Full-Code · Orchestration Frameworksactive

Agent Network Protocol (ANP)

Type: full-code  ·  Vendor: Agent Network Protocol  ·  Language: Python, TypeScript  ·  License: MIT  ·  Status: active  ·  Status in practice: experimental

Links: homepage repo

Open specification (and reference implementation) for decentralised agent-to-agent communication: cross-platform DID identity, dynamic protocol negotiation, and an application-layer description / discovery protocol so agents from different vendors can talk without a central broker.

Description. Agent Network Protocol (ANP) is an open communication protocol for intelligent agents, positioned by its authors as 'the HTTP of the Agentic Web era'. It defines a three-layer architecture: an Identity and Secure Communication Layer based on W3C DID (with the did:wba method) for cross-platform identity and end-to-end encrypted channels; a Meta-Protocol Layer that lets agents negotiate the actual wire protocol in natural language and then run AI-generated code over it; and an Application Protocol Layer with the Agent Description Protocol (ADP) and an agent discovery protocol. The reference Python SDK lives in the AgentConnect repo.

Agent loop shape. Network-protocol layer rather than a single agent loop. Each participating agent embeds an ANP client. Step 1: identity / handshake — the client signs the first HTTP request with its DID key; the peer fetches the DID document and verifies the signature, returning a token for subsequent calls. Step 2: meta-protocol negotiation in natural language to agree the wire protocol (or reuse a cached one). Step 3: discovery via ADP description documents and the discovery protocol. Step 4: application-layer message exchange over the negotiated protocol, tagged with a 2-bit Protocol Type (meta / application / natural-language / verification).

Primary use cases

  • cross-organisation agent-to-agent communication without a central broker
  • decentralised agent identity (W3C DID / did:wba) and end-to-end encrypted channels
  • agent capability description (ADP) and peer discovery
  • interop between heterogeneous agent runtimes / vendors
  • research on self-organising, self-negotiating agent networks

Key concepts

  • Three-layer protocol architecture (docs) — Identity & Secure Communication Layer + Meta-Protocol Layer + Application Protocol Layer — the spec's top-level structure separating trust, negotiation, and capability description.
  • W3C DID (did:wba method) → decentralized-agent-network (docs) — Decentralised identifiers and DID documents give each agent a cross-platform identity; ANP's did:wba method extends did:web for agent communication so peers can verify each other without a central authority.
  • Agent Description Protocol (ADP) → inter-agent-communication (docs) — Application-layer protocol by which an agent publishes a structured description document declaring its capabilities and supported application protocols.
  • Agent Discovery Protocol → decentralized-agent-network (docs) — Mechanism for one agent to locate peers and fetch their ADP description documents — the discovery surface for an open agent network.
  • Meta-protocol negotiation (docs) — Two agents negotiate the actual wire protocol in natural language, then generate code to handle it; cached negotiation results can be reused on subsequent calls.
  • Protocol Type (PT) header → performative-message (docs) — 2-bit field at the head of every decrypted ANP message labelling the speech-act category: 00 meta protocol, 01 application protocol, 10 natural language protocol, 11 verification protocol.
  • End-to-end encrypted channel (ECDHE) (docs) — Each agent's DID key pair is used both for authentication and for an ECDHE-based end-to-end encryption scheme, so intermediate forwarders cannot decrypt traffic.

Patterns this full-code implements —

Provenance

  • Last analyzed:
  • Last updated:
  • Verification status: verified