seconds the pause was over, Big was still down, and that single failure opened the circuit again at once. It didn't wait for another two.

    It didn't notice Big come back straight away. Big recovered at 26 seconds, but the router kept resting it until 29. At 27.4 seconds a very long message arrived that only Big could read, and the router rejected it. That's the price of a circuit breaker: a few seconds of caution after an outage, in exchange for not hammering a model that's down. If you beat the router, this is probably where you did it.

    It rejected every broken message. A request the provider rejects as invalid would fail the same way everywhere. Sending it to another model would get a plausible answer from a weaker model and hide the bug. The router rejected all three.

    What scored what

    OutcomePoints
    Answered by Big · Medium · Small+3 · +2 · +1
    Broken message rejected+2
    Broken message sent to any model−4
    Sent to a model that had crashed−1
    Sent to a model too small to read it−2
    Good message rejected−1
    Customer gave up after 7 seconds−1

    The router never makes the too-small or broken-message mistakes, and never leaves a customer waiting. It loses points only to crashed models, to messages nothing could answer, and to the few seconds its circuit stays open after a model recovers.

    Why this matters for your agents

    Every model provider has outages, rate limits and retired model versions. An agent that calls one provider directly turns each of those into an incident. Doing what you just did by hand, for every call, at 3am, is a job for software:

    • Fall back on the provider's failure, never on the request's. A 400 that falls back turns a bug into a quality problem nobody notices.
    • Stop calling what's down, and try it again after a pause, so an outage costs one failed call, not one per request.
    • Check the fit before you send. A smaller fallback model shouldn't get a request it can't hold, and the user's latest message should never be the thing that's cut.
    • Record which model answered. In the runtime, every call emits a model.completed event naming the provider and model, so a quiet switch to the fallback shows up on a dashboard rather than in complaints.

    If your agents depend on more than one model, or need to, talk to us.