OpenAI and four competitors have reached an agreement on a unified standard for AI agents.

      The AI competition was focused on models for two years, but this week it shifted to the infrastructure. On the brink of GPT-5's first anniversary, OpenAI and four competitors introduced Agent Plugins, an open standard that allows a single agent extension to function across different products.

      The key message is "build once, run anywhere." Currently, each agent tool has its own unique folder structure and setup requirements, but Agent Plugins replaces that complexity with a single package format. It combines two existing elements: Model Context Protocol servers, which link an agent to real-time tools and data, and Agent Skills, which are reusable instruction sets. Essentially, a plugin consists of a folder with a simple plugin.json file at the top level.

      In terms of leadership, despite OpenAI's promotion, this does not originate solely from them. The proposal was initiated by Vercel, with contributions from representatives of Amazon, Anysphere (the maker of Cursor), GitHub, Microsoft, OpenAI, and Vercel, who collectively formulated the 1.0 specification. The steering committee consists of five members: Amazon, Cursor, Microsoft, OpenAI, and Vercel. The project is openly licensed, and its proponents assert that no single company's roadmap dictates its trajectory.

      OpenAI still presented the launch as a personal achievement. With GPT-5's birthday on August 7, the company is celebrating the week by expanding its focus beyond individual models, as reported by 9to5Mac. The ChatGPT and Codex applications support the new format at launch, alongside Cursor, GitHub Copilot, Kiro, and VS Code. This development goes beyond OpenAI's own Codex plugins, which are limited to its tools.

      The standard is intentionally narrow. It specifies how a plugin should be packaged and located, but very little else. Aspects such as marketplaces, installation, permissions, sandboxing, and trust are left to individual clients. This approach makes the format simpler to adopt but also leaves significant challenges unmet. Determining whether a plugin is safe to execute remains the responsibility of each client, a concern highlighted by the earlier incident when fake Agent Skills bypassed security measures.

      Not everyone is supportive of this direction. Dax Raad, who develops the SST framework for developer tools, expressed strong opposition, labeling it "a thin standard" whose valuable elements will eventually be incorporated into client-specific extensions anyway.

      On the other hand, some have embraced the change. Developer advocate Angie Jones remarked, "We needed this," expressing a desire for a unified method to transfer her skills across various tools.

      The broader question concerns how standardizing the infrastructure will impact competition. A common format could enable a small developer to reach all major agents simultaneously, supporting the open-ecosystem argument, which is indeed plausible. Conversely, it may reinforce the few clients already possessing a user base, as a standard tends to favor those who already have established users.

      For now, the infrastructure has been agreed upon. The ongoing debate regarding the aspects that remain unaddressed, such as marketplaces and trust, has just begun.