Agent Plugins Ends the Agent Format War
Agent Plugins arrived on August 6, and it paused a fight that had defined the year. For most of the past year, the battle over how AI agents plug into the wider world looked like a classic platform war, with every major lab pushing its own way to bundle skills, tools and connections. Then five of the most important names in software agreed on Agent Plugins, a single vendor-neutral standard for packaging AI agent extensions into one portable folder that any compatible client can read. Amazon, Cursor's maker Anysphere, Microsoft, OpenAI and Vercel signed on as core maintainers, with Google joining core maintenance and GitHub helping shape the 1.0 specification.
The timing was pointed. Agent Plugins landed one day before GPT-5 turned a year old, a milestone OpenAI has used to argue that the industry has moved past the era of standalone models and into the era of agents that actually do work. A shared packaging format is the plumbing that makes that shift real. When a developer builds an extension once, Agent Plugins lets it run inside ChatGPT, Codex, Cursor, GitHub Copilot, Kiro and Visual Studio Code without rewriting anything.
That breadth is the story. Rival companies rarely agree on anything, and here a group that competes fiercely on models and cloud revenue chose to standardise the layer beneath their products. The message to developers is that the connective tissue of the agentic web is now shared infrastructure rather than a competitive weapon.
What Agent Plugins Actually Standardises
Underneath the diplomacy, Agent Plugins is refreshingly simple. A plugin is just a directory. At its root sits a plugin.json manifest that describes the package, an optional skills folder, and an optional mcp.json file. Those two folders map to the two building blocks that have defined agent development this year. The first is Agent Skills, modular instruction sets that teach a model how to complete a defined task. The second is Model Context Protocol servers, the connectors that link an agent to APIs, databases and external systems.
The clever part is what Agent Plugins does not reinvent. Rather than replace Skills or MCP, it wraps both inside a common envelope so a single download carries everything an agent needs. Vercel, which initiated the proposal, framed the format as a way to end the copy and paste chaos that developers face when moving extensions between clients. Amazon described it in its open source blog as portable agent extensions that behave the same everywhere, and Google echoed the point in its developer blog, pledging to enable the format across its products.
Governance is the other deliberate choice. The project is openly licensed, its contribution process is public, and no single company's roadmap sets the direction. That structure is meant to reassure the thousands of developers who have grown wary of building on formats that a platform owner can change or deprecate at will. Agent Plugins is trying to be a standard in the boring, durable sense of the word.
The Anthropic Problem
There is one conspicuous absence. Anthropic, the company that created Model Context Protocol in the first place, appears on neither the list of core maintainers nor the roster of launch clients. The standard that now packages MCP for the rest of the industry was assembled without the protocol's author in the room.
The gap is more than symbolic, and it shows up in the technical details. Anthropic's own plugin layout hides its manifest at a path like .claude-plugin/plugin.json and keeps its MCP configuration in a hidden .mcp.json file. Agent Plugins puts both files in plain sight at the root of the folder. The two conventions are close enough to look compatible and different enough to force a choice. Developers who want their work to run everywhere will gravitate toward the format backed by six of the largest platforms, which quietly pressures Anthropic to conform to a design it did not write.
This is how open standards can be turned into strategic instruments. MCP was Anthropic's gift to the ecosystem, and it succeeded so completely that rivals adopted it. Now a coalition led by OpenAI has absorbed that success into a broader container and set the terms for how it spreads. Anthropic can join later, but it would be endorsing a standard shaped by its fiercest competitor. Staying out preserves independence at the cost of interoperability. Neither path is comfortable.
Why the Coalition Formed Now
Standards emerge when fragmentation starts costing everyone money. Over the past year the number of agent clients exploded, and each one shipped a slightly different way to install extensions. A tool built for Cursor did not drop cleanly into Copilot, and a skill written for ChatGPT needed rework to run in VS Code. That friction slowed the exact ecosystem every platform was racing to grow. The companies backing Agent Plugins each concluded that a bigger, shared market is worth more than a walled garden they control alone.
There is also a defensive logic at work. A single portable format lowers the barrier for developers to support many clients at once, which favours the incumbents who already have distribution. If any extension can run anywhere, the competition shifts back to the models and the products, where OpenAI, Google, Microsoft and Amazon believe they can win. Standardising the plugin layer neutralises it as a point of lock-in for smaller challengers while enlarging the pie for the giants.
The GPT-5 anniversary framing reinforces the strategy. OpenAI spent the year arguing that agents, not chatbots, are the real product, and an agent is only as useful as the tools it can reach. By seeding a common format now, the coalition is trying to make the coming wave of agent development flow through infrastructure it collectively governs.
The Stakes for the Agentic Web
If Agent Plugins takes hold, the practical effects arrive quietly. Developers publish once and reach every major client. Enterprises audit a single manifest format rather than a dozen. A marketplace of portable plugins becomes plausible, much as app stores emerged once mobile settled on shared conventions. That is a meaningful acceleration for a category that has been held back by plumbing problems rather than a shortage of ideas.
The risk is that a format governed by the largest platforms entrenches their advantage under the banner of openness. Open licensing and public governance are real safeguards, yet the companies with the most developers and the deepest clouds still shape the defaults everyone else inherits. Watching who actually steers the specification, and whether independent voices retain influence, will matter more than the launch day list of logos.
For now, the headline is unusual cooperation in an industry defined by rivalry, and a pointed absence at its center. Agent Plugins may become the MCP-compatible standard that finally unifies agent extensions, or it may become the moment the field split into a large shared camp and a smaller independent one led by Anthropic. Either way, the plumbing of the agentic web just became a strategic battleground. Blumefield will keep tracking how the standard spreads, and you can follow the full arc of the agent wars at Blumefield.