Beyond detection: cooperative multi-agent reasoning for rapid onboard EO crisis response

The most resource-constrained place to run multi-agent AI is probably not a server rack in a hyperscaler data center. It is a 16-core ARM board with 32 GB of RAM, traveling at 28,000 kilometers per hour, 400 kilometers above your head.

Researchers at Thales Alenia Space, the Franco-Italian aerospace and defense company, have published a preprint on arXiv describing a hierarchical four-agent pipeline for onboard disaster detection—wildfire and flood—running on IMAGIN-e, an edge computing payload that has been operational aboard the International Space Station since March 2024. The paper was accepted at the ESA 4S Symposium 2026, the European Space Agency's biennial small satellites conference.

The lead author is Alejandro Mousist, a Thales Alenia Space researcher who previously built what his team describes as the first in-orbit agentic AI system—ASTREA, an LLM-driven autonomous thermal management agent also tested on IMAGIN-e in real flight, presented at the AISTAR 2025 conference. This is not a researcher spinning up a cloud benchmark and calling it satellite infrastructure. The hardware in question is in orbit.

The architecture the paper describes is a diamond topology. An Early Warning Agent, running a 4-bit quantized Qwen2-VL 2B vision-language model from Alibaba's open-source Qwen lineup, screens incoming Sentinel-2 RGB satellite imagery and decides whether a scene warrants deeper analysis. If it does, it routes to one of two domain specialists—a Wildfire Specialist Agent or a Flood Specialist Agent—each equipped with traditional remote sensing tools (spectral indexes like NBR and NDWI) and a shared Qwen-2.5 3B 8-bit language model to interpret results. A Decision Agent at the end of the diamond fuses the evidence and issues a final alert.

The routing is the whole point. The baseline approach—always running both specialist agents regardless of what the Early Warning Agent sees—is computationally expensive and largely wasteful on orbiting hardware that has to prioritize power and thermal budgets. The paper reports a 4.78x speedup on non-disaster scenes, where the Early Warning Agent correctly decides there's nothing to escalate, reducing compute by 73.2 percent. On actual disaster scenes, where both specialists activate, the speedup is 1.3x. That asymmetry is actually the right result: most scenes captured by an Earth observation satellite are not disasters.

The broader context here is the shift a January 2026 survey on arXiv framed as the move from static perception pipelines to autonomous, tool-augmented, goal-directed Earth observation systems. Selective routing is a core pattern in that shift—the expensive model only fires when the cheap model says it needs to. The Mousist team is applying that pattern in one of the few deployment environments where it isn't optional.

Distributed satellite AI isn't new as a concept. NASA's Starling mission, a four-satellite CubeSat swarm launched in July 2023, demonstrated inter-satellite networking and onboard autonomous decision-making using the agency's Distributed Spacecraft Autonomy software. What this paper adds is the integration of large language models and vision-language models into that onboard stack, and the role-specialization pattern that makes it tractable under the memory and compute constraints of actual space hardware.

The paper uses Prefect, a Python-based workflow orchestration tool, to coordinate the agents in the proof-of-concept implementation. Prefect is not deployable on actual satellites—the paper acknowledges this, noting that real inter-satellite links are narrowband, high-latency, and intermittent, not the REST endpoints used in the experiment. That's the right kind of honesty for a proof of concept: the diamond topology and routing logic are the transferable artifact; the orchestration layer is scaffolding that would need replacement for flight.

The test set is 27 samples, which is thin. The flood model achieves an intersection-over-union score of 0.554 on the Copernicus Emergency Management Service dataset—usable, but not production-hardened. The paper is a proof of concept, and the authors say so. The ISS deployment is for validation purposes, not operational disaster response. Those caveats matter.

From an infrastructure standpoint, the model selection is worth noting. Both models—Qwen2-VL 2B and Qwen-2.5 3B—are from Alibaba's open-source Qwen family, aggressively quantized to fit the IMAGIN-e memory envelope. The vision-language model handles the visual triage; the language model handles text-based reasoning over tool outputs. That's a clean separation that should be legible to anyone who's built a tool-calling pipeline: the VLM replaces the visual intake step, and the LLM still does the structured reasoning. The difference is that the latency budget is measured in mission windows, not user sessions, and there is no retry button.

For satellite operators and defense contractors, the agent coordination framework is the deliverable—not the specific models, which will be swapped as the field advances. The paper's contribution is demonstrating that role-specialized agents with selective routing can run within the constraints of commercially available space edge hardware, and that the pattern produces meaningful compute savings without degrading decision quality on the cases that matter. That combination—feasibility on real hardware, measurable savings on real data—is what separates this from the majority of multi-agent papers that never leave a cloud VM.

The preprint was posted to arXiv on March 20, 2026, and accepted for presentation at ESA 4S Symposium 2026.