ONR proposes that instead of keeping an explicit list of "who talks to whom" in an agent swarm, each agent gets a position in a learned space, and its connections are simply its nearest neighbours there. When the swarm's job changes or agents drop out, positions drift to lower one shared loss function and the connections follow on their own.
This page is a toy simulation of that idea with 40 made-up agents. It compares two ways to keep a swarm connected: a coordinator that rebuilds the graph and broadcasts it, and ONR-style local updates. It counts the messages each approach sends in this toy only.
No benchmark, code or paper was attached. The last claim is internally in tension: a global loss needs some shared signal, which this toy makes visible as neighbour reads.
Colours are agent roles. The ringed squares are task anchors that agents of that role should stay near.
A real version needs a definition of the loss (what "task-relevant invariant" means for real agents), a way for each agent to compute its gradient without seeing the whole swarm, and a test harness that measures message counts and latency against a plain baseline. The toy shows that local updates still cost neighbour reads every tick; whether that beats a coordinator depends on how often the topology really changes.
For a fleet of five agents on a few VPS servers, the full graph has ten possible links, so there is no topology problem for this to solve.