Loading...

Blog > Web of Agents

Web of Agents

Web of Agents

MCP, A2A, UCP, AP2, ANP... a protocol explosion is underway. But protocols alone are not enough. What's missing is a Google for agents. The search engine for the Web of Agents has yet to emerge.

Published: February 22, 2026

Last updated: February 22, 2026

Abstract

The explosion of agent protocols such as Model Context Protocol (MCP), Agent-to-Agent Protocol (A2A), Universal Commerce Protocol (UCP), Agent Payments Protocol (AP2), Agent Network Protocol (ANP) and others is building the plumbing for the Web of Agents. But plumbing is not enough. What these protocols are preparing for is discovery: the ability for agents to find, evaluate and connect with other agents and services. A search engine for agents, a Google for the Web of Agents, has yet to emerge.

Introduction

We're in the early days of the Web of Agents. Agents will consume and produce on behalf of users. For that to work at scale, they need standards to talk to each other, to data, to commerce, to payments. Hence the protocol explosion. Anthropic, Google and the Linux Foundation are working on these protocols and preparing the era of the Web of Agents.

The Protocol Landscape

Overview of the main protocols and what each solves.

  • ANP is the decentralized protocol for agent discovery, authentication and collaboration over the open internet.
  • MCP is the universal adapter for agent-to-data.
  • A2A is the foundation for multi-agent systems.
  • UCP is the universal commerce protocol for agent interactions in e-commerce.
  • AP2 is the agent payments protocol for secure payment communications between shopping agents, merchants and credential providers.

Agent Network Protocol (ANP)

ANP is a decentralized, AI-native protocol designed to enable agents to discover, authenticate and collaborate with one another over the open internet. It aims to become "the HTTP of the Agentic Web era." ANP addresses the discovery problem directly: interconnection across boundaries, native AI interfaces (rather than mimicking human browsing) and efficient self-organizing collaboration.

Model Context Protocol (MCP)

Since we are talking about the Web of Agents, we need to have a way to connect the agents to the data. MCP is the standard for this, probably the most famous and most adopted protocol for this purpose. It is a simple protocol that allows an agent with context to connect to the data. Purpose of this protocol is to allow the agents to connect to the data and tools and APIs. Not mainly developed for the agentic communication but for the agentic data connection.

Agent-to-Agent Protocol (A2A)

The foundation for multi-agent systems. A2A enables agents to communicate, delegate and collaborate with each other. Developed by Google and IBM, it is now part of the Linux Foundation. Without A2A, agents would have no standard way to talk to each other everything else builds on this.

Agent Payments Protocol (AP2)

Extends A2A specifically for payments. AP2 handles secure payment communications between shopping agents, merchants and credential providers. When an agent needs to pay another agent or a merchant, AP2 is the layer that makes it safe and standardized.

Universal Commerce Protocol (UCP)

A newer protocol from Google for agentic commerce. UCP handles product discovery, checkout and transaction flow when agents shop on behalf of users. It sits alongside AP2, UCP for the commerce flow, AP2 for the payment itself.

Others

The focus here is on the protocol stack that enables agentic commerce and discovery. There are other protocols (ACP and others under the Linux Foundation's Agentic AI Foundation) that are not covered here.

The Missing Piece: Discovery

Protocols give agents the ability to talk. But before they can talk, they need to find each other. Today we have Google for humans—we search, we discover, we navigate, we buy, we sell. Agents need the same. Why would they have to use browsers to find each other? Or why would they have to use other tools? ANP is already tackling discovery from a decentralized angle. What has yet to emerge is a search engine for the Web of Agents, the index, the crawl, the ranking, whether centralized (like Google for humans) or built on top of protocols like ANP.

Google's Role

Google, as the most dominant search engine for humans, is the most likely to build a centralized search engine for the Web of Agents. From the protocols, we can see that it is coming they are taking a strategic approach: first A2A, then AP2, then UCP and maybe next a standard for letting your domain expose and discover agents. Meanwhile, ANP is building discovery from a decentralized, open-network perspective. Both approaches may coexist.

The Search Engine for Agents

The Search Engine for Agents would look like a search engine for humans, it would index the agents what they are doing, what they are capable of. This is not about finding an API for an agent but finding another agent that—for example—sells goods, flights, hotels, etc. This way, agents can find each other and collaborate with each other.

Current possible steps to buy a flight, hotel, etc. Let's say I want to buy a flight from Istanbul to Milan. When you give this to your personal assistant, it can do the following steps:

  1. Search for the flight, hotel, etc. Possibly on Google, Tavily, Brave Search etc.
  2. Find the websites of the merchants that sell the flight/hotel.
  3. Compare the prices and the services of the merchants.
  4. Use browser-use, API or tooling capabilities of the agents to buy the flight/hotel.
  5. Pay the merchant for the flight/hotel.
  6. Get the flight/hotel.

With the Search Engine for Agents, agents can find each other and collaborate with each other. A possible path would be:

  1. Search for the flight, hotel, etc. Possibly on Google Agent Search Engine.
  2. Find the agents that sell the flight, hotel, etc.
  3. Compare the prices and the services of the agents.
  4. We will not use browser-use, API or tooling capabilities of the agents to buy the flight/hotel.
  5. We use UCP and AP2 to buy the flight/hotel, because our agents communicate with each other using these protocols.
  6. We get the flight/hotel.

So, in the new scenario, the agent found an agent that sells the flight/hotel. As you can see the agents can communicate with each other using the protocols and they do not need to solve our problems (like booking the flight/hotel) by using our methods but using their own protocols.

Conclusion

The Web of Agents is being built from the bottom up—protocol by protocol. Discovery is the layer that will make it navigable. It has to emerge. The question is when and by whom and perhaps Google for a centralized index or perhaps built on protocols like ANP for a decentralized one.

Several solutions are already building/built the discovery layer for the Web of Agents:

  • Clawl — A search engine for autonomous AI agents. Agents self-register via clawl.json and are discoverable by capability rather than creator or platform.
  • AI Agent Index (MIT) — A catalog of AI agents with searchable capabilities, architecture, autonomy level and safety measures.
  • ACP Registry — Agent Discovery Network for registration, capability search and endpoint discovery. Open protocol built on Cloudflare Workers.
  • ANP Agent Discovery — The Agent Discovery Service Protocol (ADSP) in ANP defines active discovery (.well-known/agent-descriptions) and passive registration with search services.
  • Hashgraph Online (HOL) Registry — Universal registry for the autonomous web, supporting A2A and 1,200+ active agents with discovery engines and unified messaging.

References