ANP: The Agent Network Protocol and the Open Web of Agents

Today most AI agents live inside a platform. They can talk to the tools their platform provides and little else. The Agent Network Protocol (ANP) asks a bigger question: what would it take for agents anywhere on the internet to find, trust and talk to each other, the way any browser can reach any web server?

ANP is an open-source protocol, developed by a community under the Apache-2.0 license, that aims to be a common layer for agent identity, discovery and communication. Its stated goal is to be something like “the HTTP of the agentic web”.

How ANP is layered

flowchart TB
    A["Application layer: agent description, discovery, domain protocols"]
    B["Identity and encrypted communication layer: DID authentication, end-to-end encryption"]
    C["Existing internet: HTTP, DNS, TLS, CDN"]
    A --> B --> C

• Existing internet. ANP reuses HTTP, DNS and TLS instead of inventing a new network stack.
• Identity and encrypted communication. Agents are identified with W3C Decentralized Identifiers (DIDs). ANP supports the did:wba method and did:web. A DID lets an agent prove who it is without a central login provider, and it supports encrypted messaging between agents.
• Application layer. Agents publish a machine-readable description of what they offer, so other agents can discover them. Domain protocols, such as payments or authorization, can be defined on top.

The specification also describes a meta-protocol for agents to negotiate how they will talk. That part is still a draft, so treat it as direction and not as something to depend on.

ANP vs A2A vs MCP

Protocol	Focus
MCP	Giving one agent access to tools and data
A2A	Structured collaboration between known agents
ANP	Open-internet identity, discovery and trust between agents

These overlap less than they seem. A2A assumes you know which agent to call and focuses on tasks. ANP puts more weight on identity and discovery across organizational boundaries. They could be used together.

Why identity comes first

If any agent can claim to be any other, nothing built on top is safe. DIDs give each agent a verifiable identity that does not depend on one company. That matters for:

• Trust decisions: allow, block or rate-limit specific agents.
• Auditing: record which agent did what.
• Privacy: encrypted channels keep content away from intermediaries.

What builders should do

1. Watch, do not bet. ANP is a community project. Check the maturity of each part before adopting it.
2. Learn DIDs. Even outside ANP, decentralized identity is becoming part of agent security.
3. Keep identity separate from transport in your own design, so you can adopt a standard later.
4. Never trust a remote agent by default. Verify identity, then limit permissions.

The takeaway

MCP connects agents to tools and A2A connects them to each other. ANP is a proposal for the layer beneath both: who an agent is, and how others can find and trust it. It is early, but identity is the problem every open agent ecosystem eventually has to solve.