Five Tech Giants Agree on a Shared AI Agent Plugin Standard

Adding a new capability to an AI agent has been a per-platform chore. Write the integration once for one vendor's agent, then repackage it for the next. On August 7, five companies decided to stop doing that separately.

Amazon, Microsoft, OpenAI, Vercel and Cursor announced Agent Plugins, an open format for packaging what an agent can do and what it can connect to. The whole thing is deliberately unglamorous: a folder with a manifest file named plugin.json inside it.

What version 1.0.0 actually covers

The spec recognises two components. Agent Skills are reusable instructions and workflows, the "here is how this task gets done" layer. MCP servers (Model Context Protocol) are the connections that wire an agent to real tools, databases and files. The specification is published openly on GitHub, with a separate site documenting it.

What the standard leaves out matters just as much. Agent Plugins handles packaging and discovery only: how a plugin is bundled, and how an agent finds it. Marketplaces, permissions and the runtime environment are all outside its scope. How far a plugin can reach into your company's data is still governed entirely by boundaries you set yourself.

The name missing from the list

Here is the odd part. MCP is Anthropic's protocol, and Agent Skills started as Anthropic's idea. Yet Anthropic is not among the signatories, and continues to run its own plugin system inside Cowork.

So this is not quite the industry consensus the announcement implies. Five major platforms converging on one format carries real weight, but the company that authored both underlying concepts sitting it out means the "one standard" story is, for now, incomplete.

What it changes for engineering teams

In the enterprise AI work we do, a surprising share of the effort never touches the model. It goes into plumbing: pulling stock levels out of an ERP, reading a customer balance from accounting software, writing a note back into a CRM. Repackaging that plumbing for every agent platform is an invisible but very real line item.

A shared manifest chips away at it. An integration written once becomes portable across any tool that adopts the format. If you are designing an agent integration now, it is worth structuring skill definitions and tool connections along that split from the start, even before your platform supports the spec. Porting later gets much cheaper.

Keep the caveat in view, though. Because permissions sit outside the standard, deciding what a plugin may read and what it may do without human approval remains your job. Writing those limits down before the first integration ships costs far less than retrofitting them after something goes wrong.

Sources: The Decoder, Agent Plugins Spec (GitHub)