The 'Skill Standardization' of AI Agents Has Begun: What Agent Plugins 1.0.0 Solves

One-line summary

The 'Agent Plugins' specification (v1.0.0, released in August 2026), co-developed by OpenAI, Amazon, Microsoft, Cursor, and Vercel, is a minimal standard for packaging AI agent skills in a way that allows them to be shared across multiple tools. Google has also joined.

The developer's dilemma

AI agent frameworks are proliferating. LangChain, Dify, Hermes Agent, Cursor... Each has its own 'skill' or 'tool' definition, creating the burden of rewriting the same functionality multiple times.

To give a concrete example, even if a developer implemented a skill (such as a 'convert text to image' function), if it was for LangChain, it had to be written in the LangChain format, and if it was for Dify, it had to be converted into the Dify Plugin format. In fact, there are cases where users, before Dify 1.14.0 support, created their own tools to 'convert Agent Skills to Dify Plugins'.

As this is a common problem for those of us developing AI agents, this is a highly welcome standardization.

What does the Agent Plugins specification define?

It is minimal. It only standardizes the file structure.

my-plugin/
├── plugin.json          ← スキルメタデータ（仕様を明記）
└── skills/
    └── skill-name/
        └── SKILL.md     ← 手順・リソース・指示

In plugin.json, the schema version and name are mandatory. The functional core (Agent Skills or MCP server settings) can be added optionally. In other words, it is a portable package format that only standardizes 'where to look for files'.

Why can this become a standard?

It has a lineage

MCP was released by Anthropic in November 2024 and donated to the Agentic AI Foundation under the Linux Foundation in December 2025. Agent Skills were also developed by Anthropic and moved to an open standard in December 2025 as well. In short, this specification is the 'distributable form' of the 'mechanism for agents to load external functions' that Anthropic pioneered, compiled jointly by five major companies.

Neutral governance

Vercel proposed it initially, and Amazon, Cursor, GitHub, Microsoft, and OpenAI jointly formulated the specification. Representatives from the five companies participate in the Technical Steering Committee, creating a structure that does not depend on the roadmap of any specific company.

The value for developers

• Making 'plugins' portable—the same plugin can be used across multiple clients such as ChatGPT, Copilot, and Dify.

• Reducing the need for re-implementing skills—it's not 'write once, done,' but rather the ability to use the same assets across different agents.

• Simplifying sharing within teams—once the format is unified, friction in distribution and management is reduced.

The Pitfall

This is a 'specification,' not executable software. It only functions once frameworks like LangChain or Dify implement loaders that support this specification. Furthermore, whether it will be adopted remains to be seen. If existing frameworks are satisfied with their own proprietary mechanisms, there may be little motivation to adopt a standard.

The First Step

Repository (https://agent-plugins.org/) Try reproducing the structure of 'hello-plugin' found in the README locally. By writing a plugin.json and skills/greet/SKILL.md, you can understand the 'minimum requirements' that form the core of the specification.

Question

Can this specification become the 'OpenAPI' of the AI agent skill ecosystem? —The key lies in the momentum of its adoption.

#AIAgent #Agent #Standardization #Plugins #Interoperability
#DeveloperTools #2026 #MumimaToolRoad