On routing, and what it taught us about inference

The first generation of the commercial Internet inherited a problem that the academic networks of the 1980s had only begun to face. Hosts wanted to talk to other hosts, but the set of reachable hosts was no longer enumerable in a single routing table. The community responded by inventing protocols that did not try to know everything — they only tried to know enough, locally, to make the next correct hop.

The current generation of inference frameworks faces a structurally similar problem, and has not yet drawn the same conclusion. Retrieval augmented generation, the Model Context Protocol, and the agentic toolchains built on top of both are gestures toward a single underlying need: a specialist system, somewhere in the network, knows how to answer this query better than the model in front of the user does. The question is which specialist, and how to find it. This is a routing problem.

What routing knows

Border Gateway Protocol does not require any participant to hold a global view. Each autonomous system advertises the prefixes it can reach, listens to its peers’ advertisements, and selects the best route on the basis of locally configured policy. The aggregate of these local decisions produces a coherent path between any two endpoints on the public Internet. It is not an elegant protocol. It is a robust one. It has been incrementally extended for thirty years without ever requiring a flag day.

Open Shortest Path First takes a different approach within a single administrative domain. It floods link state information so that every router can compute the same shortest path tree. The trade-off is well understood: convergence is faster, scalability per area is lower. The two protocols compose because the conditions under which each works well are crisply specified.

Multi-Protocol Label Switching sits underneath both. Once a path is selected, traffic flowing along it is labelled, and the label drives forwarding at the next hop without revisiting the routing decision. This separates the question of which path from the question of how to traverse it efficiently, and it is the basis for traffic engineering, virtual private networks, and the kind of operational guarantees that production networks need.

What inference is rediscovering

The current vocabulary in the inference layer reinvents pieces of this without reaching for the precedent. Tool calls in agentic frameworks are best-effort route advertisements with no convergence guarantee. Vector store retrieval is shortest-path computation against an opaque distance metric that nobody can audit. The Model Context Protocol is a transport without a routing plane.

None of this is wrong. The work is necessary and the people doing it are good. But there is a piece missing, and the shape of the missing piece is recognisable to anyone who has run a peering relationship: there is no mechanism by which a specialist system can announce its capabilities to a network of peers, no mechanism by which a query can be routed on the basis of those announcements, and no mechanism by which the resulting path can be verified, cached, or torn down under load.

What the protocol layer would look like

The contribution Sonata42 is working toward is not a new model, a new training method, or a new agentic framework. It is the missing protocol layer: a capability advertisement format, a route selection algorithm, a session establishment handshake, and a label-switching forwarding plane that lets specialist generative systems federate without any of them needing to hold a global view of the network.

The early sketches look more like RFC 4271 than like a model card. That is the right shape for the problem. The lessons are in the precedent, and the precedent has not yet been read carefully by the people who most need it.