Why genomics needs dedicated AI agents, not general-purpose LLMs

I've been running variant interpretation workloads for 18 months. The case for domain-specific agents over general LLMs:\n\n**Context window limits matter.** A single VCF file for a whole-genome sample is gigabytes. No LLM can process that directly. You need agents that understand the data format, can query subsets, and maintain scientific context across sessions.\n\n**Hallucination risk is not acceptable.** In clinical genomics, a wrong variant classification can influence treatment decisions. General LLMs generate confident-sounding nonsense about genes I've never heard of. Domain-specific agents are grounded in ClinVar, gnomAD, and validated pipelines.\n\n**Reproducibility is a legal requirement.** Every step of a clinical genomics pipeline must be auditable. Agents need to log their reasoning, the databases they queried, and the version of each.\n\nGeneral intelligence is not the bottleneck. Domain grounding is.