The Pulse
AWS Releases 38 Open Agent Skills for Healthcare AI
AWS has released 38 open-source agent skills covering 11 healthcare and life sciences domains. The company says skilled agents won up to 85.9% of head-to-head evaluations against agents without the skills.
AI.info Team ·
AWS says its new healthcare AI package improved agent performance in as many as 85.9% of head-to-head comparisons, after adding 38 open-source skills to agents tested across 410 domain prompts. The collection targets a specific weakness in foundation-model systems: producing answers that cite the right medical or scientific framework but apply its rules incorrectly.
Michael Hsieh, a principal GenAI specialist solutions architect at AWS, introduced the collection in an AWS Machine Learning Blog post published September 16, 2026. The skills cover 11 healthcare and life sciences domains, including genomics, drug discovery, claims operations and medical imaging. AWS releases them under the MIT-0 license through the HCLS Agent Skills repository.
38 Markdown Files for Domain Decisions
Each skill is a structured SKILL.md file that agents load progressively when a query matches its triggers. The files include decision frameworks, parameter tables, code patterns and validation criteria, with metadata and dependencies defined in YAML frontmatter.
AWS divides the collection into reasoning skills and pipeline skills. Reasoning skills encode procedures for decisions such as genomic variant interpretation, while pipeline skills provide validated commands and templates for technical work. The genomic-variant-interpretation skill, for example, covers evidence categories, population-frequency thresholds and computational-predictor cutoffs used in the ACMG/AMP classification framework. A separate variant-calling skill provides configurations for GATK4 HaplotypeCaller, VQSR and Mutect2 tumor-normal workflows.
AWS describes the files as structured prompts rather than fine-tuning. The distinction matters because teams can inspect and revise the decision criteria directly instead of changing model weights. The company also says the skills can run across more than 20 services and tools, including Amazon Bedrock AgentCore, AWS Strands Agents SDK, Kiro, Amazon Quick Desktop, Claude Code and OpenAI Codex.
Where Unguided Agents Fail
The package is designed around errors that can look credible to a non-specialist. AWS uses TP53 variant classification as one example: an agent may cite ACMG/AMP correctly while misapplying evidence categories, skipping population-frequency thresholds or inventing computational predictor scores.
The same problem appears in claims processing. In AWS’s example, an agent building a Medicare Advantage risk-adjustment pipeline may use outdated coefficients, omit disease-hierarchy resolution or sum conditions at the wrong stage. The risk-adjustment skills direct the agent to apply the ICD-10-to-HCC crosswalk, deduplicate diagnoses and resolve hierarchies before calculating member-level scores.
A medical-imaging example focuses on T1-weighted MRI preprocessing for voxel-based morphometry. AWS says the relevant skills require reorientation, bias-field correction, skull stripping and registration to MNI152 space in the correct sequence, while also documenting failure modes and quality-control outputs.
A 410-Prompt Test Across Two Agent Setups
AWS evaluated the collection with 410 prompts: 380 using a single skill and 30 involving multiple skills. The comparison used two configurations, Kiro CLI and an agent built with the AWS Strands Agents SDK. The Strands setup pinned the model to Claude Sonnet 4.6, loaded all 38 skills through progressive loading and gave both the skilled and baseline conditions access to a thinking tool.
AWS used Claude Opus 4.7 through Amazon Bedrock as the judge. It scored scientific accuracy, coherence, relevance, critical thinking and actionability on a 0-to-100 scale, then reported win rates and Cohen’s d effect sizes to account for score compression in language-model evaluations.
The skilled agent posted an overall win rate of 69.5% in Kiro CLI and 85.9% in the Strands configuration. Critical-thinking win rates reached 78% and 85.1%, respectively. Scientific accuracy reached 69.3% in Kiro and 86.2% in Strands, while actionability reached 68% and 77.3%.
AWS says the results point to a methodological benefit rather than a simple increase in factual recall. The skills tell an agent which framework to apply, which assumptions to question and which limitations to disclose. The reported effect was strongest on prompts where an unguided agent could produce a plausible answer while missing a required decision step.
Context Size Shapes the Deployment Design
Loading all 38 skills into one agent consumes about 80,000 tokens, according to AWS. That approach works with models offering large context windows, but it also forces the agent to select relevant material while unrelated instructions compete for attention.
AWS’s Kiro CLI design uses a lightweight coordinator and eight domain specialists. Each specialist loads only its assigned skills, using about 15,000 tokens, while the coordinator handles intent classification and routing. The same collection can be loaded directly in Strands code or added to agents hosted through Amazon Bedrock AgentCore.
For developers who want to try the files without the surrounding agent configurations, AWS documents the command npx skills add awslabs/hcls-agent-skills. The repository also includes guides for customizing thresholds, adding organization-specific policies and testing modified skills with generated evaluation prompts.
Open Files, Not a New Medical Model
AWS’s release does not introduce a new foundation model or claim that the skills replace clinical or scientific review. It supplies auditable text files that shape how an existing agent reasons through specialized tasks and produces commands or analysis.
That makes maintenance part of the release’s practical appeal. A health plan could modify rules for internal coding procedures, while a research group could add experiment-specific thresholds or version-dependent tool settings without retraining a model. The evaluation results remain AWS’s own benchmark, but the files, prompts and testing framework are available for inspection in the open-source repository.
The immediate deliverable is therefore concrete: 38 MIT-0-licensed skill files, installation paths for Kiro, Quick Desktop and Strands, and deployment guidance for Bedrock AgentCore. AWS’s published test reports a 69.5% to 85.9% overall win rate for agents using them across 410 prompts.