AWS published a collection of 38 open-source agent skills spanning 11 healthcare and life sciences domains, encoding domain decision procedures (like ACMG/AMP variant classification) as structured markdown documents that agents load at inference time. Agents equipped with the skills won 70–86% of head-to-head comparisons against unskilled agents, addressing a gap where models know facts but misapply clinical frameworks. The skills are MIT-0 licensed and work across AWS services and third-party coding agents.
AWS published a framework called Trust in Depth for granting AI agents more autonomy while keeping control, applying Defense-in-Depth principles to agent governance. It pairs a decade of AWS automated-reasoning production experience with new agent tooling: Bedrock AgentCore policies, AWS Agent Registry, Strands, and CloudWatch Omni. The approach lets agents act independently when rules can be formally proven and reserves human judgment for genuinely ambiguous cases.
AWS published a framework called Trust in Depth for granting AI agents more autonomy while keeping control, applying Defense-in-Depth principles to agent governance. It pairs a decade of AWS automated-reasoning production experience with new agent tooling: Bedrock AgentCore policies, AWS Agent Registry, Strands, and CloudWatch Omni. The approach lets agents act independently when rules can be formally proven and reserves human judgment for genuinely ambiguous cases.
An arXiv paper presents an autonomous LLM agent combining ReAct and RAG to classify genetic disease severity across 10,211 Human Phenotype Ontology terms using ACMG and ACOG guidelines with PubMed retrieval and claim verification. The system hit 93.55% phenotype-level accuracy and 95.2% concordance with the Mackenzie's Mission gene list, enabling automated, evidence-backed genetic screening panel design.
An arXiv paper presents an autonomous LLM agent combining ReAct and RAG to classify genetic disease severity across 10,211 Human Phenotype Ontology terms using ACMG and ACOG guidelines with PubMed retrieval and claim verification. The system hit 93.55% phenotype-level accuracy and 95.2% concordance with the Mackenzie's Mission gene list, enabling automated, evidence-backed genetic screening panel design.
An AWS post demonstrates deploying Hugging Face models to SageMaker AI using six open-source agent skills from the Hugging Face Skills repo, working with Kiro and Claude Code. Unguided agents chose outdated TGI containers and produced failed or silently broken deployments; the skills select vLLM from the AWS DLC catalog and add autoscaling, alarms, and teardown.
An AWS post demonstrates deploying Hugging Face models to SageMaker AI using six open-source agent skills from the Hugging Face Skills repo, working with Kiro and Claude Code. Unguided agents chose outdated TGI containers and produced failed or silently broken deployments; the skills select vLLM from the AWS DLC catalog and add autoscaling, alarms, and teardown.
Two August 2026 arXiv papers tackle agent skills from complementary angles. SkillAlchemy (Aug 24) studies open-world skill creation, where a creator agent must discover behavior-relevant requirements omitted from underspecified skill briefs. SPT (Aug 27) introduces Skill Pre-Training, a mid-training method applying causal language modeling to public skills, treating them as pre-training data rather than inference-time context.
Two August 2026 arXiv papers tackle agent skills from complementary angles. SkillAlchemy (Aug 24) studies open-world skill creation, where a creator agent must discover behavior-relevant requirements omitted from underspecified skill briefs. SPT (Aug 27) introduces Skill Pre-Training, a mid-training method applying causal language modeling to public skills, treating them as pre-training data rather than inference-time context.
Researchers introduce DMoA, a multi-agent framework that structures role-based interaction among LLMs to support iterative diagnostic reasoning, addressing the limits of single-turn question-answering in complex clinical settings. They evaluate it against base models on 297 rare disease cases and 1,719 challenging cases, targeting diagnostic decision support.
Researchers introduce DMoA, a multi-agent framework that structures role-based interaction among LLMs to support iterative diagnostic reasoning, addressing the limits of single-turn question-answering in complex clinical settings. They evaluate it against base models on 297 rare disease cases and 1,719 challenging cases, targeting diagnostic decision support.