The Internet of Agents

2025-05-06 post Matt Ferguson

Comparing Anthropic’s MCP, Google’s A2A, Cisoc’s AGNTCY, MIT NANDA, and Kafka

The digital landscape is pretty much on the verge of another massive transformation, moving beyond connected devices (IoT) and interconnected services (the web) towards an Internet of Agents (IoA). Imagine a world where potentially billions of specialized AI agents, built by different organizations, can discover, communicate, collaborate, and even transact autonomously.

The concept really hit home for me back in April when I attended the Imagination in Action event at MIT and was introduced to the NANDA initiative (NANDA Website). It was genuinely eye-opening. Around the same time Cisco announced AGNTCY (Outshift Blog - Announcement). Being a Cisco alum, and fortunate enough to live in the Boston area and have attended MIT courses, both these announcements were enough to solidify that we’re experiencing the beginning of something big and that agent collaboration would actually be standardized and built.

It’s clear the potential here is huge. Market forecasts suggest the AI Agents market could explode from around $5-8 billion in 2024/2025 to well over $200 billion by the early 2030s (Source: Precedence Research, Verified Market Research). That’s some serious growth! But how do we actually get there? Realizing this vision requires foundational protocols and infrastructure – the standards that let these diverse agents talk reliably and securely. Several contenders are emerging, each taking a different swing at the problem. So, let’s dive in and compare some key players: Anthropic’s MCP, Google’s A2A, Cisco’s AGNTCY, MIT’s NANDA, and see where established tech like Apache Kafka fits in.

Anthropic’s MCP: Getting Agents and Tools Talking

First up, announced originally in late 2024 and beefed up since (Anthropic News, May 2025), is Anthropic’s Model Context Protocol (MCP). Its goal? Standardizing how AI models (mostly LLMs) chat with external tools and data sources.

What it Does: Think of MCP like a “USB port for AI.” It’s an open, client-server setup (often using JSON-RPC) that gives AI apps a consistent way to connect to external capabilities. No more writing custom code for every little integration! Developers can use MCP to feed models structured context, manage how they use tools (like APIs, databases, or even local files), and handle the results.
Why Does This Matter? Well, MCP is crucial for building those fancy, modular AI apps (sometimes called “Compound AI”). It lets agents grab the specific data and tools they need on the fly, making development faster and the apps way more flexible. It’s really the bedrock for letting agents interact effectively with the world around them.
Adoption Check: MCP saw pretty rapid uptake in the developer community. Big names like AWS, GitHub, Microsoft, Google, OpenAI, Sourcegraph, and others are jumping on board, implementing MCP servers or baking it into their tools (ZDNet, Microsoft DevBlogs). While getting it truly enterprise-ready is still a work in progress for many, its wide adoption shows just how much everyone needed a standard way for agents to use tools.

Google’s A2A: Letting Agents Collaborate

Next, let’s look at Google’s Agent2Agent (A2A) Protocol, announced just recently in April 2025 (Google Developers Blog). A2A tackles a different, though related, piece of the puzzle: getting independent AI agents to actually talk and work together.

What Problem is it Solving? Basically, today’s AI agents often live in their own little walled gardens, unable to easily cooperate. A2A wants to tear down those walls by providing an open, standard “language.” Agents can use it to broadcast their skills (via an “Agent Card”), figure out tasks together, share info, and manage complex workflows that might span different companies or systems. The focus here is squarely on agent-to-agent chats.
So, Why is A2A Important? It’s vital if we ever want to see that true Internet of Agents vision come to life. Without something like A2A, trying to get Agent X from Vendor A to work with Agent Y from Vendor B means building complicated custom bridges or just giving up and sticking with one vendor. A2A promises to slash that integration headache (W&B Report), letting us mix and match the best agents for the job.
What’s the Potential Here? Although brand new, A2A has a ton of potential. Google didn’t build it in a vacuum; they worked with over 50 industry partners. It clearly addresses real needs. Plus, it uses familiar web standards (JSON-RPC over HTTP, SSE) and has solid features for discovering capabilities and handling tasks that might take a while. Keep an eye on this one.

Cisco’s AGNTCY: Aiming for the Whole Ecosystem?

Now, onto Cisco’s AGNTCY. Announced in March 2025 (Outshift Blog - Announcement, Outshift Blog - GitHub Release), this one is interesting. Cisco, of course, was instrumental in building the first internet. As I mentioned, being a Cisco alum, this announcement definitely had me leaning in.

How Does it Compare to A2A? Well, AGNTCY isn’t just a protocol. They’re billing it as an open-source collective trying to build the whole infrastructure for the IoA (AGNTCY Overview). Its goals definitely overlap with A2A, especially around agent discovery (they have an Agent Directory and schema) and communication (an Agent Connect Protocol). But AGNTCY seems broader, aiming to cover standards, directories, protocols, gateways, workflows, and even evaluation. Think of it this way: if A2A is defining how two agents talk, AGNTCY wants to define the whole playground they talk in.
So, What About Cisco’s Role? They bring undeniable networking and infrastructure know-how. And kudos to them for pushing an open, collaborative model from the start (partnering with LangChain, Galileo, etc.). But let’s be honest, the AI agent world is driven by AI innovation, not just network pipes. While Cisco’s history gives them street cred, they’re up against serious AI heavyweights. AGNTCY is still fresh, so its market momentum is TBD. Success will mean getting the broader AI community to buy into their vision and standards.

Kafka & Event-Driven Architecture: A Different Approach?

While protocols like A2A define direct agent chats, things get tricky when you try to scale. Imagine thousands, maybe millions, of agents trying to talk directly – sounds messy, right? This is where Event-Driven Architecture (EDA), often powered by platforms like Apache Kafka, enters the chat, with companies like Confluent making a strong case for it.

How Does This Fit In? Kafka itself isn’t an agent protocol per se. It’s a beast at moving huge amounts of event data around asynchronously. But Confluent argues that the direct, point-to-point model (like A2A over basic HTTP) creates bottlenecks, makes things brittle, and hides what’s actually going on when you scale up (Confluent Blog: Why A2A Needs Kafka). Their proposed solution? Use Kafka as the backbone. Agents toss events (requests, status updates, results) onto Kafka topics, and other agents just subscribe to the topics they care about.
So, Is It a Good Fit?
- Pros: This EDA approach means agents don’t need to know about each other directly – they’re decoupled! That boosts resilience. Plus, multiple systems can listen to the same event (think logging, monitoring, other agents). It potentially makes orchestrating complex flows easier and more real-time (Confluent Blog: Event-Driven Agents). Kafka’s built for high throughput.
- Cons: But, let’s be real, implementing and managing Apache Kafka can be a headache and potentially expensive, although managed services like Confluent Cloud are designed to ease that pain (Confluent Blog: Multi-Agent AI with Confluent). It definitely requires thinking differently (event-driven!) and adds another piece of infrastructure to manage.

MIT’s NANDA: The Decentralized Vision

Remember NANDA (Networked Agents And Decentralized AI), the initiative I saw at MIT’s Imagination in Action event (NANDA Website)? Let’s see where it fits.

Where Does it Fit? NANDA is aiming high: building the infrastructure for a “true Internet of AI Agents” with a big emphasis on decentralization. What’s really interesting is that it explicitly positions itself as building on Anthropic’s MCP. So, if MCP handles agents talking to tools, NANDA wants to add the missing pieces for large-scale, distributed smarts: reliable ways for agents to find each other, prove who they are, securely interact, keep track of interactions (“Trace”), maybe even create knowledge markets. Its goals clearly overlap with A2A and AGNTCY’s infrastructure plans, but with that unique decentralized flavor and a direct link to MCP.
What About Adoption? As an MIT project with industry partners (TCS, Mitsubishi Electric, HCLTech, etc.), it feels like an open project looking for traction. The focus on open protocols and decentralization definitely points that way. Whether it takes off will depend on building a strong community and showing how it works with, or improves upon, the other standards popping up.

So, Where Does This Leave Us?

Okay, that was a whirlwind tour! The Internet of Agents isn’t just hype anymore; it’s happening, and the race to define its foundations is on.

MCP is nailing the agent-to-tool connection.
A2A is giving agents a direct line to collaborate.
AGNTCY is thinking bigger, aiming for the whole open infrastructure.
NANDA offers a cool, decentralized vision built on MCP.
Kafka/EDA provides a powerful, if different, way to handle the underlying communication chaos at scale.

It’s super important to remember this field is moving incredibly fast. You’ve got countless other platforms and players integrating agent features, as lists like the one on Slashdot show (Slashdot MCP Alternatives).

For me personally, the journey from seeing NANDA unveiled at MIT (NANDA Website) to tracking Cisco’s AGNTCY launch (Outshift Blog - Announcement) as an alum highlights just how dynamic and crucial this early phase of standardization is. It’s going to be really fascinating to watch which protocols gain traction, how they might work together, and how they ultimately shape the true Internet of Agents. Stay tuned!