What the First AI Plugin Standard Reveals About AI's Future

Six major companies released a shared specification for AI agent plugins on August 6, 2026. OpenAI, AWS, Microsoft, GitHub, Anysphere (Cursor), and Vercel jointly published Agent Plugins 1.0.0, an open and vendor-neutral standard. Google quickly joined as a core maintainer, while Anthropic created the underlying technologies the standard packages together.

The standard is compact and allows developers to package AI agent skills and tool connections into a single, portable bundle. A compliant plugin is a directory containing a manifest, skills folder with reusable instructions, and an MCP server configuration. The promise is simple: build a plugin once, and it will work across various AI clients like ChatGPT, Cursor, Copilot, and VS Code.

The standard solves the packaging problem by bundling two established technologies - Agent Skills and MCP servers. However, it deliberately omits security, trust verification, publisher identity, code signing, sandboxing, permission models, and runtime behavior, leaving these aspects to individual platforms. The omission of security measures is significant, as AI agents are already facing trust issues in production environments.

The standard acknowledges the gap between the impressive capabilities of AI models and their reliability in real-world scenarios. Despite the impressive performance of AI agents, there are numerous incidents where they have failed silently or performed inappropriate actions due to a lack of proper constraints and verification mechanisms.

The emergence of the standard was driven by the widening gap between AI agents' capabilities in demos and their actual performance in production. While the standard does not solve the trust problem, it acknowledges its importance and signifies the beginning of the infrastructure phase for agent-based AI.