Skip to content

PRD.md

Latest commit

 

History

History
97 lines (70 loc) · 4.09 KB

PRD.md

File metadata and controls

97 lines (70 loc) · 4.09 KB

📄 PRD: Project "Aura" - Autonomous Reliable Agentic Commerce

Version: 2026.1.0 Stack: Google ADK (Python), UCP, AP2, Kagent (Kubernetes), Gemini 2.5 Flash Status: Hackathon Prototype (Google AI Agent Labs Oslo 2026 — Team 6)

1. Executive Summary

Aura is a multi-agent, cloud-agnostic procurement system designed to automate the B2B commerce lifecycle. Unlike traditional "shopping bots," Aura integrates Real-time Compliance (KYC/AML) and Verifiable Intent to allow agents to discover, negotiate, and settle transactions autonomously on behalf of an enterprise. It bridges the gap between modern AI agents and rigid Core Banking Systems (BMS).

2. System Architecture (Multi-Agent Squad)

Aura operates as a MAS (Multi-Agent System) using the Google ADK orchestration layer. Each agent is an independent entity that can be deployed as a Kubernetes Agent resource.

Agent Responsibility Key Protocol/Tool
Architect Root LlmAgent; parses user intent; manages sub-agent state. adk.LlmAgent
Scout Discovers vendors via Universal Commerce Protocol (UCP). ucp_tools.discover_vendors
Sentinel Executes KYC/AML checks against core banking (BMS) logic. compliance_tools.verify_vendor_compliance
Closer Handles secure payment via Agent Payments Protocol (AP2). ap2_tools.generate_intent_mandate + settle_cart_mandate

3. Functional Requirements

FR-01: Agentic Discovery (UCP)

  • Action: The Scout agent must query /.well-known/ucp endpoints of potential vendors.
  • Validation: It must parse the dev.ucp.shopping capability manifest to find product pricing and availability.
  • Output: Return a structured list of VendorEndpoint objects to the Architect.

FR-02: Compliance-First Vetting (The Sentinel)

  • Action: Before a cart is finalised, the Sentinel must cross-reference the VendorID against the internal BMS (Core Banking) compliance database.
  • Requirement: Return a "Compliance Hash" for valid vendors; immediately block and flag vendors like "ShadowHardware" (mocked blacklist).

FR-03: Verifiable Intent & Settlement (AP2)

  • Action: The Closer agent must generate an Intent Mandate (W3C Verifiable Credential).
  • Execution: Simulate a secure checkout using a Cart Mandate signed via the AP2 protocol, ensuring the payment is routed through a compliant banking gateway.

FR-04: Cloud-Agnostic Kubernetes Deployment

  • Platform: Deploy via Kagent (Cloud Native Agentic AI framework).
  • Resiliency: Agents must be defined as Agent Custom Resources (CRDs) to allow for cross-cloud portability (GCP/AWS/On-prem).

4. Technical Specifications

Data Model: The AP2 Intent Mandate

{
  "type": "IntentMandate",
  "id": "<uuid>",
  "issued_at": 1741694400,
  "vendor": {
    "id": "v-001",
    "name": "TechCorp Nordic"
  },
  "constraints": {
    "max_amount": 5000.00,
    "amount": 3897.00,
    "currency": "USD",
    "compliance_required": true,
    "compliance_hash": "<64-char-hex>"
  },
  "proof": {
    "type": "ecdsa-p256-signature",
    "value": "MOCK_SIGNATURE_HASH",
    "created": 1741694400
  }
}

Infrastructure Target

  • Runtime: Python 3.12 (FastAPI for A2A comms, port 8080)
  • Orchestration: google-adk with SequentialAgent for the main flow
  • Model: Vertex AI Gemini 2.5 Flash (gemini-2.5-flash)
  • GCP Project: ai-agent-labs-oslo-26-team-6
  • Region: us-central1
  • Deployment: kagent YAML with ModelConfig targeting Vertex AI

5. Implementation Prompts (Copilot Agent Mode)

  1. Project Init: Scaffold with agent-starter-pack, requirements.txt, .env, agents/ and tools/ folders.
  2. The Sentinel Logic: Implement verify_vendor_compliance(vendor_name) — reject "ShadowHardware", return 64-char ComplianceHash for others.
  3. The Orchestrator: Chain Scout → Sentinel → Closer via SequentialAgent. Flow proceeds to Closer only if Sentinel passes.
  4. K8s Containerisation: Multi-stage Dockerfile + kagent.yaml following kagent.dev/v1alpha2 spec.
⚡