#35 · Tools & infrastructure

LLM agents orchestrate genomics tools; biological adjudication remains unreliable

A common thread across these preprints is an LLM agent layer placed over deterministic bioinformatics software. BaseCamp applies an agentic LLM framework to decisions in DNA sequencing pipelines, where established tools still perform alignment, variant calling and annotation. Genome Skeptic pairs deterministic tools with an LLM controller for bacterial tet gene annotation and reports that the controller orchestrates tasks well but adjudicates biological questions poorly, a distinction that frames the central open problem. Murmurent addresses the infrastructure beneath such agents for biomedical labs, adding multi-user projects, specialized agents, tiered memory and traceability. What is new is less the use of LLMs than the explicit separation of orchestration competence from biological judgment, and the engineering needed to make agent workflows auditable in shared lab settings.