An AI agent can now read a plain-English request like "prioritize low latency for this video call" and turn it into actual routing decisions on a satellite network.
Researchers built STR-Agent, a framework that uses a large language model to manage quality-of-service routing in low-Earth-orbit (LEO) satellite networks. The system has four parts working together: one module translates natural-language requests into structured routing rules, another executes those rules using network tools, a third module remembers past outcomes, and a fourth reflects on real-time congestion to adjust its strategy on the fly. The team also built a custom perception model and a fine-tuning dataset specifically for understanding LEO satellite service requests. In simulations on a Walker-Delta constellation, a common satellite arrangement pattern, STR-Agent cut end-to-end delay by up to 60% compared with a conventional routing baseline called DQ-Dijkstra, and fine-tuning pushed intent-understanding accuracy from 45.4% to 92.45%.
This matters because LEO constellations like Starlink and Kuiper are becoming real infrastructure, not just experiments, and their routing has always been rigid: fixed objectives applied to a network whose topology changes by the minute. Letting an LLM translate what a service actually needs, then re-adapt as congestion shifts, is a meaningfully different approach than hand-tuned heuristics.
The catch is that this is simulation work, not a deployed system, and it adds an LLM's latency and compute overhead to a network already fighting the clock.