A2A Inter-System Business Flow Demo

A telecom call center demo demonstrating end-to-end business process integration across multiple subsystems using the A2A (Agent-to-Agent) Protocol.

---

Overview

A customer calls in via WebRTC. CC-server receives the call event and orchestrates CRM and Billing subsystems via A2A (JSON-RPC 2.0 over HTTP) to complete a full business loop: customer lookup, bill inquiry, AI-powered NBO recommendation, identity verification, and order placement — all surfaced on the agent dashboard in real time.

---

Architecture

[CC-client] — WebRTC offer/answer + call-start event
      │
      ▼
[CC-server] — Orchestrator + WebRTC signaling relay
      │
      ├── A2A: query_customer ──────────────────► [CRM-server / Profiling Agent]
      │
      ├── A2A: query_bill ──────────────────────► [Billing-server / Usage Agent]
      │
      ├── A2A: get_ai_model_status (optional) ──► [CRM-server / AI Management Agent]
      │
      ├── A2A: get_nbo ────────────────────────► [CRM-server / Recommendation Agent]
      │                                                      └── Ollama (qwen2.5:7b)
      ├── A2A: verify_identity ────────────────► [CRM-server / Profiling Agent]
      │
      ├── A2A: create_order ───────────────────► [CRM-server / Order Agent]
      │
      └── A2A: send_notification ──────────────► [CC-server / Communication Agent]
                                                        │
                                                        ▼
                                                  [CC-gui] — Agent dashboard

---

Project Structure

app/
├── Billing/
│   ├── Billing-server/     # FastAPI — Usage Agent (query_bill) [port 8003]
│   └── Billing-gui/        # Vue 3 — bill details view [port 5175]
├── CC/
│   ├── CC-server/          # FastAPI — orchestrator + WebRTC signaling + Communication Agent [port 8001]
│   ├── CC-client/          # Vue 3 — customer WebRTC caller [port 5172]
│   └── CC-gui/             # Vue 3 — agent dashboard + WebRTC receiver [port 5173]
└── CRM/
    ├── CRM-server/         # FastAPI — Profiling, Recommendation, Order, AI Management agents [port 8002]
    └── CRM-gui/            # Vue 3 — customer 360 view [port 5174]

openspec/                   # Spec-driven development artifacts
├── project.md              # Full project overview and architecture
├── AGENTS.md               # AI coding assistant instructions and conventions
├── specs/                  # Authoritative specs per subsystem
└── changes/                # Change proposals and delta specs

---

Prerequisites

Tool	Version	Install
Python	3.11+	python.org
uv	latest	pip install uv
Node.js	18+	nodejs.org
npm	bundled with Node	—
Ollama	latest	ollama.com

Pull the NBO model used by the Recommendation Agent:

ollama pull qwen2.5:7b

---

Services & Ports

Service	Port	Role
CC-server	8001	Orchestrator, WebRTC signaling relay, Communication Agent
CC-client	5172	Customer-side WebRTC caller (Vue 3 GUI)
CC-gui	5173	Agent dashboard, WebRTC receiver (Vue 3 GUI)
CRM-server	8002	Profiling, Recommendation, Order, AI Management agents
CRM-gui	5174	Customer 360 view (Vue 3 GUI)
Billing-server	8003	Usage Agent
Billing-gui	5175	Bill details view (Vue 3 GUI)

WebRTC signaling: ws://localhost:8001/ws/signal (served by CC-server)

---

Getting Started

1. Start Ollama

ollama serve

2. Start backend servers

Each backend uses uv. Run from the respective server directory:

# CC-server
cd app/CC/CC-server
uv run uvicorn src.main:app --port 8001 --reload

# CRM-server
cd app/CRM/CRM-server
uv run uvicorn src.main:app --port 8002 --reload

# Billing-server
cd app/Billing/Billing-server
uv run uvicorn src.main:app --port 8003 --reload

Seed data is loaded automatically at startup via seed.py.

3. Start frontend GUIs

Run from each GUI directory:

npm install
npm run dev

GUI	URL
CC-client	http://localhost:5172
CC-gui	http://localhost:5173
CRM-gui	http://localhost:5174
Billing-gui	http://localhost:5175

---

Demo: WebRTC Call Flow

1. Open CC-gui at http://localhost:5173 — agent dashboard, waiting for a call
2. Open CC-client at http://localhost:5172 — customer caller
3. Enter a phone number (e.g. 13800000001) and click Call
4. CC-client sends a WebRTC offer → CC-server relays it → CC-gui answers → P2P audio/video established
5. CC-client sends call-start { phone } → CC-server runs the A2A business flow:
   • Looks up customer profile (CRM)
   • Queries bill summary (Billing)
   • Checks NBO AI model health (CRM)
   • Fetches NBO recommendations via Ollama (CRM)
6. CC-gui displays: customer info, bill summary, and ranked NBO recommendations
7. Agent selects an offer → CC-gui sends POST /order/create → CC-server verifies identity, places order, and sends notification
8. CC-gui displays order confirmation and notification status

---

A2A Skills Reference

Agent	Endpoint	Skill	TMF API
Profiling Agent	POST /profiling/a2a	query_customer	TMF629 / TMF637
Profiling Agent	POST /profiling/a2a	verify_identity	TMF720
Recommendation Agent	POST /recommendation/a2a	get_nbo	TMF701 / TMF620
Order Agent	POST /order/a2a	create_order	TMF622
AI Management Agent	POST /ai-management/a2a	get_ai_model_status	TMF915
Usage Agent	POST /usage/a2a	query_bill	TMF677 / TMF678
Communication Agent	POST /communication/a2a	send_notification	TMF681

Agent Cards are discoverable at GET /<agent-prefix>/.well-known/agent.json.

All A2A calls use JSON-RPC 2.0:

{
  "jsonrpc": "2.0",
  "method": "tasks/send",
  "params": { "skill": "<skill_name>", "input": { ... } },
  "id": "<uuid>"
}

---

TMForum Alignment

Skill interfaces align with TM Forum Open API data models (TMF620, TMF622, TMF629, TMF637, TMF677, TMF681, TMF701, TMF720, TMF915). Field names and status enumerations follow TMF conventions where practical. See openspec/project.md for the full TMForum field mapping table.

---

Tech Stack

Layer	Technology
Backend language	Python 3.11+
Backend framework	FastAPI
A2A protocol	Google A2A SDK (JSON-RPC 2.0 over HTTP)
LLM backend	Ollama (qwen2.5:7b)
Data store	SQLite (per subsystem, demo only)
Async HTTP client	httpx (AsyncClient)
Backend dependency mgr	uv + pyproject.toml
Frontend framework	Vue 3 (Composition API, <script setup>)
Build tool	Vite
UI library	Element Plus
State management	Pinia
HTTP client (frontend)	Axios
WebRTC	Browser-native RTCPeerConnection (no STUN/TURN)
Frontend language	TypeScript
Frontend pkg manager	npm

---

Development

This project uses OpenSpec spec-driven development. All changes should go through the openspec/ workflow:

• Read openspec/AGENTS.md before making any changes — it contains critical architecture rules, directory conventions, and the A2A skills registry.
• Read openspec/project.md for the full system overview and TMForum alignment tables.
• Use /opsx:propose, /opsx:apply, and /opsx:archive to manage changes (see AGENTS.md for details).

Spec Annotation Tags

Markdown files in openspec/ support inline annotation tags processed by Claude Code (see CLAUDE.md):

Tag	Action
@claude: <instruction>	Execute the instruction at that location
@revise: <requirement>	Rewrite the adjacent paragraph to meet the requirement
@expand: <direction>	Expand and elaborate on the adjacent content
@delete	Remove the annotated content block entirely
@todo	Mark as incomplete — prompts for input before proceeding
@discuss: <question>	Pause for analysis and options before editing
@keep	Do not modify this content under any circumstance
@note: <remark>	Informational only — preserved, no modification triggered

Demo seed data

Each server seeds its SQLite database at startup with 3 test customers:

Phone	Customer	Plan
13800000001	Customer A	Plan-50G
13800000002	Customer B	Plan-100G
13800000003	Customer C	Plan-200G