bench get-app https://github.com/aws-samples/sample-multi-agent-procure-to-pay

[![Listed on Frappe Gems](https://frappegems.com/api/method/frappe_gems.seo.badge?app=aws-samples%2Fsample-multi-agent-procure-to-pay)](https://frappegems.com/gems/apps/aws-samples/sample-multi-agent-procure-to-pay)

ARIA — Agentic Procure-to-Pay on Amazon Bedrock AgentCore

Buy something in plain English. Watch eight AI agents run the entire purchase — sourcing, approval, PO, receipt, invoice, payment — and never touch a form.

ARIA is a reference implementation of an autonomous Procure-to-Pay platform on AWS. You type "I need 200 hex bolts." A fleet of specialized agents — orchestrated by Amazon Bedrock AgentCore, gated by Cedar authorization, and checked by a Bedrock Automated Reasoning guardrail — turns that sentence into a validated, budget-checked, supplier-scored purchase order, then follows the order all the way to an optimized payment. Every decision is reasoned, audited, and reversible.

The system of record is ERPNext, reached through a swappable canonical adapter — so the same agents can drive SAP, Infor, or Workday without a line of agent code changing.

> Sample code, for non-production use. Work with your security and legal teams to meet your organizational requirements before deploying.

Ask for what you need — the agent files the request

---

Meet the fleet

Eight agents, each an expert at one step of the pipeline. They hand off to each other like a well-run procurement team — except it happens in seconds, around the clock.

#	Agent	What it owns
1	Chat	Turns natural language into a Material Request in ERPNext
2	Requisition	Validates items, catches duplicates, compares pricing, checks the budget
3	Sourcing	Scores suppliers on price, delivery, quality, and capacity — recommends the winner
4	PO Management	Generates the Purchase Order with the chosen supplier
5	Receiving	Validates goods receipts against the PO when items arrive
6	Invoice Matching	Runs the three-way match: PO vs. Goods Receipt vs. Invoice
7	Payment	Times payment to capture early-pay discounts when they beat the cost of cash
8	Workflow	Chains requisition → sourcing → PO into one hands-off run

The judgment call that matters: requests ≤ $5K at LOW risk auto-approve; anything bigger or riskier escalates to a human — with the agent's full reasoning attached, so the approver decides in seconds, not hours.

---

See it work

A user asks for 200 hex bolts. Here's the whole arc:

"I need 200 hex bolts"
        │
        ▼
  ┌─────────────┐   creates Material Request in ERPNext
  │  Chat Agent │
  └─────────────┘
        │
        ▼
  ┌──────────────┐  items valid? duplicates? price sane? budget OK?
  │ Requisition  │
  └──────────────┘
        │
        ▼
  ┌──────────────┐  scores every supplier → recommends the best
  │  Sourcing    │
  └──────────────┘
        │
        ▼
   ≤ $5K + LOW risk ? ──── yes ──▶ auto-approve
        │
        no
        ▼
   human approver (with agent reasoning attached)
        │
        ▼
  ┌──────────────┐  cuts the PO → validates receipt → 3-way match → pays on time
  │ PO · Receipt │
  │ Invoice · Pay│
  └──────────────┘

Every hop is guardrail-validated and written to an audit trail you can replay.

---

Data

All data in this project is synthetic and fabricated for demonstration. There is no real, customer, production, or personally identifiable data anywhere in the repository or the demo environment.

• Seed data (utilities/data/) — static JSON that populates the demo ERPNext instance: 22 fictional suppliers, 60 catalog items, item groups, budgets, payment terms, and a handful of sample requisitions. Supplier names and contact domains are invented and illustrative; they do not correspond to real companies or mailboxes.
• Demo users — the personas (Maria, Sarah, Priya, …) are fictional and use reserved @example.com addresses. Passwords are supplied by the operator at setup time and are never committed.
• Simulation engine (utilities/simulation/, optional) — generates additional in-flight documents (requisitions, receipts, invoices) so the demo feels alive. It produces data two ways:
  1. Structural randomness — Python's random and Faker jitter quantities, prices, timestamps, and names. This is non-cryptographic and used only for demo variety, never for tokens, IDs, or any security decision.
  2. LLM-generated narrative — llm_generator.py calls Amazon Bedrock (Claude, via the structured-output Converse API) to produce realistic but entirely fictional document content. Prompts contain only the synthetic catalog above; no real or sensitive data is sent.

The generated documents live only in your own ERPNext instance and DynamoDB lifecycle table (ephemeral, TTL'd); nothing is transmitted to third parties beyond the Amazon Bedrock model invocation.

---

Try it in five minutes (locally)

No AWS deploy required. The local/ harness runs the entire stack on your machine — it emulates DynamoDB, S3, and Secrets Manager, shims the AgentCore Runtime, Gateway, and Memory APIs, and runs all seven agents in-process. The only thing that talks to real AWS is Bedrock (the actual model calls), so you get genuine agent reasoning without provisioning a thing.

cp local/env.example local/.env      # add AWS creds for real Bedrock
make -C local erpnext                 # local ERPNext (one-time setup wizard)
make -C local up                      # emulators + shims + 7 agents + backend
make -C local ui                      # the SPA (Vite), in another terminal

Open the SPA, sign in as a demo persona, and run a requisition end to end. Full walkthrough, port map, and limitations: local/README.md.

---

Architecture

Five design decisions define how ARIA stays safe, portable, and auditable:

Decision	Why it matters
ERPNext is the single source of truth	Agents never store procurement data. DynamoDB holds only ephemeral lifecycle state (runs, decisions) with TTL — no shadow copy to drift or leak.
Canonical adapter pattern	ERPAdapterBase → ERPNextAdapter. Implement the same 22-method interface and the identical agent fleet drives SAP, Infor, or Workday.
Per-user ERP identity	X-P2P-User-Email maps to per-user API keys; ERPNext enforces each person's real permissions. Agents can't do what the user can't.
Three independent safety layers	Cedar (authorization on tools) → prompt rules (reasoning guidance) → Automated Reasoning guardrail (formal output validation). Defense in depth, not one brittle check.
Agents never touch the database directly	Every read and write flows MCP Gateway → Cedar → Adapter Lambda → ERPNext REST. One choke point to authorize, log, and reason about.

Deep dive — full data-flow diagrams, agent workflow state machines, and the safety model: ARCHITECTURE.

---

Project structure

├── backend/                    # Python — API Lambda + AgentCore agents + ERP adapter
│   ├── main.py                 # API Lambda entrypoint (FastAPI + Mangum)
│   ├── agentcore_app.py        # AgentCore Runtime entrypoint (7 agents)
│   ├── agents/                 # Agent prompts and tools (one file per agent)
│   ├── adapters/               # Canonical API + ERPNext adapter + field maps
│   ├── services/               # DynamoDB, auth, lifecycle, textract
│   └── tests/                  # pytest suite
├── frontend/                   # React + TypeScript + Tailwind CSS
│   └── src/pages/              # Role-based pages (Chat, Dashboard, Decisions, …)
├── infra/                      # AWS CDK (TypeScript) — two stacks
│   ├── lib/p2p-agentic-stack.ts  # Cognito, Lambda, AgentCore, Gateway, Cedar
│   ├── lib/erpnext-stack.ts      # VPC, EC2, RDS, Redis, ALB
│   └── policies/                 # Cedar RBAC policies
├── local/                      # Local dev harness (emulators + AgentCore shims)
├── utilities/                  # Seed data + simulation engine
│   ├── scripts/                # ERPNext setup, seeding, verification
│   └── simulation/             # EventBridge-driven demand + invoice generation
└── docs/                       # Architecture + diagrams

---

Prerequisites

Tool	Version	Purpose
Node.js	18+	Frontend build, CDK CLI
Python	3.13+	Backend, seed scripts
AWS CDK CLI	2.x	Infrastructure deployment
Docker	24+	Agent containers, ERPNext
AWS CLI	2.x	Account with Bedrock + AgentCore access

---

Deploying to AWS

Deployment is partially automated — CDK provisions every AWS resource, and the numbered scripts in utilities/ create users, API keys, and seed data. A few steps still require the AWS or ERPNext console; each is called out below.

Before you start

These are the easy-to-miss prerequisites. Confirm each before running CDK.

1. Bedrock model access. Request access to Claude Sonnet 4.6 (us.anthropic.claude-sonnet-4-6) in the Bedrock console for your region. Without it, agents fail at runtime, not deploy time.
2. Strong passwords ready. No defaults are baked in — the scripts and CDK refuse to run without them:
   • erpnextAdminPassword and dbRootPassword in infra/cdk.context.json (copy from cdk.context.example.json).
   • ERPNEXT_USER_PASSWORD and ERPNEXT_PASSWORD in utilities/.env (copy from .env.example).
3. Container build tool. The agent image builds with finch. On macOS without Docker Desktop, install Finch and run finch vm start first.
4. (Optional) Custom domain. For real hostnames, register a domain and create a Route53 hosted zone, then pass --context hostedZoneName=. Omit it for a CloudFront-only deploy with HTTP-only ERPNext.

1. Deploy infrastructure

cd infra && npm install

# Stack 1: ERPNext — VPC, EC2, RDS MariaDB, ElastiCache Redis, ALB (external + internal)
npx cdk deploy ErpNextStack --context hostedZoneName=your-domain.example.com

# Stack 2: ARIA Platform — Cognito, API Gateway, Lambdas, AgentCore (7 runtimes),
#   MCP Gateway, Cedar policies, Bedrock Guardrail, DynamoDB, S3/CloudFront
npx cdk deploy P2PAgenticStack --context hostedZoneName=your-domain.example.com

CDK builds the ARM64 agent container from backend/Dockerfile, pushes it to the CDK-assets ECR, and creates every resource. Add --profile for non-default AWS profiles.

> Both stacks must finish before the next steps. Script 02 reads the Cognito User Pool ID from P2PAgenticStack outputs, and the frontend deploy reads the AgentCore ARNs.

2. Configure ERPNext (manual — wizard required)

ErpNextStack brings up an empty ERPNext site. Frappe's setup wizard has no fully-headless mode, so complete it once in the browser:

1. Open https://erp. (or the EC2 IP if you skipped the domain).
2. Run the wizard with exactly these values — the seed script depends on them:
   • Company: Apex Manufacturing Group · **A