/ Architecture

The OpenAN architecture.

From the A2A-T SDK to telecom-grade extensions and multi-agent use cases — how OpenAN connects the service layer and the network layer for end-to-end closed-loop autonomous networks.

Project Scope

A multi-agent collaboration framework for OSS, EMS, and the network layer.

OpenAN sits between the service layer (OSS copilots, optimization and fault agents) and the network layer (RAN, fixed access, IP, core). The A2A-T SDK extends Google's A2A protocol with telecom-grade primitives so agents in either layer can register, discover, negotiate, and act safely on the network.

3 Open source components
4 A2A-T extensions
2 Transport protocols
OpenAN Architecture Diagram
/ Seed Components

Three components,
open sourced together.

In June 2026 at DTW-Copenhagen, OpenAN open sources its three seed components. From September to December, scenario-based solution packages are released through joint innovation with operators.

01 / SDK

A2A-T SDK

A telecom-grade extension of the A2A protocol, designed for deterministic, network-scenario-compliant cross-layer agent collaboration. Supports JSON-RPC and HTTP+REST.

JSON-RPC HTTP+REST Apache 2.0 A2A-T-SDK-Python A2A-T-SDK-Java
02 / Discovery

Registry Center

The "agent labor market" — a yellow-pages service that finds and manages agent addresses, registers skills, handles health management, and uses LLM-aided retrieval to map intents to the right agent.

Agent Cards Skill Registry RAG Retrieval
03 / Workflow

Orchestration Center

A hub that orchestrates agents into reliable, reusable workflows via a workflow canvas, condition editor, property manager, and a workflow market — bridging manual design and autonomous programming.

Workflow Canvas Autonomous Programming Knowledge & RAG
/ A2A-T Extensions

Telecom-grade primitives on top
of the A2A protocol.

A2A is the foundation for inter-agent communication. For telecom, A2A-T adds four extensions so agents can describe tasks precisely, share network events efficiently, negotiate around ambiguity, and act on the network only with explicit authorization.

Task-T Schema

A structured task schema that defines a deterministic prompt for each task, eliminating the ambiguity of free-form requests between agents.

Notification-T Eventing

Network event subscription and publication. Optimized for the data/event reporting patterns that telecom O&M relies on at scale.

Negotiation-T Control

Controllable, structured negotiation between agents to resolve ambiguity and confirm feasibility before action is taken.

Authorization-T Security

Dynamic security: agent card and skill limitations plus explicit authorization before any agent makes network-changing operations.

Agent Cards Base

A2A base fields — name, description, URL, provider, version, capabilities, authentication, default input modes, and skills — fully reused by A2A-T.

Transport Protocol

JSON-RPC for fixed-format, one-way messages; HTTP+REST for stateful, long-lived connections over URL-addressable resources.
/ Multi-Agent Use Case

Home broadband
complaint handling.

When a customer calls in with "no internet for two days" or "YouTube is slow", OpenAN orchestrates the right agents end-to-end — billing, CDN, RAN, and vendor-specific fault agents — and hands off to a human expert only when needed.

Workflow

From call center to NOC, with humans in the loop.

The Home Broadband Complaint Handling Agent recalls a workflow from the Orchestration Center, retrieves the matching agents from the Registry Center, and coordinates them over A2A-T.

Different complaints trigger different workflows: a billing issue invokes the customer-care and billing agents; a CDN issue invokes customer-care, the CDN agent, and a vendor FAN agent.

  1. Intent capture — Call center receives the customer complaint and forwards intent to the Complaint Handling Agent.
  2. Workflow recall — The Orchestration Center selects the workflow that matches the complaint type.
  3. Agent retrieval — The Registry Center returns matching agent cards (customer care, billing, CDN, vendor FAN, etc.).
  4. Multi-agent execution — Agents collaborate over A2A-T to diagnose and resolve the issue.
  5. Expert handoff — When the workflow cannot resolve the case, the NOC expert is brought in with full context.