Published

AI in practice · Article

Agent protocols: what already has a specification, and what physical AI has not settled yet

Adriano Meirinho 3 min read

A published specification is a useful test. For software agents, that test is getting easier to apply. There are open documents that say how an agent reaches a tool, how one agent works with another, and how a personal agent presents itself to a company. For physical AI, the official pages I reopened describe a real stack, just not the same kind of document.

I also carry a recent reference point. I was at XPeng’s campus in Guangzhou. In its official results, the company presents itself as a global Physical AI company, with its own stack of Turing AI chips, world foundation models and integrated software and hardware, and a portfolio of intelligent electric vehicles, robotaxis and humanoid robots. That is the thesis the company puts on paper. It does not, by itself, give the field a shared agent protocol for that physical layer.

On the software side, the layers already have a home and a text. The Model Context Protocol, MCP, is an open protocol that its official specification describes as a standardized way to connect language-model applications to external data sources and tools. The same document says it takes some inspiration from the Language Server Protocol, which standardized how development tools add support for programming languages. The comparison is about ecosystem standards, and it does not introduce a new language.

Horizontally, Agent2Agent, A2A, is an open standard for communication between independent agent systems, including agents built with different frameworks or by different vendors. The specification published on the project’s site is 1.0.0. A2A’s own documentation sets the fit with MCP in plain terms: MCP connects an agent to tools, APIs and data; A2A connects it to other agents. The two are meant to be used together. IBM Research’s Agent Communication Protocol, ACP, has officially merged into A2A under the Linux Foundation, and is no longer a separate active line of development.

Those projects now sit under the same umbrella. The Linux Foundation created the Agentic AI Foundation, AAIF, and MCP is one of the founding projects. A2A, which Google Cloud had already transferred to a Linux Foundation project, is now a hosted project at AAIF. In the same set sits AGENTS.md, an open Markdown format for guiding coding agents, a README for the agent, stewarded by the same foundation. It belongs with project instructions, not with agent-to-agent conversation.

The map also includes Personal Agent Protocol, Poppy, in the 1.0 version published by Sierra. It describes how personal agents interact with companies through business agents, websites and APIs and, as Sierra reports, is largely built with existing standards such as MCP, OpenAPI and Universal Commerce Protocol (UCP). That is a different slice, the personal agent at a company’s door, and I cover the launch itself in this other piece: https://meirinho.com.br/en/blog/sierra-poppy-personal-agent-protocol-1-0/.

The page for each of these protocols, on this site, is at https://meirinho.com.br/en/ia/protocolos/.

When the subject turns to physical AI, the official sources support another layer, just as real, and of another kind. NVIDIA defines physical AI as what enables autonomous machines, such as robots in factories and robotaxis, to perceive, understand and perform complex actions in the physical world, combining models, real-world data, simulation and accelerated computing. In that work it publishes platforms and models: Cosmos, presented as an open platform for world models, and Isaac GR00T, an open stack with workflows and foundation models for general-purpose robot policies. For simulation, it describes SimReady as an OpenUSD-based framework, with OpenUSD as a foundational 3D format governed by the Alliance for OpenUSD, carrying the physical properties an asset needs in order to be simulated.

At the robotics base, Open Robotics still presents ROS as a software development kit for robot applications, and Open-RMF as an open system for interoperability between robot fleets and building infrastructure. There are also data efforts such as Open X-Embodiment, which gathers trajectories from dozens of embodiments in a standardized format for research on policies across robots. And there is ISO 21423, already published, on communications and interoperability for industrial mobile robots, with an explicit scope: an industrial workplace, not entertainment, consumer use, military or medical use, or use on public roads. None of those pages, with the same rigor as an agent specification, describe a shared protocol for agents in the physical world, or a programming language written only for AI.

The useful work, for now, is to keep the map honest: the agent protocols that already have a home and a spec, and the physical AI tools and standards that exist beside them. That is the picture I will keep updating.

Author

Adriano Meirinho · Co-founder & CAIO (Chief AI Officer) at Celcoin. About