- LLM-driven agents now pull satellite imagery from cloud catalogs using plain English.
The paper introduces a three‑agent stack: Guardrail enforces policy, General‑QA interprets user intent, and Recommender‑Analyst builds schema‑aware API calls. The system translates natural‑language queries into structured requests, then returns remote‑sensing data for tasks like flood monitoring or climate studies. Preliminary adversarial tests show that prompt‑level safety cues reduce obvious failures, but rare, high‑impact API manipulation bugs still occur.
This matters because it bridges the gap between conversational AI and Earth‑observation services, potentially automating data pipelines that currently require manual scripting. However, the lingering safety gaps highlight that autonomous data retrieval cannot rely solely on prompt engineering; robust guardrails are needed to prevent misuse or costly errors.
Even with the Guardrail, the prototype reminds us that AI‑driven access to critical infrastructure remains a work in progress.