The Great Agent Land Grab: Why the New Plugin Standard is a Battle for the Distribution Layer
AI-generated image · US National Wire
Amazon, Microsoft, OpenAI, and others are pushing a 'write-once, run-anywhere' spec for AI agents. While they call it interoperability, the real prize is controlling the marketplace where these skills are traded.
### The Illusion of Interoperability
On the surface, the announcement of Agent Plugins 1.0 looks like a victory for the open web. As The Register first reported, a coalition of the industry's heaviest hitters—including Amazon, Microsoft, OpenAI, Cursor, and Vercel—have joined forces to establish a universal format for AI agent skills and tools. The pitch is a "write-once, run-anywhere" container that allows developers to port their agentic capabilities across different platforms with ease.
By bundling instructions (via the Agent Skills standard created by Anthropic in 2025) and connectivity (via the Model Context Protocol, or MCP) into a single directory structure, the industry is attempting to tidy the "wild frontier" of agent development. Vercel posted the initial draft, and the Linux Foundation’s Agentic AI Foundation (AAIF) has embraced the project.
But as a columnist who follows the money, I see a different story. This isn't about the benevolence of sharing; it's about who owns the pipes.
### Following the Money: The Distribution Layer
To understand the stakes, you have to look at what the Agent Plugins 1.0 spec *doesn't* do. The specification handles the technical packaging—the JSON manifests, the schema, and the directory structures—but explicitly leaves the user interface, security boundaries, and the creation of marketplaces to the agent clients and platforms.
This is the critical distinction. By agreeing on a common language for the "skills," the titans of AI are ensuring that no matter who writes the tool, it must run on their infrastructure to reach the user. Currently, the platforms supporting this standard include ChatGPT, VS Code, Cursor, GitHub Copilot, Codex, and Kiro. By locking in a standard that these specific platforms support, the core development team—composed of engineers from Amazon, Cursor, Microsoft, OpenAI, and Vercel—is effectively defining the boundaries of the agent economy.
### The Technical Moat and Governance
The architecture of Agent Plugins 1.0 is designed for predictability, not necessarily for innovation. Mirroring Next.js practices, it uses a "file-system-as-configuration" approach. While developers can add experimental hooks into a "namespace" directory, the AAIF has stated that the 1.0 release scope is deliberately "narrow" to keep the format predictable.
Even competing efforts are being consolidated; Google recently folded its Agent2Agent (A2A) protocol into the AAIF. While the spec is managed under an open governance model and licensed under CC-BY-4.0, the core development team is composed of the very companies that stand to profit from the distribution layer.
### Opinion: The New App Store War
*Opinion: We have seen this movie before. In the early days of the smartphone, the battle wasn't just about hardware; it was about who controlled the App Store. By controlling the distribution layer, Apple and Google could tax the entire creator economy.*
*The push for Agent Plugins 1.0 is the AI equivalent. By standardizing the 'plugin,' these companies are ensuring that the 'agent' becomes a commodity. The 'write-once, run-anywhere' promise is the bait; the hook is the marketplace. The companies providing the UI and security boundaries are the ones who will ultimately decide which agents thrive and which ones are buried. This isn't a technical tidying of the frontier; it's a land grab for the AI economy's primary distribution channel.*