A2A agentic payments: The full vision and implementation challenges

In today’s payment innovation landscape, a new generation of tools is beginning to emerge: intelligent agents powered by large language models (LLMs), capable of executing transactions autonomously. Their greatest potential is unlocked when operating over direct, account-to-account (A2A) payment schemes, removing traditional intermediaries and enabling much more seamless purchase experiences.

As A2A payments gain ground as an efficient alternative to traditional card-based systems, the emergence of these agents raises an even bolder question: what if we don’t just change how we pay, but also who makes the decision and executes the purchase?

This possibility is already becoming relevant in e-commerce, where the friction between purchase intent and execution remains high. In this context, an ambitious and concrete proposal is emerging: to build intelligent shopping assistants capable of searching, evaluating, and executing purchases within parameters defined by their users.

This innovation challenges the classic four-party payment architecture — user, merchant, issuing bank, and acquiring bank — suggesting an evolution toward a model with up to six actors, adding the intelligent agent (as transactional executor) and its tech provider or platform (as the interaction facilitator).

The core promise of these agents is clear: to give our language models the ability to efficiently and autonomously analyze massive amounts of information to make purchasing decisions. We envision assistants capable of identifying the best offers in terms of price, quality, and delivery time — freeing people from manual search and comparison tasks.

The vision even extends to the obsolescence of conventional checkout. In the near future, these agents are expected to interact directly with merchant systems to complete transactions without needing a graphical payment interface. This shift has the potential to optimize the user experience, allowing us to invest our time in more meaningful activities instead of navigating repetitive and cumbersome processes. Going to the store is no longer the problem: the tedious part now is spending hours scrolling without deciding what to buy.

However, moving from theory to practice requires solving structural challenges. The key question is: how do you manage the flow of funds so that an agent can complete a purchase autonomously and securely?

At Prometeo, we’ve begun addressing this challenge through experimentation with so-called “Agent Accounts”: This concept involves creating virtual, nominative bank accounts linked to the agent’s owner, to which a specific balance can be assigned for autonomous transaction execution.

This proposal introduces two key advantages that make the use of automated agents for payments viable:

• Identification of the agent’s owner: A nominative bank account is used — meaning an account where the account holder is formally identified and registered. This establishes a direct connection between the agent and its owner, facilitating accountability and enabling precise auditing of its operations.
• Enforced operational limits: By assigning a specific balance to the agent’s account, transaction limits are automatically set. This allows for control over the agent’s activity scope and helps mitigate risks in cases of failure or misuse.

With this infrastructure in place, the next logical question arises: what kinds of operations can an agent perform within an A2A payments environment?

In our exploration, we’ve focused on the three core operations enabled by Prometeo’s Model Context Protocol (MCP):

1. Pay-In (Receiving Funds): The agent can receive payments directly into its account, just like any other banking user. It is also programmed to automatically notify the managing system when funds are received. This enables use cases such as receiving refunds, service payments, or reimbursements — giving the agent the ability to handle both income and outflows.
   
2. Pay-Out (Issuing Payments): Once the agent identifies a product or service that meets the predefined criteria, it can issue a payment from its account to a third party. This makes the agent a complete transactional actor, capable of making decisions and concluding operations on its own, without human intervention.
   
3. Transaction Verification: Adding an extra layer of security is crucial — especially in the context of A2A payments, where transactions settle in real time and reversing them requires manual processes outside the standard flow. This verification step helps validate conditions before executing the payment, reducing risks due to errors or misuse.

Once the agent has identified the desired product and has the necessary funds, the next challenge arises: can it actually interact with the merchant platform and complete the purchase without human intervention?

Here’s where the gap between vision and reality becomes clear.

We have analyzed the Model Context Protocols (MCP) of leading platforms like Shopify, Mercado Libre, Stripe, and PayPal. Surprisingly, as of May 31, 2025, none of them have enabled operational solutions that allow autonomous agents to make native purchases within their ecosystems.

The absence of defined standards or built-in functionalities from major payment and e-commerce platforms poses a major challenge: how can we automate and optimize interaction between intelligent agents, products or services, and the current payment infrastructure?

This dilemma echoes the early days of open banking, when traditional financial systems also lacked programmable interfaces that allowed third parties — like fintechs — to access data or perform operations securely and automatically. For years, that lack of access was a critical barrier to innovation.

The emergence of pioneering startups like Prometeo, Plaid, and Tink marked a turning point. These companies identified the structural need and began building APIs — application programming interfaces — that served as technological bridges to banking systems. In the early stages, these APIs relied on user-authorized software robots capable of checking balances, extracting data, or executing programmed transactions. This created an access layer that allowed a new generation of companies to develop innovative and inclusive financial services on top of what had previously been a closed infrastructure.

Today, in the world of e-commerce, we face a similar moment. Intelligent agents encounter the same lack of native connectors once faced by fintechs. And while the path is still undefined, there are signs that — once again — need and experimentation will drive the first breakthroughs.

Until standard interfaces are established that allow language models (LLMs) to interact seamlessly with online retail platforms — including catalogs, payment systems, and conditions like pricing or return policies — agents must operate on the web as it exists today: designed for humans, not machines.

To do this, they rely on a familiar tool in the world of automation: headless browsers — browsers without a graphical interface that simulate real user behavior programmatically.

These tools allow intelligent agents to browse websites without screens or clicks, and perform actions like:

• Extracting relevant information from e-commerce platforms (such as pricing, availability, or delivery times),
• Feeding that structured data to LLMs,
• And generating automated decisions based on the user’s context and preferences.

The agent can then execute concrete actions: selecting products, scheduling deliveries, or completing payment.

This way, the entire purchase cycle is automated — from expressing the initial need, through searching and comparing options, to making an informed decision and executing the transaction. All without direct human intervention.

The idea of an intelligent agent that can search, decide, and pay for you sounds incredible. But between vision and reality lie several obstacles that can’t be solved with a single line of code.

First, technical fragmentation. Every country has its own A2A system with different rules. What works in PIX doesn’t apply in SPEI, and so on. There’s currently no easy way to scale these agents across multiple markets without building a Frankenstein of integrations. But this is exactly where platforms like Prometeo make a difference — we’ve built a shared layer of access and pan-regional interoperability to bring order to all this banking chaos.

Second, the legal vacuum. Who’s liable if an agent makes a mistaken transfer? There’s still no specific regulation. In Mexico and Brazil, the law assumes that if your credentials were used, it’s your fault — even if it was the bot’s error. And while Brazil is advancing principles of liability for AI systems, applying them to payments is still in the early stages. That uncertainty keeps most players on the sidelines. It’s exactly the kind of gap that calls for practical safeguards — which is why we’ve designed a setup where agents operate through pre-funded, owner-linked accounts, adding traceability and control while regulation catches up.

Third, security. A2A payments settle in real time, and while some reversals are possible, they’re not easy. This demands fraud detection within milliseconds. What’s needed are more adaptive models and pre-transfer validations — such as verifying the recipient’s bank account — to ensure the money goes to the right place before sending it

Fourth, the black box. Today’s language models don’t explain why they do what they do. But if they’re going to move our money, they’ll have to be accountable. How do we audit an agent? How do we avoid bias or unfair decisions?

These challenges aren’t deal-breakers, but they do create friction — and that opens up space for innovation.

What if the future of payments isn’t just about speed or UX — but about redefining control itself?

Intelligent agents aren’t just tools that execute. They interpret, decide, and act. That changes the game.

The road ahead is messy: fragmented systems, legal grey zones, and unanswered questions around trust. But, like open banking, this won’t be solved by waiting — it’ll take builders willing to move without a map.

Maybe what’s coming isn’t just a new way to pay, but a new logic for how decisions happen online — where intent and execution converge.

The real work isn’t just technical. It’s about accountability, design, and earning trust.

And if anyone’s still asking “internet is dead?” — building agents that act on real human intent might be the best proof it’s very much alive.