🧠 The Agentic Payments Map

Welcome to Fintech Brainfood, the weekly deep dive into Fintech news, events, and analysis. You can subscribe by hitting the button below, and you can get in touch by hitting reply to the email (or subscribing then replying)

Weekly Rant 📣

🧠 The Agentic Payments Map

ACP, UCP, A2P, AXTP, x402.

If your eyes just glazed over, you’re not alone.

The agentic payments conversation has become an alphabet soup of competing protocols, each claiming to solve the problem of how agents pay for things. And that’s before you look at anything the card networks are doing, with their own emerging standards.

Why so many? Because they’re not all solving the same problem.

There are commerce protocols (how agents buy stuff from merchants). There are B2B protocols (how agents pay invoices). And there are agent-native protocols (how agents pay each other for compute, data, and API access at machine speed).

Different problems. Different protocols. One confusing mess.

So I made you a map.

To understand why there are so many protocols, and why they’re not as redundant as they appear, we need to split a few things apart:

1. The internet itself runs on protocols that work together

2. Agents are getting a similar hierarchy, but it looks different

3. The internet never got a payment layer

4. Agents expose this gap because they’re a new actor

5. But first: the trust layer that makes payments possible

6. Commerce protocols are the most mature

7. B2B payment flows have different needs

8. Agent-to-agent flows are where it gets exciting

9. How does all of this fit together?

1. The internet has many protocols but one experience

When you’re doomscrolling you’re probably not thinking about TCP/IP, HTTPS, or the video codec streaming content to you. You’re probably not thinking at all.

That’s the point. When protocols work together well, the experience is seamless. But the path there is often messy.

The internet relies on a stack of canonical protocols to function:

TCP/IP (Transmission Control Protocol/Internet Protocol)

• IP handles addressing and routing—so your request gets to the right place

• TCP ensures reliable delivery—so your email isn’t garbled and no bits get lost

DNS (Domain Name System)

• Translates human-readable domain names (like google.com) into numerical IP addresses

HTTP/S (Hypertext Transfer Protocol/Secure)

• HTTP requests and transfers web pages and media

• HTTPS adds encryption (via TLS/SSL) for secure communication

New protocols get added quietly all the time—gRPC for efficient application communication, WebSocket for persistent connections. Around these sit certificate authorities, CDNs, and a whole swathe of infrastructure that makes it work.

The internet isn’t static. It evolves.

2. The internet never got a payment layer

HTTP status code 402 is “Payment Required.” It was reserved in 1997. Never implemented.

The internet was designed to move information, not money. Commerce got bolted on top through specialized financial messaging systems—your browser talks HTTPS to the merchant, who talks to a payment gateway, who talks to financial networks like Visa or ACH.

There’s no single, universally adopted protocol that standardizes the entire lifecycle from “add to cart” to “settled payment.”

This is precisely why ACP, UCP, and the others are emerging—they aim to create a structured language for agents to communicate commerce intent.

3. Agents expose this gap because they’re a new actor

The plumbing that worked for humans at keyboards doesn’t work for software making decisions at machine speed.

Consider what happens when an agent tries to buy something on your behalf:

• It’s a new customer. How does an LLM decide which store and product to choose? Merchants now have to optimize for agents, not just humans. Welcome to “Agent Engine Optimization” or AEO.

• It’s using a new channel. The chat interface is the new checkout. Your conversion funnels, A/B tests, and cart abandonment emails? Largely irrelevant when there’s no page to load.

• It creates new fraud risks. How do you know this agent is using a legitimate card from an authorized user, and not a stolen credential at machine speed?

These are genuinely hard problems. And they’re being attacked from multiple directions by multiple players, which has led to an explosion of new protocols.

But before we get to payment protocols, there’s a layer that has to come first: trust.

4. The trust layer: knowing who you’re dealing with

MCP and A2A handle how agents communicate. But as the ERC-8004 specification puts it: “these agent communication protocols don’t inherently cover agent discovery and trust.”

Before agents can transact, they need to establish legitimacy. This is especially critical for merchants, who’ve spent years blocking bots. Now they need to let the right bots through.

Two protocols are emerging to solve this:

ERC-8004 creates on-chain registries for identity, reputation, and validation. The spec explicitly shows how an agent’s registration file can list MCP endpoints, A2A agent cards, ENS names, and DIDs together. It’s designed to work with existing communication protocols, not replace them.

Visa TAP takes a different approach—verifiable signatures that prove an agent is: A Visa-trusted agent with commerce intent (not a scraper or attacker). Acting on behalf of a recognizable consumer (via loyalty accounts or device identifiers). Carrying valid payment credentials that merchants can verify

The takeaway: Trust is a prerequisite for payment. These protocols answer “should I believe this agent?” before anyone asks “how should this agent pay?”

5. Commerce has the most new protocols

Commerce means buying goods or services as a business or human. With agentic commerce, you outsource that purchase moment to an agent. The transaction still involves buying a thing—like a Fallout Vault Boy figurine or a PlayStation subscription—but now the agent handles discovery, selection, and payment.

Firstly, two protocols are emerging to standardize this flow:

These protocols achieve similar consumer experiences in different ways:

• UCP (Google) is an orchestration framework. Merchants publish “capability manifests” that agents discover and negotiate with. Think of it like DNS for commerce—agents find merchants and learn what they can do.

• ACP (OpenAI) is a transaction standard. It defines exactly how to structure a cart, generate a payment token, and complete checkout. More prescriptive, less flexible.

The takeaway: If you’re a merchant wanting to appear in both ChatGPT and Gemini, you’ll likely need to support both. For now. Your trade-off appears to be that UCP requires more work but offers greater flexibility across the entire flow. ACP works with a lot of what’s already there.

Secondly, there are those designed to enforce at the payment network level.

The takeaway: In practice, both standards are likely very similar—the core difference is that tokens work with their respective networks. These network-level tokens ensure consumer protections, chargebacks, and fraud are handled as merchants expect.

6. B2B payment flows have different needs

Commerce gets all the attention, but B2B payments are where the volume actually lives. Think invoices, supplier payments, payroll—the unsexy plumbing of business. Here, the agents tend to talk directly to the rail.

Agents handling B2B flows face different challenges:

• Payments are larger and less reversible

• Invoices require matching, approval workflows, and audit trails

• The rails are slower (ACH, wire) but also more flexible

Stablecoins offer card-like guarantees of payment success plus programmability, but there isn’t yet a universally adopted standard with real traction.

(If you’re wondering why x402 appears in my diagram for this layer: it does offer escrow-like features, but it isn’t widely adopted as an internet standard, and the companies building for B2B have more complex needs than it currently solves.)

7. Agent-to-agent payments have huge potential

This is where it gets interesting.

The internet has valuable resources locked behind API keys and subscriptions. If you want Google Maps data or ChatGPT in your application, you create an account, pre-fund it, get a key, and then you can access the service.

That model doesn’t scale for billions of agents writing code, buying from each other, and needing resources on demand.

Today this shows up as two friction points:

• The “ran out of tokens” problem: Your agent is coding and hits a limit. You have to manually top up before it can continue.

• The “need an API key” problem: Your agent needs a service, asks you to sign up, create credentials, and hand them over.

These friction points mean agents can’t truly be autonomous. They’re your junior dev, but you don’t trust them with the expenses card or company credentials yet.

Agent-native protocols aim to fix this:

AP2 deserves special attention:

• It’s an abstract protocol that works with both x402 and UCP—these aren’t mutually exclusive

• It’s part of Google’s broader A2A (Agent-to-Agent) communication protocol for enterprise security and coordination

• It uses verifiable credentials to create “cart mandates” (what the agent can buy), “intent mandates” (what the human wanted), and “payment mandates” (the stored payment credential)

Imagine trillions of fractional, instant micropayments for compute, data, and API calls. Imagine autonomous economies where agents transact for every resource consumed, creating massive new transaction volume for products that aren’t properly monetized today.

It sounds like sci-fi. But so did self-driving cars, robot dogs, and AI solving novel math problems. Those are here now.

8. How does all of this fit together?

In a word: messily. For now.

I created this I have two key takeaways:

First, a Google stack is emerging. Google’s approach (A2A → AP2 → UCP) creates a coherent stack that works for both commerce and non-commerce payments.

Second: These protocols operate at different levels of abstraction:

• Agent communication layer: How will we talk? Standardizing agent-to-agent messaging (MCP, A2A).

• Trust layer: Should I believe this agent? Identity and reputation (ERC-8004, Visa TAP).

• Mandate layer: Do we have credentials to pay? Verifying authorization (AP2 mandates, VIC/MAP tokens).

• Transaction flow layer: What are we paying for? Managing discovery, negotiation, and checkout (ACP, UCP).

• Authentication layer: Is this transaction legitimate? Security, fraud prevention, reversals.

• Payment rail layer: How do we complete the payment? Execution on cards, ACH, stablecoins.

These standards are emerging, incomplete, and not yet widely adopted. In five years they could all have disappeared—remembered like WAP and Betamax, forgotten standards lost to time.

But that assumes AI agents go away. Which seems unlikely.

If you’re a merchant, payments company, or bank, you can’t ignore this. But it’s also nearly impossible to pick winners. Here’s how I think about it:

1. Google has created internet standards before and will do it again. A2A, AP2, and their children will likely stick around, especially given Google’s dominance in search and growing share of LLM chat.

2. Merchants should focus on commerce first. Supporting ACP and UCP is how you appear in both ChatGPT and Gemini, the two dominant consumer LLMs.

3. Trust infrastructure matters more than you think. ERC-8004 and Visa TAP are solving problems that will only get harder as agent volume scales. Watch this space.

4. B2B payment flows are a massive, wide-open opportunity. Stablecoins have adoption but no clear standard yet. Expect that to change soon.

5. Agent-native payments are a generational opportunity. Stablecoins are the obvious fit—fast, cheap, 24/7, programmable. x402 is a start, but it isn’t ready for prime time.

The next era of agentic payments will be defined by how these protocols fit together and inherit capabilities from each other. Just as A2A → AP2 → UCP creates a coherent stack, expect similar compositions to emerge for B2B and agent-native flows.

The path there will be messy. Standards will compete. Some will die.

But the underlying shift—software that can discover, negotiate, and pay for resources autonomously—is coming regardless of which protocol wins.

Now at least you have a map.

ST.

4 Fintech Companies 💸

1. Radius - The payments network for AI Agents

Radius enables up to 2.5m transactions per second with sub second settlement. It comes with an SDK to quickly integrate monetization capabilities to your AI agent or for them to autonomously buy access to things like paywalled datasets or APIs. To achieve this speed the radius network gives its own agents a signed transaction, it then locks the balances of both users, changes the state of the data across its partitions and then unlocks those balances.

🧠This is quite different to how blockchains work, and much more of a classic PKI approach to signing transactions and data management. It’s differentiated engineering. But I wonder if that alone is enough to get traction. Go to market is still such an unknown in the agentic commerce space. And right now the big names have all of the momentum. I wonder if and how that changes?

2. Onsetto - Business Account Switching

Onsetto helps a bank’s customers switch their accounts payable, payroll, and operating payment flows into a new account during onboarding. Businesses switching get a progress bar of status, and an AI assistant to check in with at any stage. 

🧠Opening new business accounts is easy, getting the balance, treasury and relationship is harder. An account that is dormant just costs money. Relationship bankers burn calories trying to get the account active and it never happens. Moving business account is 100x harder than a consumer so there’s a ton of inertia. Anything that helps that is a good thing.

3. GrowthPal - AI Deal sourcing for M&A

GrowthPal identifies early and mid-stage company candidates for M&A based on strategic intent, the sector context and willingness to transact. It helps create the mandate (to source a company), identify high-fit targets, and position growth ideas for the combined entities. 

🧠 They’re selling to corporate development teams like Razorpay and Mindtree, but you have to imagine the M&A bankers would love this. But the product also removes some of the mystique of the M&A banker. Thirsty for the transaction, deep in relationships. Both the most friendly and commercial bunch on the planet. 

4. Meld - The stablecoin payments network

Meld supports 54 local payment methods, in 184 countries to enable companies to accept and deliver stablecoins from a single API. It adds value by scoring on-ramps by conversion, and helps customers get the best possible price for their crypto or stablecoin transfer. It helps route transfers across blockchains to maximize conversion and minize cost.

🧠 Orchestration is a tough business, but a network is a great business. There are probably 50 stablecoin orchestration companies out there. But only a couple building networks, partnering with every wallet, every orchestrator, every on ramp and local payment method. That’s a powerful place to be (Mesh is a similar business).

Things to know 👀

1. Nubank receives conditional approval to establish a US Bank.

Nubank says the charter allows them to expand their operational footprint and the “launch of deposit accounts, credit cards, lending and digital asset custody.” (Emphasis mine). The US organization will be led by co-founder Cristina Junqueira who has relocated to the US. Nubank now enters an organizational phase and is pending approval from the FDIC and Federal Reserve.

Here’s Cristina speaking at Fintech Nerdcon 2025 about how and why:

🧠 Nubank is a serious and well-organized tech company that happens to be a fully licensed bank in markets where it operates. With a $77bn market cap, 60% of Brazilian adults bank with Nubank. 1 in 4 banked Mexicans use Nubank. If they achieve anything near that in the US, it will be a major disruption.

🧠 What now? 12 months to capitalize. 18 months to open. Still needs FDIC and Fed approval. By mid 2027 we will start to see if its’s working.

🧠 Who’s the target market? "Consumers that know the brand" — read: the 65M+ Hispanic Americans, 92% of whom already use fintechs.

🧠 Monzo, Bunq, Revolut, and others have tried, stumbled, and tried again. Nubank, Klaros Partner, and Davis Polk worked incredibly fast to get this done. And have much more experience in market expansion with full licences than peers.

🧠 Conditional approvals are coming thick and fast, but will they last? One fear I’ve heard is its ok to get your charter, but what about your organization, or even your first exam? What the policy level is doing will be different to how the staff examiners view things.

🧠 Staying power. Nubank is a listed company with a track record on lending, loss ratios, TPRM, regulatory reporting, and much more. Even if after the midterms we have a slightly less tech-forward administration, I’d never bet against Nubank’s success. They’re a force of nature.

2. CFTC Chairman statement on events contracts walks back prior guidance.

In a statement, Chairman Michael S. Selig said, "I have directed CFTC staff to withdraw the 2024 event contracts rule proposal that would prohibit political and sports-related event contracts." He directed staff to withdraw (1) the 2024 proposed rule that would have prohibited political and sports related event contracts, and (2) the 2025 staff advisory that warned registrants about offering sports event contracts amid ongoing litigation.

Instead, Selig says the CFTC will move forward with new rulemaking to set clear standards for event contracts and give market participants more certainty. He also directed staff to reassess the CFTC’s role in ongoing court fights, stating that where jurisdiction is at issue, the Commission has a responsibility to defend its exclusive jurisdiction over commodity derivatives

🧠 This is amid multiple states actively litigating to classify sports events as gambling. Maryland, New Jersey, Nevada, New York, and Ohio, along with Native American tribes, argue that event contracts are "sports betting in disguise" that evade state licensing, taxes, and consumer protection regulations.

🧠 States vs Federal is an age-old issue, but at stake is tax revenue. They’re asking who will manage licensing, age gating, responsible gambling controls, taxation, and enforcement. What’s the right perimeter here, and when is it gambling and when is it not?

🧠 There’s already a gold rush for prediction markets; this will only accelerate that. Wall St sees a huge opportunity to bring consumers into new asset classes. Fanduel and other sportsbooks are now coming into prediction markets. They will be everywhere.

🧠 I’m ambivalent to how regulation comes, but I do care deeply that consumer protections are uniform. If these products are going mass market, uniform protections beat a patchwork—but only if the CFTC actually writes and enforces them. 

🧠 Passing a rule is one thing, keeping it is another. We saw how the Trump administration dismantled 1033 pretty quickly. A future administration could potentially do the same for prediction markets in time.

3. Stablecoin network Mesh raised $75M Series C at a $1B valuation.

Mesh* says it's raising, bringing its total funding to over $200 million. Mesh solves the increasing fragmentation in the crypto industry, supporting many blockchain networks, in many jurisdictions, with many wallets. This means consumers can pay with any asset (e.g., Bitcoin or Solana) and the merchant receives their preferred asset (e.g., USDC).

🧠 Mesh* is adjacent to many stablecoin infrastructure companies. They connect exchanges (e.g., Binance, Coinbase) with payment companies (SHIFT4), with Fintech companies (Revolut, PayPal), and wallets (MetaMask).

* Mesh is a sponsor of the Tokenized podcast agentic commerce series.

Good Reads 📚

1. Who gets to give people money?

Alex gives a detailed history of the “can stablecoins have yield” or not debate following Coinbase's withdrawal at the 11th hour from a proposed bill in the Senate Banking Committee. Alex takes the view that stablecoin issuers not directly passing yield, but it going through 3rd parties like Coinbase, is a loophole and not what GENIUS intended. But that equally the banks claims of “dangerous and undersierable” draining of deposits is really the big banks pretending to be small banks but really defending their margin., 

Crucially, the current draft allows rewards tied to activity like using a wallet, making a payment, providing liquidity, or other forms of activity.

Tweets of the week 🕊

That's all, folks. 👋

Remember, if you're enjoying this content, please do tell all your fintech friends to check it out and hit the subscribe button :)

Want more? I also run the Tokenized podcast and newsletter.

(1) All content and views expressed here are the authors' personal opinions and do not reflect the views of any of their employers or employees.

(2) All companies or assets mentioned by the author in which the author has a personal and/or financial interest are denoted with a *. None of the above constitutes investment advice, and you should seek independent advice before making any investment decisions.

(3) Any companies mentioned are top of mind and used for illustrative purposes only.

(4) A team of researchers has not rigorously fact-checked this. Please don't take it as gospel—strong opinions weakly held

(5) Citations may be missing, and I’ve done my best to cite, but I will always aim to update and correct the live version where possible. If I cited you and got the referencing wrong, please reach out