According to TheRegister.com, the AI industry is drowning in an “alphabet soup” of competing protocols designed to standardize how AI agents communicate. The open-source Model Context Protocol (MCP), originally developed by Anthropic in late 2024, has become a de facto standard for agent-to-tool communication, adopted by giants like OpenAI and Google. For agent-to-agent talk, Google’s A2A protocol, contributed to the Linux Foundation in 2024 and later merged with IBM’s ACP, is gaining similar traction. Meanwhile, newer protocols like the Universal Tool Calling Protocol (UTCP), the Agent Network Protocol (ANP), and Google’s preview-stage Agent to UI (A2UI) protocol are all vying for attention in specific niches. The situation has become so fragmented that the Linux Foundation formed the Agentic AI Foundation (AAIF) last month to provide vendor-neutral oversight.
The USB-C wannabe and its rivals
Here’s the thing about standards: everyone wants to make their own. MCP is being billed as the “USB-C of agentic systems,” which is a fancy way of saying it wants to be the one universal plug everything uses. And, to be fair, it’s got the momentum. When Anthropic’s rivals like OpenAI start using your protocol, you’re doing something right. But is it the *best* technical solution? That’s far less clear.
The criticism, as highlighted by UTCP proponents, is pretty straightforward. MCP adds a layer—a client-server architecture where the tool needs an MCP wrapper. UTCP argues, “Hey, if a tool already has an API, why make the AI talk to a *second* API just to talk to the first one?” It’s a simpler, more direct approach that could be more secure and performant. But in tech, the best tech doesn’t always win. Momentum, backing, and ecosystem matter more. Right now, MCP has that in spades for tool-calling, even with its documented security concerns.
When agents need to gossip
But tools are only half the story. The real promise of agentic AI is multiple specialized agents working together like a team. You don’t want one monolithic AI trying to do your taxes, book travel, and write code. You want a tax agent, a travel agent, and a coding agent that can hand off tasks. That’s where protocols like A2A come in. It’s basically the chat room for your AI workforce.
What’s interesting is how Google is playing this. They’re not just pushing one protocol; they’re building a whole suite. A2A for agent chats, MCP adoption for tools, and then A2UI for generating dynamic user interfaces. They’re trying to own the plumbing of the agentic world. And by donating A2A to the Linux Foundation, they’re making a smart play for legitimacy and broad adoption, pulling in Microsoft and AWS. It’s a classic “embrace, extend, but maybe don’t extinguish” open-source strategy.
The wild west of specialized tasks
Now, just when you thought you had a handle on the core protocols, the domain-specific ones pop up. Of course we need a protocol for e-commerce! Google’s Universal Commerce Protocol (UCP) and its related Agent Payments Protocol (AP2) are direct responses to the very real, very hilarious fear of AI agents going on shopping sprees. Remember the story about the AI spending $31 on a dozen eggs? Yeah, no one wants that.
This is where it starts to feel a bit ridiculous. Do we need a new protocol for *every* possible thing an agent might do? Probably not, but that won’t stop companies from proposing them. It’s early days, and everyone is staking a claim on what they think will be critical infrastructure. The risk is massive fragmentation, where building a useful agent requires supporting a dozen different, incompatible specs.
Can anyone make this make sense?
So, what’s the endgame? Chaos, probably, for a while. But the formation of the Linux Foundation’s AAIF is the first real sign of the industry trying to get a grip. Merging A2A and IBM’s ACP was a good first step. The hope is that the Foundation can do what it often does: provide a neutral ground where competing interests can argue it out and arrive at something that doesn’t suck for everyone.
Basically, we’re in the “VHS vs. Betamax” or “HD-DVD vs. Blu-ray” phase of agentic AI protocols. No developer or company wants to bet big on the standard that fizzles out. The protocols that win won’t necessarily be the most elegant—they’ll be the ones with the biggest backers and the least friction to adopt. For now, if you’re building, you’re likely looking at supporting MCP and A2A as your safe bets, while keeping a wary eye on the long tail of niche contenders like ANP or NLIP. It’s a mess, but it’s the messy, noisy process of building something new. And honestly, it’s more interesting than having one company dictate everything from the start.
