Last decade’s cloud‑native design rests on a 20‑year assumption: state lives in the database, compute stays stateless. Modern LLM agents break that mold with long‑running tasks, built‑in context, and interactive control, forcing developers to rethink routing and persistence.

Agents no longer fit the request‑response mold. A 10‑minute job, multi‑turn dialogue, or tool‑chain execution requires state that exists in the agent’s memory, not a database. This hidden state clashes with the stateless service model, making traditional load balancers and database‑centric routing brittle.

Durable execution systems such as Temporal, Inngest, and Restate add resilience but still hide the routing problem behind polling. Clients must query a database endpoint for updates, turning the database into a makeshift message bus and adding latency, load, and a poor user experience.

A missing primitive emerges: a routable transport name that identifies a process, not a server. Pub/sub channels solve this by letting both client and server connect to a named channel, preserving state across disconnects. This approach keeps LLM interactions efficient, cost‑effective, and reliable without re‑threading data through a database.