AWS Releases Open-Source Agent Skills for Healthcare AI Reasoning

In a post published on the AWS blog, Michael Hsieh, Principal GenAI Specialist Solutions Architect at AWS, introduced a collection of 38 open-source agent skills spanning 11 healthcare and life sciences (HCLS) domains. According to the post, AI agents built on foundation models often err in applying healthcare decision frameworks, even when they have been exposed to guidelines in training and within the system prompt. When an agent is asked to classify a TP53 missense mutation using ACMG/AMP criteria, it may cite the correct framework but misapply evidence categories, skip population frequency thresholds, or fabricate computational predictor scores. This gap creates silent failures across variant interpretation, claims adjudication, clinical trial design, and imaging analysis, where outputs appear sound but apply incorrect criteria with regulatory and patient safety consequences.

Solution Overview and Skill Structure

The skills in the HCLS Agent Skills collection are structured Markdown documents (SKILL.md) that encode domain decision procedures into a format that AI agents can consume at inference time via progressive disclosure. Following the open Agent Skills standard, each skill declares triggers, dependencies, and metadata in YAML frontmatter, followed by decision frameworks, parameter tables, code patterns, and validation criteria. All 38 skills are released under the MIT-0 license and cover 11 domains, including genomics, drug discovery, claims operations, and medical imaging.

The skills in the repository are classified into two types:

1. Reasoning skills: Encode methodology and decision frameworks that guide how the agent thinks. For example, the genomic-variant-interpretation skill encodes the full ACMG/AMP classification framework, including evidence categories, population frequency thresholds, and computational prediction cutoffs.
2. Pipeline skills: Encode tool-specific commands, validated parameters, and code templates that produce runnable artifacts. For example, the variant-calling skill provides GATK4 HaplotypeCaller commands with correct annotation groups, VQSR tranche sensitivity targets, and Mutect2 configurations for tumor-normal pairs.

Three characteristics distinguish skills from other approaches to domain specialization: they are auditable, portable, and straightforward to maintain. Every decision criterion is human-readable in Markdown format and is not hidden within model weights. A skill operates across more than 20 services (including Amazon Bedrock AgentCore, AWS Strands Agents SDK, Kiro, Amazon Quick Desktop, Claude Code, OpenAI Codex, and others) without customization for each service. Annual medical policy changes or new experiment criteria can be reflected by editing a text file, without requiring model retraining.

Prerequisites and Deployment Patterns

To use the skills, one of the supported AWS services is required (such as Kiro or Kiro CLI, the AWS Strands Agents SDK with access to foundation models in Amazon Bedrock, the AgentCore harness, or Amazon Quick Desktop), or a coding agent of choice (such as Claude Code or OpenAI Codex), along with Python 3.10 or higher with the uv management tool, and Git to clone the repository. Installing only the skills is possible using the command npx skills add awslabs/hcls-agent-skills.

The post outlines three deployment patterns:

• Single-agent use in Amazon Quick Desktop: The agent selectively triggers relevant skills based on trigger patterns in the user query. For example, a question regarding the RAF impact of coding E11.9 versus E11.42 triggers the risk-adjustment skill and returns specific HCC mappings, hierarchy resolution, and quantified RAF deltas instead of a generic suggestion to review documentation.
• Multi-agent architecture in Kiro CLI: Loading all 38 skills into a single agent consumes approximately 80,000 tokens, creating a context engineering challenge where irrelevant content competes for attention. Kiro CLI resolves this using a lightweight coordinator agent (with no loaded skills) that routes queries to eight domain specialist agents, with each specialist loading only its relevant skills (approximately 15,000 tokens per specialist).
• Integration with AWS Strands SDK and deployment to Amazon Bedrock AgentCore: The SDK provides native skill loading in Python code. The AgentCore environment allows skills to be injected at the environment level for hosted agents and provides managed hosting, auto scaling, security boundaries, and observability capabilities without infrastructure management.

Practical Use Cases

The post presents three sample use cases drawn from the evaluation suite:

1. Evaluating drug repurposing candidates for a rare fibrotic disease (IPF): Researchers wishing to evaluate approved drugs that modulate TGF-β1 signaling through TGFBR1. Before skills, the agent provides a general literature review lacking structured ranking criteria. After being equipped with the drug-repurposing and translational-research skills, the agent applies the DGIdb query framework, ranks candidates according to an evidence hierarchy where direct target engagement outweighs pathway-level evidence and phenotypic association, maps TGFBR1 inhibition to IPF pathological processes, and evaluates clinical translatability according to T0→T1 criteria.
2. Building a CMS-HCC risk adjustment pipeline in healthcare claims operations: For a Medicare Advantage plan with 12,000 members calculating RAF scores using CMS-HCC Model V28 coefficients. An agent without skills misses hierarchy resolution or uses outdated V24 coefficients. With the risk-adjustment and claims-billing-rules skills, the agent generates custom SQL queries with deduplication, applies V28 hierarchy resolution (e.g., HCC 18 supersedes HCC 19, and HCC 326 supersedes HCC 327), performs demographic segmentation, and explains that skipping hierarchies inflates scores and creates audit liability under CMS RADV reviews.
3. Preprocessing T1w MRI scans for imaging research: In a study with 45 healthy adults requiring preparation for voxel-based morphometry (VBM) analysis in an FSL/ANTs environment. Without skills, the agent might order bias field correction after skull-stripping, biasing brain masks. With the radiology-preprocessing and imaging-study-design skills, the agent details the correct operational order, explains that intensity inhomogeneity at brain boundaries causes algorithms to remove excess or insufficient tissue if correction is not performed first, and provides a complete Bash script with error checking and quality control.

Empirical Evaluation Results

A pairwise evaluation was conducted across 410 domain prompts (380 single-skill and 30 cross-skill) under two configurations: Kiro CLI (using the Auto model) and Strands SDK (with Claude Sonnet 4.6 in Bedrock). Both conditions were provided with a think tool, testing a baseline agent against an agent with access to all 38 skills via progressive loading. Claude Opus 4.7 in Amazon Bedrock served as the LLM judge across five dimensions: scientific accuracy, coherence, relevance, critical thinking, and actionability.

Key findings from the evaluation include:

• Overall win rate: Skills won 69.5% of comparisons in Kiro CLI (Cohen's d = 0.39 effect size) and 85.9% in Strands Agent (d = 0.97 effect size).
• Critical thinking: The strongest improvement was recorded in the critical thinking dimension, with a 78.0% win rate in Kiro CLI (d = 0.65 effect size) and 85.1% in Strands (d = 1.03 effect size).
• Scientific accuracy: A win rate of 69.3% in Kiro CLI (d = 0.34 effect size) and 86.2% in Strands (d = 0.85 effect size).
• Actionability: A win rate of 68.0% in Kiro CLI (d = 0.37 effect size) and 77.3% in Strands (d = 0.56 effect size).
• Baseline performance correlation: A negative correlation of r = -0.59 in Kiro CLI and r = -0.61 in Strands was found between baseline response quality and the magnitude of skill benefit. On prompts where the baseline agent struggled (score below 80), skills achieved an 87% win rate in Kiro (an 8.7-point lift) and 96% in Strands (a 7.9-point lift).
• Variance reduction: Skills reduced the standard deviation of response scores by up to 61.9% in Kiro CLI and up to 52.1% in Strands. For example, in the clinical data domain in Kiro CLI, the standard deviation dropped from 6.8 to 3.3 (a 51% decrease), indicating higher consistency in outputs.

Customizing and Extending Skills

The repository includes three guides for customizing and building new skills:

• CUSTOMIZING.md: Addresses modifying, extending, and creating skills from scratch, including adding organization-specific rules such as LCD codes or internal protocols.
• SKILL_DESIGN_GUIDE.md: Documents evidence-based patterns for writing effective skills, including decision trees, threshold tables, common pitfalls lists, and response structures.
• QUALITY_CHECKLIST.md: A pre-merge quality checklist covering frontmatter, structure, and content quality.

Additionally, the repository provides an evaluation framework capable of generating 30 prompts per skill and running parallel comparative evaluations that generate an HTML dashboard to review scores and distributions. Because skills are local text files that do not create AWS resources themselves, they require no teardown process after use.