#9 · ML / AI methods

LLM agents orchestrate bounded, provenance-tracked biology workflows

The common thread is software that encodes domain structure into reproducible, auditable workflows rather than a single sequence-inference method. Fármaco orchestrates scalable drug-screening workflows over a provenance-tracked molecule memory, benchmarked on 22 tasks. GLESKit links gene-level databases, tools and literature into retained workspaces that yield reviewable reports, evaluated on GeneReportBench. LPT constrains LLM agent loops to a fixed orchestration backed by a curated ~14000-paper literature database for de novo protein binder design. BioResearchBench supplies a 50-task benchmark of expert-written research questions for evaluating frontier models. What is new is bounded, provenance-aware orchestration of LLM agents. Performance claims should be read against each paper's own benchmarks, since task domains differ.