Mapping Agentic Protocol standards to India’s DPI Stack

I recently came across this post by one of my favorite Tech Blogger – Cobus Greyling’s on LinkedIn. Cobus created a visual of all the Protocol standards that are developed (being) by different organizations, non-profits, enterprise consortiums etc., for Agent based interaction. It covers

  • Agent consent / permissions
  • Agent service discovery
  • Agent-to-agent coordination and handoffs
  • Agentic commerce and payment settlement

Essentially a full ecosystem of interoperable protocols for supporting the Agentic Commerce workflows.

Image Reference - LinkedIn

This got me thinking, how will this stack collaborate and co-exist within the India’s DPI stack. That took me down a rabbit hole of trying to map each layer from this diagram into a native offering from India’s  DPI stack.

Let’s dive in.

India DPI Stack – A primer

India has built a vast array of Technology layer that embeds into lives of citizen in multiple ways every day. A few examples:

  • Identification
  • Commerce
  • Education
  • Banking
  • Healthcare
  • Retirement / pension etc.,

The most important components of India’s DPI layers include and not limited to - Beckn/ONDC, UPI, Account Aggregator/DEPA, DigiLocker, Bhashini and more in pipeline for healthcare services etc.,

I started mapping, how Cobus’s visual maps into India offering at a high level.

protocol stack · 7 layers, global and India, with a possible future agent-trust layer

 

The protocols are layers, not rivals, except in the purchase layer, where ACP, UCP and Beckn/ONDC coexist through adapters. A possible future "agent trust" layer (Know Your Agent, KYA, with a reported NPCI agent registry) would sit between consent and settlement; it is not a current standard.

Use case Analysis

We want further and started analyzing for specific use cases that are most common use cases in the day in the life of a citizen. Namely

  • Quick Commerce
  • Bill Pay
  • Banking

Quick Commerce

Request: "Get me milk, eggs and a birthday candle in 15 minutes, under ₹600."

  1. Permission (PAP + DEPA-style consent): the agent presents its limits: spend cap, categories, substitutions allowed or not.
  2. Tooling (MCP): the agent reaches catalog and cart tools through an MCP server, which in India can wrap a Beckn buyer app.
  3. Discovery (Beckn/ONDC or UCP): Beckn broadcasts a location-aware search to nearby dark stores and returns results through asynchronous callbacks. UCP works where the seller publishes a UCP profile or a gateway fronts the Beckn catalog.
  4. Coordination (A2A): the shopping agent negotiates substitutions with the seller's agent and hands the slot to a logistics agent.
  5. Cart and checkout (UCP, ACP or Beckn): the agent builds the cart, confirms the price and sets the address.
  6. Authorization and payment (AP2 then UPI): the signed mandate records the ₹600 cap; a UPI Reserve Pay block or one-time mandate executes it.
  7. Tracking and issues: order status, rider tracking, refunds.

 

Where it strains

  • Latency: Beckn's async callbacks must be collapsed into a quick answer for a 10-minute promise.
  • Stock freshness: inventory changes by the minute, so re-validate the cart at confirm.
  • Substitutions: the mandate must cover them explicitly, or the agent must ask again.

Quick commerce payment with KYA

 

The only new steps are registration (1 to 2), the credential check (9 to 10) and the optional suspension at the end. Discovery, cart and order handling are unchanged, because KYA is payments trust layer and not a commerce rail. A bank-side step may also be needed if NPCI's reported design of an NPCI-side agent talking to each bank's agent goes ahead.

Open questions to watch

  1. Will NPCI confirm the UAP and its agent registry, and on what timeline?
  2. Will the registry cover only UPI, or later cards and bill payments too?
  3. Will NPCI's registry interoperate with the Visa, Mastercard and International KYA framework (reported 10 September 2026), or will agents need to register in both?
  4. How will agent identities relate to ONDC's existing registry of buyer and seller apps?
  5. Who bears liability when a registered agent exceeds its authority?

Bill Pay within a cap

Request: "Pay my electricity bill when it's due, up to ₹4,000." The purchase rails drop out; this flow uses consent, MCP and settlement only.

 

Banking - find a better personal loan

Request: "Find me a better personal loan." This flow uses consent-based data sharing, A2A between agents, and identity checks at the lender.

 

Consent expires on its stated window; the agent cannot fetch statements after that without a new approval.

Where banking strains

  • Higher stakes: moving money and sharing financial data carry stricter rules than ordering groceries.
  • Delegation limits: UPI Circle-style caps (reported at ₹5,000 per transaction and ₹15,000 per month) may be too low for lending.
  • Liability: NPCI's liability framework for agent errors is not yet public.
  • Compliance: RBI authentication and e-mandate rules and the DPDP Act apply; confirm with counsel.

Who collaborates with whom

Protocol

Quick Commerce

Banking

PAP + Consent

Spend rules, substitution

Standing permission, data sharing consent

MCP

Catalog and cart tools

Account, bill and statement tools

A2A

Seller negotiations, logistics handoff

Comparing lender offers

Beckn/ONDC

Discovery and order rails

Rarely used

UCP/ACP

Alternate checkout rails

Rarely used

AP2+UPI

Cart-capped payments

Billpay, EMI, Mandates

Consent, mandates and UPI carry over between both scenarios; only the commerce rails change. In a possible future, KYA would add an agent-identity check at the settlement step in both.

Potential future state: KYA and an NPCI agent registry

Status: not current. This section is a projection from reported NPCI development. NPCI has not formally confirmed its Unified Agent Protocol (UAP) or the agent registry reported by Reuters in September 2026, and no published NPCI "KYA standard" was found. Know Your Agent (KYA) is the emerging idea that agents should be verified the way customers are under KYC: identity, authentication, limits, consent and revocation. Treat every row below as a possible direction, not a commitment.

KYA would add one check at the payment step and a way to suspend an agent.

Step

Today (as drawn above)

Possible future with KYA

Agent identity

Asserted by the agent

Registered agent credential checked against an NPCI registry, or a car-network KYA framework

Limits

Set in the consent and the AP2 mandate

Also enforced by the registry and the bank, so the agent cannot exceed them

Payment execution

AP2 mandate, then UPI Reserve Pay or a one-time mandate

Mandate plus agent credential, executed as delegated authority under UAP (reported to build on UPI Circle)

Revocation

The user revokes consent

The user revokes consent, or the registry suspends the agent

Audit

Mandate record and UPI reference

Mandate record plus a registry and transaction audit trail

Liability

Undefined

Depends on NPCI’s liability framework, which is not public

Potential future state: CBDC/Tokenized Assets and Agentic payments

Status: not current. The Reserve Bank of India (RBI) has been advancing concept papers and experiments exploring central bank digital currency (CBDC) and tokenization. CBDC initiatives could provide a foundation for a more streamlined approach to assigning and managing digital wallets associated with AI agents, including defining the controls under which those agents would be permitted to transact. As companies develop AI agents and agentic commerce begins to take shape, CBDC could play an increasingly important role in supporting payments initiated by these systems. Although not central to the Agentic Protocol stack, CBDC could become central to the whole process of Agentic payment. Its potential value would lie not only in transaction speed and convenience, but also in the ability to establish clear controls over wallet access and payment authorization. In this prospective model, an agent-linked wallet could support transactions within predefined permissions, aligning payment execution with the boundaries established for the agent’s commercial activities.

Conclusion

What started as an attempt to map a protocol diagram to India’s DPI stack leads to a simple takeaway: this is less about choosing one winning protocol and more about understanding how the layers fit together. In the workflows explored here, consent defines what an agent may do, MCP connects it to tools, A2A supports coordination, commerce protocols handle discovery and checkout, and mandates connect authorized actions to payment settlement. India’s DPI components provide the underlying rails; agentic protocols offer a way to connect those rails to user intent.

The three use cases also show why there is no single stack for every task. Quick commerce needs discovery, fresh inventory, substitutions and logistics. Bill pay follows a narrower path from consent to settlement. Banking adds consent-based data sharing, identity checks and stricter delegation constraints. The common thread is not maximum autonomy, but clearly bounded authority: what the agent can access, what it can spend, when it must ask again and how permission can be withdrawn.

The proposed KYA and NPCI agent-registry layer remains a future possibility in this analysis, not an established standard. It could add agent verification and suspension without replacing the commerce or payment rails, but interoperability, implementation details and liability remain open questions. Ultimately, the opportunity explored here is to make India’s DPI accessible through agents while keeping consent, accountability and user control intact. The real test is not simply whether an agent can complete a transaction, but whether it can do so within the authority the user granted.