Two labs shipped agent standards on the same day
On 7 August OpenAI published Agent Plugins as an open standard and Anthropic's Claude Agent SDK opened an in-process MCP server. Coverage is calling it protocol year one, which is generous, but the direction is real.
On 7 August two things landed together: OpenAI published Agent Plugins as an open standard, and the Python build of Anthropic's Claude Agent SDK opened an in-process MCP server. Two competitors shipping interoperability work on the same day is not coordination, but it is not coincidence either.
The emerging division of labour is clearer than the branding suggests. MCP is the calling protocol — how an agent reaches a tool. Agent Plugins is a packaging protocol — how a capability is bundled and distributed. A2A is the collaboration protocol — how agents talk to each other. Three layers, three problems, and until recently every vendor was solving all three privately.
The in-process MCP server is the more technically consequential item. Running the server inside the agent's own process removes the transport hop that made local tool calls slow and awkward to secure. It is the difference between a tool call being a network event and being a function call.
Standards land when the incumbents decide the interface is not where they compete. That is the signal in this pair — both labs have concluded the money is in the model and the product, not in owning how a tool gets invoked.
What remains unsettled is authorisation. Every protocol in this stack describes how agents connect, and none of them settles what an agent is permitted to do once connected. That gap is exactly where this summer's agentic misalignment work has been pointing.
Boao Intelligent — AI Agents Enter 'Standardization Protocol Year One': OpenAI Agent Plugins + Anthropic Claude Agent SDK → · Digital Applied — AI Agent Protocol Ecosystem Map 2026: MCP, A2A, ACP, UCP →