PigGPA: A Domain-Constrained Agent for Porcine Genomic Breeding with Verifiable Tool Execution

Abstract

Large language model agents can plan and execute scientific analyses, but successful task completion does not establish whether analytical decisions remain consistent across independent sessions or whether computations follow the parameters specified beforehand. We investigated these limitations using four genomic analysis tasks. General-purpose agents showed limited consistency in analytical decisions across repeated sessions (57.5-58.8%). Incorporating domain knowledge improved analytical performance, with reusable skills increasing mean evaluation scores from 73.44 to 89.06/100 , but did not resolve decision variability or ensure agreement between specified and executed parameters. To address these limitations, we developed an approach combining domain-constrained analytical skills with call-level execution contracts. Each tool invocation is linked to a pre-execution parameter declaration through SHA-256 hashing, with actual command-line arguments and execution records subsequently checked against the declared parameters. We implemented this approach in PigGPA, an autonomous agent for porcine genomic breeding that integrates statistical genetics, sequence-based functional prediction and multi-omics annotation. PigGPA completed all benchmark tasks with full agreement between declared and executed parameters, while its core quantitative-genetic results closely matched established software (r = 0.999-1.000). An application to loin muscle depth further demonstrated an integrated workflow from genetic association analysis to functional interpretation. These findings show that domain knowledge improves analytical performance, while direct verification of tool execution provides an additional basis for reproducible scientific computation.