ACP-Compliant Checkout Rail for Autonomous AI Buyer Agents
Translates raw autonomous AI agent purchasing intents into bounded, gated, and auditable orders on the Razorpay rail. Enforcing server-authoritative catalog pricing, deterministic guardrails, 30-min inventory soft-holds, and zero-hallucination checkouts.
The Vulnerability in Direct Agentic Payments
Giving autonomous buyer agents direct API keys to merchant payment gateways exposes balance sheets to prompt injections, cart hallucinations, and race conditions. AgentPay Bridge installs an immutable verification layer between agent intents and money movement.
Autonomous agents inject arbitrary client prices (e.g. Rs 1 for a Rs 4,999 SKU). Gateways lacking catalog authority blindly charge the requested amount without cross-referencing inventory databases.
Client unit prices are strictly discarded. Catalog lookups and Indian GST slabs (12%, 18%, 28%) are calculated deterministically on the server with proportional discount allocation.
Prompt-injected or looping agents can execute excessive bulk purchases or stack unverified promotions beyond merchant risk tolerances.
Hard merchant limits (max 50% discount, Rs 50,000 order total, 10 units/SKU) and sliding-window anomaly velocity checks hard-block malicious transactions before session completion.
Multi-turn agent negotiations double-book scarce stock, or abandon checkout sessions mid-flow leaving merchandise locked indefinitely.
Sessions soft-hold stock during negotiation. Uncompleted sessions automatically expire after 30 minutes, and a dedicated sweeper safely returns inventory to stock.
Rejected sessions and price corrections vanish into unformatted log streams, preventing human merchants from auditing automated agent behavior.
Every state mutation, agent identity, rejection reason, and before/after cart total is written immutably to Firestore and viewable in real time.
Engineered for Protocol Precision and Safety
Dual-engineered for AI developers building standardized ACP agent clients and merchants demanding bulletproof balance-sheet protection.
Server-Authoritative Pricing and Indian Tax Slabs
Client-submitted unit prices are ignored. The adapter queries the server catalog for authoritative prices and executes Indian GST slab calculations (12%, 18%, 28%) with proportional pre-tax discount allocation, ensuring zero paise drift across multi-item orders.
Deterministic Guardrails and Anomaly Engine
Hard mathematical ceilings (Max 50% discount, Rs 50,000 order total, 10 units/SKU). A sliding-window anomaly scoring engine tracks agent velocity to prevent abusive flash-buying without relying on unpredictable LLM evaluators.
Soft-Hold Inventory and Expiry Sweeper
Stock is soft-reserved upon checkout session creation to prevent double-booking. Sessions not completed within 30 minutes expire automatically, triggering the sweeper endpoint to safely release merchandise back to available inventory.
Cryptographic Rails and Reliable Webhooks
Inbound Razorpay webhooks are verified with HMAC-SHA256 constant-time signature comparison. Outbound webhook events dispatch to agent listeners with exponential backoff retries and dead-letter queue logging on delivery failures.
Simulate Autonomous Buyer Scenarios
Experience live checkout session state transitions, price-tampering defense, discount ceiling enforcement, and TTL sweeper mechanics directly from the adapter's test suite.
Protocol Sandbox Simulator
Simulate autonomous buyer transactions against the payment rail
Standard Autonomous Checkout Flow
Autonomous buyer agent negotiates 2 items, applies valid promotional discount, attaches tokenized payment method (pm_tok_*), and bridges to Razorpay Orders API.
{
"id": "cs_e44bcc7178a943f8",
"status": "completed",
"line_items": [
{
"product_id": "prod_bolt_001",
"quantity": 2,
"unit_price": 499,
"tax_rate": 0.18
},
{
"product_id": "prod_bolt_004",
"quantity": 1,
"unit_price": 2499,
"tax_rate": 0.18
}
],
"totals": {
"subtotal": 3497,
"discount": 100,
"tax": 611.46,
"total": 4008.46,
"currency": "INR",
"tax_breakdown": [
{
"rate": 0.18,
"subtotal": 3397,
"tax": 611.46
}
]
},
"payment_provider": {
"provider": "razorpay",
"razorpay_order_id": "order_TXqxPXnQFo5pt2"
},
"is_anomalous": false,
"anomaly_score": 0
}Deterministic Finite State Machine (FSM)
Checkout sessions transition strictly through deterministic states. Completed, rejected, refunded, and cancelled states are terminal and cryptographically locked against further mutations.
Cart Initialized
Authoritative totals calculated. 30-min TTL timestamp assigned. Soft-hold inventory reserved.
Negotiation & Patching
Agent modifies item quantities or applies coupon codes. Guardrail bounds re-evaluated on each turn.
Delegated Token Attached
Full shipping address provided. Agent attaches delegated token (pm_tok_*).
Razorpay Order Settled
Inventory committed. Razorpay order created (order_*). Immutable audit entry written.
Triggered on guardrail breach (discount > 50%, total > Rs 50k, qty > 10). Session locked and unpayable.
Triggered by explicit agent cancellation or 30-min TTL expiry sweep. Soft-held stock returned to available catalog.
Post-payment reversal calling Razorpay Refund API. Locks session into audited refunded state.
Engineered for Absolute Determinism
Autonomous agents require mathematically predictable boundaries. Explore the full three-zone runtime topology and the seven-state deterministic checkout FSM.
System Topology & Component Interactions
End-to-end data flow bridging autonomous AI buyer agents through the AgentPay Bridge internal engine matrix into Razorpay settlement rails and Google Cloud Firestore.
- >Discovery & Catalog: GET /.well-known/agent.json and GET /products
- >Authoritative pricing invariant to client prompt injection
- >Deterministic guardrails: 50% discount ceiling, Rs 50k order cap, 10 units/SKU
- >30-minute stock soft-hold mutex with automated TTL sweeper
- >Cryptographic Razorpay webhook verification and append-only audit stream
Integrate with Any Autonomous Buyer Agent
Standardized HTTP endpoints matching Agentic Commerce Protocol v2026-04-17. Compatible with LangChain, AutoGen, CrewAI, or standalone custom buyer scripts.
# 1. Register Autonomous Buyer Agent
curl -X POST https://razorpay-acp-adapter-922729192321.asia-south1.run.app/agents/register \
-H "Content-Type: application/json" \
-d '{"name": "Aura Autonomous Buyer Agent #42"}'
# 2. Discover ACP Capability Manifest and Search Catalog
curl https://razorpay-acp-adapter-922729192321.asia-south1.run.app/.well-known/agent.json
curl "https://razorpay-acp-adapter-922729192321.asia-south1.run.app/products?category=general&in_stock_only=true"
# 3. Create Checkout Session with Authoritative Pricing
curl -X POST https://razorpay-acp-adapter-922729192321.asia-south1.run.app/checkout_sessions \
-H "X-API-Key: acp_agent_8f0a394bc..." \
-H "Idempotency-Key: idemp_9428591a" \
-H "Content-Type: application/json" \
-d '{
"line_items": [{"product_id": "prod_bolt_001", "quantity": 2}],
"buyer": {"name": "Aura Agent", "email": "aura@buyer.internal"}
}'Inspect Live Autonomous Transactions
Open the live audit dashboard to observe autonomous buyer agents negotiate carts, trigger deterministic guardrails, and bridge transactions into Razorpay orders.