An Empirical Study of Agent Skills for Healthcare: Practice, Gaps, and Governance

Technical Summary: An Empirical Study of Agent Skills for Healthcare: Practice, Gaps, and Governance

Problem Statement
Healthcare automation is inherently constrained by local procedures and organizational variations, meaning agent capabilities rarely transfer unchanged across different settings. While AI agents can combine context, tools, and reasoning to support clinical workflows, their behaviors cannot be treated as fixed. There is a need for a mechanism to package procedures for reuse and adaptation without rebuilding entire systems. "Agent skills"—self-contained directories containing reusable procedures—have emerged as a procedural layer to address this. However, the current state of the public healthcare skill ecosystem, its alignment with research priorities, and its governance regarding clinical risk remain largely uncharacterized.

Methodology
The authors conducted the first empirical analysis of healthcare agent skills using a corpus derived from ClawHub, one of the largest public agent skill platforms.

• Data Collection: From a snapshot of 58,159 publicly listed skills (April 20, 2026), the authors identified 557 healthcare-related skills using a GPT-5-mini classifier applied to skill names and developer descriptions. Healthcare skills were defined as those concerned with clinical care delivery, medical operations, or life sciences, excluding general wellness or lifestyle coaching without clinical framing.
• Annotation Protocol: Each of the 557 skills was annotated along ten dimensions covering function, deployment context, autonomy, and safety. The annotation process utilized GPT-5.4 as a scalable classifier over the full SKILL.md files (including name, description, and Markdown body). The taxonomy was refined through manual inspection to reduce ambiguity in healthcare-specific cases.
• Dimensions Analyzed: The study examined artifact size, authorship, adoption, intended users, linguistic/geographic framing, functional distribution, healthcare lifecycle stage, input modalities, autonomy levels, clinical impact, general technical risk, and the presence of explicit safety-boundary statements.

Key Results
The analysis reveals three primary findings regarding the current state of healthcare agent skills:

1. Divergence Between Practice and Research:

   • Focus: Public healthcare skills are predominantly patient-facing, concentrating on post-diagnosis workflow automation, research support, and personal health monitoring. In contrast, healthcare-agent research literature heavily emphasizes diagnosis- and treatment-oriented reasoning.
   • Function: Workflow automation is the most common skill category (33.2%), whereas diagnosis is the leading category in research papers (40.0%) but appears in only 22.4% of skills.
   • Users: The largest intended-user category for skills is patients and general consumers (37.1%), surpassing clinicians (20.3%) and researchers (22.4%).

2. Uneven Coverage of the Healthcare Lifecycle and Inputs:

   • Lifecycle: Skills are concentrated in System Operations (293 skills), particularly Research & Development (27.1% of the corpus), and Patient Monitoring & Rehabilitation (33.0% of patient-facing skills). Critical stages like acute diagnosis and treatment planning are underrepresented.
   • Inputs: Skills favor general-purpose inputs compatible with standard LLM interfaces. Structured forms (143 skills) and text conversations (115 skills) are the primary modalities. Specialized clinical inputs are sparse: medical images and physiological signals appear as primary inputs in fewer than 2% of skills (9 each), and genomic data in only 13.

3. Misalignment Between Technical Risk and Clinical Impact:

   • Autonomy vs. Impact: The ecosystem is dominated by L3 (Delegated Execution) skills (244 total), where agents complete bounded digital tasks that may influence user decisions without real-world actuation. However, high-autonomy skills (L4/L5) exist; notably, one L5 skill acts as an "emergency guardian" capable of executing rescue actions.
   • Risk Assessment: General technical risk frameworks classify most skills as "Safe" or "Moderate Risk." However, this does not reliably capture clinical risk. Many skills with limited tool access still produce outputs that influence clinical judgment.
   • Governance Gaps: Only 163 of 557 skills include explicit boundary statements or disclaimers. Most lack clear scope definitions, leaving agents with limited criteria for qualifying their outputs.

Significance and Claims
The paper positions healthcare skills as a critical procedural layer that is currently not adequately captured by existing benchmarks or risk frameworks.

• Benchmarking Gap: Current benchmarks often favor computable outcomes (e.g., diagnostic accuracy), whereas deployable skills depend on local workflow fit and downstream actions, making them harder to evaluate outside specific deployment contexts.
• Governance Implications: The authors argue that skill governance must separate clinical impact from technical permission risk. A skill may be technically "safe" (no destructive code execution) but clinically consequential (shaping patient or clinician judgment).
• Future Direction: The study suggests that for agent skills to mature in healthcare, the ecosystem requires clearer quality signals (validation status, provenance), explicit boundary statements, and lower barriers for clinicians and domain experts to contribute procedural knowledge, rather than relying solely on technically active developers.

The authors conclude that treating agent skills as concrete, reviewable artifacts offers a practical surface for evaluation and governance, distinct from the implicit behavior of model weights. They acknowledge limitations, noting the corpus is from a single platform at a single time point and that annotations infer properties from descriptions rather than observed deployment behavior.