Improving HCLS AI reasoning with open-source agent skills

AWS is trying to fix a very specific failure mode in healthcare AI: models that know the right framework but still use it badly. A system can recite ACMG/AMP rules for a TP53 variant and still skip population thresholds, mix up evidence categories, or invent predictor scores. The result looks polished and can still be wrong in ways that matter for safety and regulation.

The company’s answer is a collection of 38 open-source agent skills across 11 healthcare and life sciences domains. They’re markdown files called SKILL.md, with YAML frontmatter for triggers, dependencies, and metadata. The idea is simple enough: encode the decision procedure itself, not just fragments of reference text. AWS says the collection is meant to work across more than 20 services, including Amazon Bedrock AgentCore, AWS Strands Agents SDK, Kiro, Amazon Quick Desktop, Claude Code, and OpenAI Codex.

AWS splits the skills into two groups. Reasoning skills carry the methodology, like the genomic-variant-interpretation skill that lays out ACMG/AMP classification. Pipeline skills carry the technical bits, such as GATK4 HaplotypeCaller commands, VQSR tranche sensitivity targets, and Mutect2 tumor-normal setups. That distinction matters because it gives agents both the judgment and the runnable steps, without pushing the whole thing into model weights or into a RAG lookup.

The deployment story is broad. You can install the skills with a CLI, attach them to Kiro, wire them into the AWS Strands Agents SDK, or use them in AgentCore and Quick Desktop. AWS also leans on progressive disclosure, so the agent only activates the relevant skill when a query matches the trigger pattern. In Quick Desktop, for instance, a question about RAF impact can pull in the risk-adjustment skill and return HCC mappings and quantified RAF deltas instead of a vague “review documentation” answer.

The part that will get people’s attention is the evaluation. AWS says agents with these skills win 70–86 percent of head-to-head comparisons against the same agents without skills, depending on the harness. The strongest results show up in critical thinking, where the win rate lands at 78–85 percent with a reported d of 0.65–1.03. That’s not a magic trick. It’s a sign that a lot of “AI reasoning” problems are really “did the model follow the right procedure” problems.