Exploring GLP-1 Drug Safety from Multiple Angles: A Toolkit for Evidence Review and Pharmacovigilance

Assessing GLP-1 receptor agonist safety requires evidence from clinical studies, spontaneous reports, and biological mechanisms. We demonstrate a multi-agent pipeline built on Claude Science, with reusable tools for risk-of-bias assessment, GRADE certainty, and analysis of the FDA Adverse Event Reporting System (FAERS). The tools connect source quotations, statistical code, and human review of unresolved assessments. The pipeline reproduces a published trial meta-analysis (risk ratio 0.84, 95\% CI 0.54--1.32). The same 912 semaglutide reports yield reporting odds ratios from 0.51 to 3.29 across ten comparison groups; with the comparison group fixed, the ratio changes from 0.85 before July 2023 to 2.98 afterwards. A small human audit finds numeric extraction accuracy of 2/11 versus 6/11 across two models, despite 5/5 direction agreement for both. The demonstration connects evidence review, drug monitoring, and exploratory repurposing with explicit uncertainty and error reporting. The pipeline and reusable skills are released under the Apache 2.0 licence.