Wiki · Concept · Last reviewed June 25, 2026

Agent Network Protocol

Agent Network Protocol, or ANP, is an open-source protocol stack for interoperable AI agents, centered on decentralized identity, agent description, discovery, messaging, and payment flows.

Category: Concept / Agent Infrastructure Published: June 25, 2026 Modified: June 25, 2026 Last reviewed: June 25, 2026 Tags: ANP, AI Agents, Agent Identity, Discovery, Messaging, DID

Definition

Agent Network Protocol is an open-source communication protocol stack for AI agents. The official ANP site describes it as an "open protocol stack for the Agentic Web" covering decentralized identity, discovery, messaging, and payment protocols for interoperable agents. The GitHub repository frames ANP as a suite for agent identity, naming, discovery, negotiation, secure messaging, and application-level collaboration.

ANP is adjacent to Model Context Protocol and Agent2Agent Protocol, but it is not the same object. MCP focuses on tool and resource access. A2A focuses on task exchange between agents. ANP is broader and more web-like: it tries to define how agents identify themselves, publish descriptions, discover one another, exchange messages, and use domain protocols such as payments.

Architecture

The ANP 1.1 documentation presents a three-layer architecture. The bottom layer reuses ordinary Internet infrastructure such as HTTP, DNS, TLS, certificate authorities, CDNs, and search. The middle layer supplies identity and communication infrastructure, including DID:WBA identity, WNS handles, authentication, key material, secure messaging, and cross-domain communication foundations. The application layer includes Agent Description, Agent Discovery, and domain-specific protocols for payment, authorization, authentication, commerce, and other workflows.

The released ANP 1.1 suite, according to the project README, covers did:wba identity, WNS handles, agent description, agent discovery, end-to-end instant messaging, and an ANP adaptation of AP2 for agent payments. The same README says the meta-protocol remains draft rather than released.

The Agent Description Protocol, ANP-07, is released at version 1.1. It defines how an agent publishes public information, supported interfaces, capabilities, and other details. The specification uses JSON as a base format, supports linked-data and semantic-web features, adopts did:wba as a security mechanism, and allows interoperability with existing protocols such as OpenAPI and JSON-RPC.

The Agent Discovery Protocol, ANP-08, is also released at version 1.1. It uses JSON-LD and defines active discovery through .well-known URI paths plus passive discovery through submission of agent descriptions to search services. The did:wba method specification, ANP-03, is released at version 1.1 and extends did:web for agent communication, including request signatures based on HTTP Message Signatures and body integrity through Content-Digest.

Governance and Safety

ANP makes agent identity a first-order governance surface. If a remote agent can publish a description, accept messages, claim capabilities, and participate in payment flows, then the institution using it needs evidence about who operates it, which DID or handle identifies it, what keys are bound to it, which interfaces are exposed, and what version of the description was trusted.

The largest risk is confusing discoverability with trust. An agent description is not a certification, and a DID is not a safety case. A malicious or compromised agent can publish plausible capabilities, leak data through messaging, induce overbroad delegation, or participate in commerce flows whose authorization evidence is incomplete. Search-style discovery also raises ranking and poisoning problems: which agents are surfaced, why, and by whom?

ANP's messaging and payment layers make auditability more important. The instant-messaging overview is marked draft and describes cross-domain direct messages, group messages, end-to-end encryption, attachments, routing, relaying, and federation. The payment protocol is released at version 1.1 and adapts AP2 concepts such as mandates, receipts, signatures, and agent payment roles. Those are useful records, but they do not settle consumer protection, liability, refund, fraud, or procurement policy.

Evidence Record

A governed ANP deployment should preserve the agent description document, DID document, WNS handle where used, public keys, proof or signature status, exposed interfaces, discovery path, protocol version, messaging profile, payment mandate or receipt where relevant, delegated scopes, human authorization requirement, and incident contact. For high-impact workflows, logs should also record which search or registry returned the agent and why that agent was selected.

Source Discipline

Use the ANP site and repository for project claims, current release-line claims, document status, and architecture. Use the individual ANP specifications for claims about did:wba, agent description, discovery, messaging, and payments. Treat the IETF "Framework for AI Agent Networks" document as an Internet-Draft and therefore work in progress; the draft itself says Internet-Drafts may be updated, replaced, or obsoleted and should not be treated as final standards.

Spiralist Reading

ANP is a proposed address system for machine social life. It asks what an agent is when it is no longer just a local tool: a name, a key, a description, a set of interfaces, a mailbox, a payment endpoint, and a record of responsibility.

Spiralism reads that stack with caution. Agent networks will not be governed by identity alone. They will be governed by which identities are discoverable, which descriptions are believed, which messages cross domains, which payment proofs count as consent, and which logs survive when something goes wrong.

Open Questions

Sources

Return to Wiki