Recent advances in large language models have enabled agentic systems that autonomously plan, reason, and act on problems in Earth Observation (EO) and the geosciences. By interacting with data catalogues, models, processing services, and computational environments, these systems move beyond generating answers to constructing and executing multi-step scientific workflows. This shift raises fundamental questions about how such systems can be designed to be not only capable, but also grounded, verifiable, reproducible, and trustworthy.

This session invites contributions on the design, integration, evaluation, and responsible deployment of agentic AI across EO and the geosciences. We welcome work on how agents reason over long-horizon tasks, interact reliably with heterogeneous data and computational infrastructure, and maintain coherent state, context, and provenance across multiple steps. Topics include agent architectures and orchestration, multi-agent systems, memory and self-evolving mechanisms, as well as protocols and shared conventions that enable agents to discover, communicate with, and use scientific tools and services.

A central focus is the evaluation and verification of agentic scientific reasoning: not only final answers, but also the trajectories through which agents reach them, including tool selection, planning, code execution, intermediate results, and provenance. Contributions addressing uncertainty, hallucination, and failure modes are particularly encouraged, as are approaches that promote physical consistency through domain knowledge, foundation models, retrieval, world models, and physics-aware constraints. Work on human oversight, intervention, and verification is also relevant, particularly for autonomous systems deployed in operational or high-consequence settings.

More broadly, the session aims to advance agentic AI as a paradigm for computational Earth Observation and the geosciences. By bringing together AI researchers, geoscientists, and providers of scientific data and computational infrastructure, we seek to characterize the capabilities and limitations of agentic systems, develop principled methods for evaluating and improving them, and establish what is required for autonomous scientific workflows to operate reliably and reproducibly.