Knowledge-Guided Skills Agent for Microbial Colony Detection in Automated Laboratory Testing

Automated microbial colony detection must remain reliable under heterogeneous imaging conditions while preserving deployed detector parameters. We propose ColonyAgent, a knowledge-guided preprocessing framework that combines quality assessment, retrieval of reusable skills and experiences, LLM-assisted workflow planning, and label-free feedback refinement. Experiments on AGAR and a self-built dataset use YOLOv8, RT-DETR, and Faster R-CNN with three independently trained checkpoints per architecture. Relative to direct inference, ColonyAgent improves mean image-macro mAP@0.5 by 0.1419-0.1706 on AGAR and 0.0766-0.1136 on the self-built dataset, and achieves the highest mean among tested external controllers in all four available dataset-detector comparisons. A separate ten-run paired analysis for YOLOv8 and RT-DETR yields Holm-adjusted $p< 0.05$ in all four M4-M1 comparisons. Counting MAE improves or matches direct inference; processing requires about 39-49 s/image, supporting delay-tolerant laboratory use.