AWS measures its 38 healthcare agent skills against the same agents without them

A head-to-head win rate answers one question: shown two outputs from the same model, which one did the judge prefer. AWS did not disclose the judge or the number of cases ^[5]. Sixteen points separate the ends of the published band ^[1]. AWS attributes that range to variation in agent harness setup ^[3], so the same markdown file can land at either end depending on whether it runs under Kiro CLI, a Strands agent, or the AgentCore harness ^[16].

Critical thinking is where AWS reports the largest effect: a 78 to 85 percent win rate, with Cohen's d between 0.65 and 1.03 ^[4]. Those two effect sizes are 0.38 standard deviations apart ^[3]. For either end to transfer, your tasks have to sit inside the 11 domains, your judge has to behave like theirs, and your baseline has to be as weak as theirs. Check the baseline condition first. If your system prompt already carries the population frequency thresholds, the unskilled arm is stronger and the gap closes. AWS's own diagnosis is that the misapplication happens even when the guidelines are in the training data and in the system prompt ^[6].

Each file declares triggers, dependencies and metadata in YAML frontmatter under the Agent Skills open standard ^[9], and progressive disclosure pulls the body into context when a trigger matches ^[8]. The trigger strings decide whether the ACMG/AMP evidence categories and computational predictor cutoffs are present when a variant question arrives ^[10]. A query phrased outside them runs the unskilled path.

Coverage is thin per domain. Thirty-eight skills across 11 domains averages under four each ^[4], and the domains named include genomics, drug discovery, claims operations and medical imaging ^[17]. The pipeline skills are specific enough to show what one file buys: variant-calling ships GATK4 HaplotypeCaller commands with annotation groups, VQSR tranche sensitivity targets and Mutect2 tumor-normal configurations ^[11].

RAG retrieves limited passages from indexed documents; these files encode the decision procedure and its error conditions. AWS is careful to say they are not fine-tuning but structured prompts that activate contextually on trigger patterns ^[12].

To run any of it you need Kiro or Kiro CLI, the AWS Strands Agents SDK with Amazon Bedrock model access, the AgentCore harness with an existing agent implementation, or Quick Desktop. AWS also names Claude Code and OpenAI Codex as workable harnesses ^[16].