Agents Router

A single lightweight process hosting the weather, time, and currency agents as three sub-applications behind one path-prefixed address, instead of three separate deployments. It does no reasoning, no model routing, and holds no registry or budget of its own — it’s the same three agents' own code, just packaged to run as one process and reached at one internal address.

Nothing outside the cluster can reach this router, and nothing inside the cluster can reach it except the orchestrator — enforced at the network-policy level, not by any key or token on the call itself. That’s a deliberate simplification from the old gateway’s per-app virtual-key model: there’s exactly one legitimate caller, so a credential on top of network policy would be redundant.

Before this router existed, weather, time, and currency each ran as their own standalone deployment — three processes, three sets of infrastructure, for genuinely tiny workloads. Bundling them behind one process cuts that overhead to one deployment while leaving each agent’s own code and behavior untouched; from the orchestrator’s point of view, a delegated call looks exactly the same as it did when each agent ran standalone.

Unlike the shared gateway it partially replaced, this router doesn’t do reasoning, doesn’t hold a model registry, doesn’t scope or budget per-app credentials, and doesn’t trace calls. The orchestrator’s own reasoning, and every other AI/TTS call in the fleet, goes straight to the outside provider instead — see LLM Gateway.