The Internet of Agents Is Being Built—But It's Not Yours (Yet)

Let's not get distracted by the hype.

Yes, AI agents are real. Yes, they're powerful. But if you look closely at who's building them—and how—you'll see a familiar pattern: centralized power, platform lock-in, and vague gestures toward openness that crumble the moment you read the fine print.

Everyone's racing to define the so-called "Internet of Agents." But right now, it's shaping up to be an internet for Big Tech, not by the people. We've been here before. And this time, we should know better.

Big Tech's Agent Playbooks

Let's break down what the big players are actually doing.

Google: Mariner, A2A

Google is all-in on agentic infrastructure—but only on their terms.

Project Mariner is a Gemini-powered agent that can browse the web, fill forms, shop, book travel, and more. It lives inside Chrome and works in tandem with Google Search and Apps. Initially running in your browser tab, Google has now updated it to run on virtual machines in the cloud, allowing it to handle up to 10 tasks simultaneously while you work on other things.

They've also introduced the A2A (Agent-to-Agent) protocol—an open spec that enables agents to discover, communicate, and collaborate via JSON-RPC and Server-Sent Events. Google launched this with support from over 50 technology partners including Atlassian, Box, Cohere, Salesforce, and major consulting firms like Accenture and McKinsey.

It's technically open. But its implementation so far? Mostly within the Google ecosystem. There's no public, cloud-agnostic A2A network in the wild. Open spec ≠ decentralized infrastructure.

Microsoft: Copilot Studio and Multi-Agent Orchestration

Microsoft is positioning agents as core to enterprise infrastructure.

Copilot Studio now supports multi-agent orchestration in public preview. Agents built with Azure, GitHub Copilot, and Microsoft 365 can work together to complete business tasks across apps. It's polished and powerful—for companies already living inside the Microsoft stack.

Microsoft gives over 230,000 organizations access to Copilot Studio, including 90% of the Fortune 500. They're also pushing the Microsoft 365 Agents SDK, supporting Agent2Agent (A2A) protocol and Model Context Protocol (MCP) integration.

But everything routes through Azure. Every interaction relies on Microsoft Entra. This is Microsoft's "Internet of Agents"—it's just not the Internet of Agents.

Salesforce: Agentforce

Salesforce's Agentforce is a low-code platform for building enterprise AI agents across sales, service, and marketing. These agents run on top of the Atlas Reasoning Engine, access real-time customer data, and extend via the AgentExchange marketplace.

It's all optimized for the Salesforce CRM layer. And like everything else Salesforce builds, it's designed to work best if you already buy into their vertical stack.

The Pattern Emerges

Here's what we're seeing across the board:

• Agents that work beautifully—within specific ecosystems
• "Open" protocols that somehow require proprietary infrastructure
• Enterprise-grade features that lock you into cloud providers
• APIs that are technically accessible but practically vendor-dependent

The real web—the one that changed everything—wasn't built by companies. It was built on protocols. On weird RFCs and open-source software and people hacking on servers in their dorm rooms.

What's Missing: A Real Agentic Commons

If we want the Internet of Agents to be anything more than another SaaSified sandbox, we need infrastructure that isn't owned.

Enter: Agent Network Protocol (ANP)

ANP is the closest thing we've seen to a truly decentralized protocol for agents. It defines a trustless, cryptographic framework for agent discovery, communication, and verification—without assuming a central provider.

The project aims to become "the HTTP of the agent internet era." Its three-layer architecture includes:

• Identity and Secure Communication Layer: Based on W3C DID specification, enabling decentralized authentication
• Meta-Protocol Layer: For negotiating communication protocols between agents
• Application Protocol Layer: Based on semantic web specifications for describing agent capabilities

It's not a wrapper for OpenAI APIs. It's not a backend for a corporate dashboard. It's a real protocol—like DNS or HTTP, but for autonomous agents. Built in public, by a distributed community, and free from cloud dependencies.

This is what we should be funding. This is what we should be building on.

My Take

We don't need another cloud. We need coordination.

The agentic future can't be five mega-clouds yelling at each other through their own protocols. That's not the internet. That's a fragmented empire.

The real breakthroughs will come from community-built protocols that:

• Don't rely on centralized identity systems
• Don't require proprietary clouds
• Allow agents to discover, trust, and collaborate without permission

So if you're a founder, a developer, or a builder dreaming about agents that work anywhere, talk to anyone, and act on your behalf—not a corporation's—then start here: contribute to ANP. Build on it. Extend it. Remix it.

The internet of agents should be as open and permissionless as the web that came before it.