Researchers built a drone swarm where every aircraft runs its own small language model - and the hard part wasn't the AI, it was getting the drones to stop talking so much.
The paper describes a fully distributed UAV swarm architecture: each drone runs an independent small language model (SLM) in a continuous reason-act-observe loop, with no central coordinator. Knowledge is stored as structured atomic notes split into core, local, and peer-specific memory, and a deterministic gossip engine decides what to share with which neighbor based on how new and relevant that information is to them. The team tested this with ten UAVs in a simulated search-and-rescue mission. Their approach finished every single run, while a swarm that flooded all information to everyone completed only 70-85% of runs, and letting the language model itself decide what to forward failed every time.
That gap matters because distributed swarms exist precisely to avoid a single point of failure - but without a coordinator, something still has to manage who knows what, and that something has real costs. The selective-gossip approach roughly halved inference-token consumption and cut transmitted data versus flooding, while also producing a more accurate survivor count, which is the actual point of a search-and-rescue mission.
Ten drones in a simulation is a modest testbed, and simulated radios don't drop packets the way real ones do over a disaster zone. Still, the result argues that smarter memory management, not bigger models or more bandwidth, is what makes agentic swarms actually finish the job.