Multi-Agent AI Purchasing System

Cross-framework Agent-to-Agent (A2A) commerce demo — Google ADK concierge · CrewAI burger seller (vLLM) · LangGraph pizza seller (Ollama) on AMD Instinct GPUs

---

Overview

Prove that heterogeneous agent frameworks can collaborate on a purchasing task through Google’s open A2A protocol — with 100% local Llama 3.1 inference on AMD GPUs (no cloud model APIs required for the demo path).

Agent	Framework	LLM backend	Port	Auth
Purchasing Concierge (root)	Google ADK + LiteLLM	Ollama llama3.1	Gradio UI 8087	—
Burger Seller	CrewAI	vLLM OpenAI-compatible · Llama 3.1-8B	10001	HTTP Basic
Pizza Seller	LangGraph ReAct + MemorySaver	Ollama	10000	Bearer

vLLM OpenAI base defaults to http://localhost:8088/v1 with --gpu-memory-utilization **0.6** so Ollama can share the Instinct GPU.

Portfolio signal for agentic systems / multi-agent orchestration / on-prem LLM serving: A2A discovery + task lifecycle, dual auth schemes, tool-calling, and AMD/ROCm ops docs.

Numbers below come from committed configs, menus, scripts, and tests. No Instinct latency/throughput leaderboard is published in-repo — do not invent one.

---

Results & repository facts (traceable)

Topology defaults

Fact	Value	Source
Agents	3 (ADK · CrewAI · LangGraph)	docs/ARCHITECTURE.md
Seller ports	Pizza 10000 · Burger 10001	agent configs
vLLM serve port	8088	scripts/start_vllm.sh
Gradio UI port	8087	.env.example / app.py
vLLM GPU mem util	0.6 (40% headroom for Ollama)	start_vllm.sh
Default models	meta-llama/Llama-3.1-8B-Instruct (vLLM) · llama3.1:latest (Ollama)	configs
Package	adk-multi-agent-system v1.0.0 · MIT · Python ≥3.11	pyproject.toml
Tracked files	34	git tree
Languages	Python 48,833 B · Shell 5,196 B	GitHub API
pytest	16 cases (burger 6 · pizza 5 · purchasing 5)	tests/
CI matrix	Python 3.11 + 3.12 · ruff · coverage	.github/workflows/ci.yml

%%{init: {'theme':'base'}}%%
xychart-beta
  title "Default service ports"
  x-axis ["Pizza A2A", "Burger A2A", "vLLM", "Gradio UI"]
  y-axis "Port" 8000 --> 10100
  bar [10000, 10001, 8088, 8087]

Loading

%%{init: {'theme':'base'}}%%
pie showData title Language composition (bytes)
    "Python" : 48833
    "Shell" : 5196

Loading

Menu catalogs (committed pricing in IDR ×1000)

Burger menu (4 items)	Price	Pizza menu (5 items)	Price
Classic Cheeseburger	85	Margherita	100
Double Cheeseburger	110	Pepperoni	140
Spicy Chicken Burger	80	Hawaiian	110
Spicy Cajun Burger	85	Veggie	100
		BBQ Chicken	130

xychart-beta
  title "Menu price tags (IDR K) — burgers"
  x-axis ["Classic", "Double", "Chicken", "Cajun"]
  y-axis "IDR K" 0 --> 120
  bar [85, 110, 80, 85]

Loading

xychart-beta
  title "Menu price tags (IDR K) — pizzas"
  x-axis ["Margherita", "Pepperoni", "Hawaiian", "Veggie", "BBQ"]
  y-axis "IDR K" 0 --> 150
  bar [100, 140, 110, 100, 130]

Loading

Session state keys (ADK)

Key	Type	Purpose
session_id	UUID str	Correlation across remote A2A tasks
session_active	bool	Seller task in flight
active_agent	str	Currently engaged seller

AMD hardware matrix (documented)

GPU	VRAM	Notes
MI250X	2×64 GB HBM2e	Recommended dual-model
MI300X	192 GB HBM3	Large-model headroom
MI210	64 GB HBM2e	Supported
RX 7900 XTX	24 GB	Consumer · reduced headroom

Explicitly not claimed

• No checked-in end-to-end order latency / tokens-sec on Instinct
• Demo auth secrets in .env.example are placeholders — replace before any network exposure

---

Architecture

flowchart TD
  U[User · Gradio :8087] --> R[Purchasing Concierge<br/>Google ADK + LiteLLM]
  R -->|Ollama llama3.1| O[Ollama :11434]
  R -->|A2A Basic · POST /tasks/send| B[Burger Seller<br/>CrewAI · :10001]
  R -->|A2A Bearer · POST /tasks/send| P[Pizza Seller<br/>LangGraph · :10000]
  B -->|OpenAI-compatible| V[vLLM Llama-3.1-8B · :8088<br/>gpu-memory-utilization 0.6]
  P --> O
  B --> TB[create_burger_order]
  P --> TP[create_pizza_order]

Loading

sequenceDiagram
  participant U as User
  participant R as ADK Concierge
  participant C as Agent Card GET
  participant S as Seller A2A server
  U->>R: food request
  R->>C: /.well-known/agent.json
  C-->>R: AgentCard + skills
  R->>S: POST /tasks/send {id, sessionId, message}
  S->>S: LLM + order tool
  S-->>R: TaskStatus COMPLETED / INPUT_REQUIRED / FAILED
  alt INPUT_REQUIRED
    R-->>U: escalate for clarification
  end
  R-->>U: streamed text / order result

Loading

flowchart LR
  subgraph TaskState["A2A TaskState"]
    A[SUBMITTED] --> B[WORKING]
    B --> C[COMPLETED]
    B --> D[INPUT_REQUIRED]
    B --> E[FAILED]
    B --> F[CANCELED]
  end

Loading

Deep dives: docs/ARCHITECTURE.md · docs/A2A_PROTOCOL.md · docs/AMD_GPU_SETUP.md

---

Why this design

Choice	Rationale (from docs)
Three frameworks	Demonstrate A2A interoperability, not vendor lock-in
vLLM for burger	CrewAI needs OpenAI-style tool calling + Llama 3.1 JSON chat template
Ollama for pizza + root	Lighter / GGUF; shares Instinct with vLLM at 0.6 GPU mem util
Basic vs Bearer	Auth-agnostic A2A — each seller enforces its own scheme
JWKS push path	Both sellers expose /.well-known/jwks.json for signed push notifications

---

Repository layout

Multi-Agent-AI-Purchasing-System-with-Google-ADK-AMD-Instinct-GPUs/
├── agents/
│   ├── purchasing_agent/   # ADK root + Gradio app + A2A connections
│   ├── burger_agent/       # CrewAI + Basic-auth A2A server :10001
│   └── pizza_agent/        # LangGraph + Bearer A2A server :10000
├── docs/{ARCHITECTURE,A2A_PROTOCOL,AMD_GPU_SETUP}.md
├── scripts/{start_all,start_vllm,start_ollama}.sh
├── tests/                  # 16 pytest cases
├── utils/                  # A2A client/server helpers
├── requirements.txt · pyproject.toml · .env.example
└── .github/workflows/ci.yml

---

Quickstart

git clone https://github.com/ArchanaChetan07/Multi-Agent-AI-Purchasing-System-with-Google-ADK-AMD-Instinct-GPUs.git
cd Multi-Agent-AI-Purchasing-System-with-Google-ADK-AMD-Instinct-GPUs

python -m venv .venv && source .venv/bin/activate   # Win: .\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
cp .env.example .env   # set HF_TOKEN; rotate demo passwords

# AMD / ROCm host with vLLM + Ollama installed:
export HF_TOKEN=...
bash scripts/start_all.sh
# UI → http://localhost:8087

pytest tests/ -v

Standalone backends: scripts/start_vllm.sh · scripts/start_ollama.sh.

---

Testing & CI

Suite	Focus
test_burger_agent.py	Order tool · TaskState mapping · fallback
test_pizza_agent.py	Order tool · INPUT_REQUIRED / COMPLETED mapping
test_purchasing_agent.py	Text parts · Agent Card resolve / skip unreachable
Actions	Matrix 3.11/3.12 · pytest-cov · ruff check/format

flowchart LR
  Push[git push] --> CI[GitHub Actions]
  CI --> Py[Python 3.11 / 3.12]
  Py --> Test[pytest --cov=agents]
  CI --> Lint[ruff check + format]

Loading

---

Tech stack & keywords

Layer	Technology
Orchestration	Google ADK, Gradio UI
Seller agents	CrewAI, LangGraph / LangChain
Inference	vLLM (ROCm), Ollama, LiteLLM
Protocol	A2A HTTP · Agent Cards · TaskStatus
Security	Basic · Bearer · JWKS / python-jose
Hardware docs	AMD Instinct MI210 / MI250X / MI300X · ROCm 6.x
Quality	pytest · asyncio · ruff · GitHub Actions

Keyword surface: Python · multi-agent · agentic AI · Google ADK · A2A protocol · CrewAI · LangGraph · vLLM · Ollama · Llama 3.1 · AMD Instinct · ROCm · tool calling · Gradio · local LLM · distributed agents · pytest · CI/CD

---

Roadmap

• Capture measured TTFT / order E2E latency on Instinct and commit JSON evidence
• Remove stray root Null artifact
• HTTPS termination + production secret management for seller auth

---

Multi-Agent AI Purchasing System · MIT · v1.0.0
github.com/ArchanaChetan07/Multi-Agent-AI-Purchasing-System-with-Google-ADK-AMD-Instinct-GPUs