---
schema_version: "newruntime-agent-readable-v0.1"
type: "trend_pattern"
slug: "skills-become-portable-capability-layer"
title: "Skills become a portable capability layer"
description: "Instructions, scripts, tools, and resources are being packaged as discoverable capabilities that can move across sessions, agents, and model providers."
thesis: "Agent skills are emerging as a portable capability layer, but their value depends on progressive disclosure, provenance, security review, and behavioral evaluation."
status: "published"
confidence: "medium"
first_seen: "2026-06-28"
last_verified: "2026-07-16"
supporting_signals: ["tg-2707","tg-2666","tg-2673","tg-2559"]
related_posts: ["claude-code-subagents-memory","agent-api-design"]
counter_evidence: ["Provider-specific tools, hooks, context models, and permission systems still limit true portability between agent runtimes.","A poorly designed or malicious skill can reduce task performance, consume context, or execute unsafe commands."]
revision_trigger: "Revise the thesis if cross-harness reuse remains rare or if evaluated skill packages fail to outperform simpler local instructions and typed tools."
topics: ["agent-skills","evals","context-engineering"]
source_urls: ["https://github.com/benchflow-ai/skillsbench","https://platform.claude.com/docs/en/agents-and-tools/agent-skills/best-practices","https://github.com/shadcn/improve","https://pydantic.dev/articles/pydantic-ai-capabilities"]
routes: {"html":"https://newruntime.com/patterns/skills-become-portable-capability-layer/","markdown":"https://newruntime.com/patterns/skills-become-portable-capability-layer.md","json":"https://newruntime.com/patterns/skills-become-portable-capability-layer.json"}
source_format: "markdown"
---

# Skills become a portable capability layer

## What is forming

A skill packages procedural knowledge with the resources needed to execute it.
That can include instructions, scripts, reference files, tool access, and
acceptance checks. The package can stay dormant until a task needs it, avoiding
the cost and interference of loading every capability into every prompt.

## Supporting evidence

- SkillsBench evaluates packages across tasks, models, and harnesses rather
  than judging instruction quality by inspection.
- Claude documents progressive disclosure and testing as core skill-design
  practices.
- Improve encodes repository audit and model-routing behavior as an inspectable
  skill folder.
- Pydantic AI capabilities combine instructions, tools, settings, and lifecycle
  hooks behind a deferred-loading interface.

## Operational consequence

Treat skills like dependencies. Pin a version, preserve provenance, inspect
scripts, document permissions, run task-level evals, and keep a no-skill
baseline. Portability is useful only when the package remains understandable
and safe outside the environment where it was authored.