4 Rivals Cooperate | How Agent Plugins Is Changing the Future of AI Standardization
機械翻訳 / Machine-translated
A
AI Friends
5日前・13分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
In August 2026, a major development shook the AI industry. Four rival companies — Google, OpenAI, Microsoft, and Amazon — agreed to adopt the same AI agent framework. This article explains the background and implications in plain terms.
Agent Plugins is a system that creates a "common container" for AI agents. An AI agent refers to an AI that autonomously carries out tasks based on instructions from humans.
For example, there are many different AI tools out there — ChatGPT, Gemini, Copilot, and more. Until now, developers had to build separate plugins (add-on features) for each individual tool. In other words, they had to recreate the same functionality over and over again.
Agent Plugins solves this problem. It is a "common standard" that allows a plugin built once to work across any AI tool — significantly reducing the burden on developers.
Technically, it is structured around a file called plugin.json. This file describes the plugin's name and capabilities; a folder called skills contains the AI's behavioral patterns; and a file called mcp.json holds the configuration for connecting to external services.
Agent Plugins 1.0.0 was jointly announced on August 6, 2026, by five companies: OpenAI, Microsoft, Amazon, Cursor, and Vercel. Then, on the following day, August 7, it was revealed that Google had joined as well.
Kevin Hou, a Senior Staff Engineer at Google's DeepMind (its AI research division), came on board as a core maintainer (a central developer on the project).
This is highly unusual. It means companies that are normally in competition agreed to build the same standard together — a sign of just how important each company recognizes AI agent standardization to be.
In fact, Google has already updated two of its developer tools, Agents CLI and Data Agent Kit, to support the Agent Plugins format. Skills for connecting to Google services such as BigQuery (a data analytics service), Spanner, and Cloud SQL can now be used with other AI tools as well.
The biggest advantage of Agent Plugins is that it dramatically cuts development time. Previously, developers had to rebuild the same functionality separately for ChatGPT, Copilot, Claude, and so on.
With Agent Plugins, a plugin built once can be reused across multiple AI tools. This means developers can focus on bringing new ideas to life without being bogged down by compatibility issues.
There are benefits for end users too. For instance, a useful plugin you've been using in ChatGPT could potentially be used in Gemini as well. The barrier to switching between AI tools is lowered.
That said, challenges remain at this stage. Agent Plugins 1.0.0 standardizes only the basic format; details around permissions management, security features, and distribution methods are left to each individual company. It is hoped that future version updates will address these issues.
To understand Agent Plugins, it helps to know about two underlying mechanisms: MCP and Agent Skills.
MCP (Model Context Protocol) is a "connectivity standard" that allows AI to access external services and data. For example, it is used when sending an email or searching a database. Think of MCP as the "hands and feet" of an AI.
Agent Skills, on the other hand, is a system for teaching an AI "how to do things." For example, knowledge such as "legal documents should be written in this format" or "code reviews should follow this procedure" is described in Markdown (a notation system for writing text). Think of Agent Skills as the "brain" of an AI.
In short, MCP is the tool for connecting to the outside world, and Agent Skills is the textbook for learning how to do the work. Agent Plugins serves as the "container" that bundles these two together into a single package that any AI tool can use.
These developments are also moving forward in Japan. In April 2026, a forum chaired by Professor Hideaki Takeda of the National Institute of Informatics compiled a set of recommendations on how to utilize AI agents in the manufacturing sector.
The Japanese government also published "Guidelines on the Procurement and Utilization of Generative AI for the Evolution and Innovation of Public Administration" in June 2026. These guidelines define an AI agent as "an AI system that perceives its environment and acts autonomously to achieve a specific goal."
Furthermore, the U.S. National Institute of Standards and Technology (NIST) announced the "AI Agent Standards Initiative" in February 2026 — an international standardization effort focused on interoperability, safety, and reliability for AI agents.
In other words, standardization efforts like Agent Plugins reflect a global trend. Japanese companies and government agencies are also preparing to keep pace with this movement.
Agent Plugins 1.0.0 is a significant step forward, but it is not perfect. Currently, only the basic "shape of the container" is standardized, while the details of the contents still vary by company.
There are several issues that will need to be resolved going forward — such as plugin distribution methods, permissions management, security measures, and sandboxing (a mechanism to prevent plugins from performing dangerous operations on their own).
Mechanisms for verifying the trustworthiness of plugin creators, as well as unification of user interfaces (operation screens), will also be needed. These are outside the scope of Agent Plugins 1.0.0 and are scheduled to be addressed in future versions.
Even so, the fact that four competing companies cooperated to build a common standard is a landmark event for the AI agent industry. As more AI tools adopt Agent Plugins going forward, the burden on developers will decrease, and users can expect to see more useful plugins available to them.
Google has stated that it intends to "expand Agent Plugins support to even more Google products." Other companies are showing similar moves, and the AI agent market is expected to undergo major changes from the second half of 2026 through 2027.
This article is a cross-post from AI Friends.
