README.md

We propose OpenEarth-Agent, the first remote sensing agent architecture designed for open environments. By performing adaptive workflow planning and tool creation conditioned on data and task contexts, OpenEarth-Agent accommodates diverse EO data and tasks. Through multi-stage tool and cross-domain knowledge integration, it effectively executes full-pipeline EO across multiple application domains.

We construct OpenEarth-Bench, the first remote sensing agent benchmark oriented towards open environments. Comprising $596$ full-pipeline cases sourced from real-world applications across various domains, it serves as a robust platform for evaluating agent performance in open environments.

Extensive experiments on OpenEarth-Bench and cross-benchmark evaluations on Earth-Bench validate the effectiveness of OpenEarth-Agent. Notably, on Earth-Bench, OpenEarth-Agent demonstrates the ability to create functionally equivalent professional tools, and in certain instances, tools with superior data adaptability. Operating with merely $6$ integrated tools, it achieves performance comparable to agents utilizing full tool calling, and significantly outperforms existing agents when provided with the complete toolset.