Cost-Aware and Self-Validating Multi-Agent Engineering of Computer-Vision Pipelines: An Optimization-Based Framework with Applications to Remote-Sensing Image Analysis

Remote-sensing image analysis increasingly requires workflows that combine heterogeneous sensors, geospatial preprocessing, deep-learning models, uncertainty estimation, and domain-specific quality control. Conventional computer-vision pipelines are commonly designed manually, making them expensive to develop, difficult to reproduce, and inflexible when sensor quality or task requirements change. Recent agentic-AI systems have demonstrated the feasibility of automatically planning and executing computer-vision and AutoML workflows, but computational cost, geospatial validity, recovery behavior, and provenance remain insufficiently integrated. This paper proposes COSA-CV, a Cost-aware and Self-validating Agentic Computer-Vision framework for automated engineering of remote-sensing image-analysis pipelines. COSA-CV consists of specialized agents for task interpretation, dataset auditing, remote-sensing context analysis, pipeline planning, model and tool routing, image preprocessing, inference, post-processing, validation, error diagnosis, recovery, uncertainty estimation, cost monitoring, and report generation. Pipeline selection is formulated as a constrained multi-objective optimization problem that jointly considers predictive quality, processing latency, GPU usage, memory consumption, energy, tool-call cost, reliability, and human-review burden. The workflow is implemented as a stateful graph with conditional edges that dynamically select optical, SAR, or multimodal processing routes according to image quality, validation results, uncertainty, and computational budget. The evaluation includes two case studies: Sentinel-2 land-cover classification using EuroSAT and multimodal flood-region extraction using SEN12-FLOOD. The numerical results reported here are an illustrative worked example rather than measurements from real experiments, used only to demonstrate the framework’s evaluation protocol and reporting format. In this hypothetical setting, COSA-CV is illustrated as reaching 94.7% accuracy and a 94.3% macro-F1 score for land-cover classification, and 85.2% Dice, 74.3% IoU, and 70.1% boundary F-score for flood-region extraction, with reduced computational cost and invalid-output acceptance relative to non-optimized multi-agent baselines. A final implementation must replace these illustrative values with results from real experiments before any of these comparative claims can be considered validated.