A new research framework lets drone swarms plan and act as a team without funneling every decision through one central computer.
Researchers built AeroWeaver, an open-source harness that pairs large language model agents with individual drone skills like searching, inspecting, or tracking a target. Instead of one brain computing joint moves for the whole swarm, each drone runs a role-conditioned local agent that picks from a governed set of executable skills. The system logs state-action-reward experience by role and uses that history to refine which skill a drone picks next, without retraining the underlying model. The team reports that testing kept skill execution valid and let multiple drones coordinate without a central agent generating joint actions from a global view.
Most LLM-agent swarm work still leans on a single planner reasoning over the entire mission, which is both a bottleneck and a single point of failure if that link drops. AeroWeaver's bet is that turning language-model reasoning into per-drone, per-role skills keeps a swarm working even when no single agent holds the full picture. That is a harder, more useful problem than getting one model to narrate a plausible-sounding flight plan.
Training-free adaptation from a set of tested runs is a promising demo, not proof it survives wind, jamming, or the messier edge cases real missions throw at it.