MAG It Faster: Bioinformatics Pipeline Efficiencies Attained Through AI-Assisted Coding

Co-assembling and binning hundreds of gigabases of nanopore metagenome reads exposed issues in published bioinformatics tools that do not appear at the scale of their test data. Through AI-assisted maintenance during 2026 we addressed 17 issues in eight published tools: Flye, metaMDBG, LorBin, CheckM, IslandPath-DIMOB, Whokaryote, Bakta and inStrain. All of the affected code predates 2026, so the set records how research software was written before AI assistance and where it breaks. Upstream outcomes range from a fix adopted and released by the maintainer to fixes waiting on repositories with no commits for almost four years. Inside a production pipeline, on a 101 GB freshwater nanopore dataset and with builds that differ only in the third-party code under test, the Flye path from reads to refined bins took 51.0 h before the fixes and 12.2 h after, with identical bins; on a 3.79 GB subset the two builds were within 11%. Most of the affected code was reasonable when written and became slow or wrong through growth in data, drift in the software ecosystem, and composition with other tools. We give practical checks for each cause, an instruction file (AGENTS.md) that encodes them for coding agents, and an example of a ground-up rewrite held to the original's output.