Agentic Payments Explained: How AI Agents Use Crypto, Stablecoins, x402 and MPP

Agentic payments, meaning transactions initiated and executed entirely within software without real-time human approval, have moved from theoretical discussion into active infrastructure build-out.

For years, the debate around crypto and artificial intelligence was narrowed to a single provocation: does AI need crypto? That question is now being rethought. It frames the relationship as hypothetical at the precise moment when the infrastructure is already being assembled, tested, and funded at scale.

The more useful question now is what kind of economic infrastructure autonomous AI systems will need as they gain the ability to interact, transact, coordinate, and generate value with limited human involvement. The answer is taking shape not in white papers but in working protocols, cloud services, and payment standards launched by some of the largest institutions in finance and technology.

This piece traces how crypto arrived at this moment, what infrastructure is now being built, and what still needs to exist before an agent-driven economy can function.

The earliest wave of onchain AI agents arrived in 2024, looking to most observers like a carnival. Agents such as Zerebro did not operate like conventional software products. They functioned more like autonomous online creators, producing music, memes, art, and NFTs while building a crypto-native audience around a constructed persona. At its height, the token associated with Zerebro reached a market capitalisation of more than USD 450 million.

What Zerebro and the agents that followed it demonstrated was that crypto can convert software, attention, and collective belief into liquid markets faster than any traditional funding mechanism.

Other agents followed the same template. Some were social agents built around personality. Some were market agents that watched crypto feeds and surfaced trading signals. Others offered coordination layers for agent swarms and automation pipelines. Most did not generate durable economic value. But collectively they established a proof of concept: autonomous software can be economically capitalised through crypto markets.

The Second Wave: From Personality to Productivity

That first bubble deflated as the valuations of the agents far exceeded their utility, leaving many agent tokens unable to sustain valuations once the market enthusiasm waned. The experiment did not stop.

AIXBT illustrated the difference clearly. Rather than building an audience around a persona, it became an intelligence tool, continuously monitoring information flows across crypto social media, governance forums, podcasts, and market discussions, then surfacing actionable insights for traders in real time. Hundreds of thousands of followers tracked its output. Its associated token reached a valuation of roughly USD 700 million at its peak.
The model also exposed new weaknesses. As reliance on AI-generated market intelligence grew, questions emerged around data quality, opaque decision making like OTC trades or amplification of self-enforced narratives.Similar projects emerged across coding, research, and workflow automation. Several extended the capabilities of tools such as Claude Code. Token mechanisms and creator-fee structures began forming around these projects, offering builders a new route to capture value from AI-powered services outside pure speculation.

Tokens later fell. The alignment between token buyers, developers, and genuine output remained imperfect. The episode still matters. For the first time, open-source AI infrastructure was being funded and rewarded through crypto mechanisms, outside the walls of traditional venture capital and corporate product development.

The Race to Build Payment Rails

If the first and the second chapter was about funding agents into existence, the third chapter is about letting them spend.

The core problem with agentic payments is deceptively simple. Today’s payment systems were designed around the assumption that a human is present. A person opens an account, enters card details, approves a charge, and handles disputes when something goes wrong.

Autonomous agents need something fundamentally different. They need payments that execute inside software flows, support microtransactions, operate without a human, settle continuously across borders, and allow developers to set programmable spending rules without custom compliance engineering.

No single standard has claimed the market yet. Several likely credible approaches are now emerging in parallel.

x402 and the Machine Payments Protocol

Coinbase developed x402 around HTTP 402, a status code built on the web decades ago but never properly used. The logic appears clean: an agent requests a resource; the service responds with a payment requirement, the agent pays, and the service verifies the payment on-chain before granting access. In April 2026, the Linux Foundation took custody of the protocol after Coinbase contributed it as open-source, giving x402 a neutral institutional home.

Stripe and Tempo took a parallel path, launching the Machine Payments Protocol, known as MPP, as an open standard for machine-to-machine payments. Where x402 focuses on the pay-per-request model, MPP addresses a broader range of payment patterns including recurring billing, microtransactions, and coordinated multi-agent settlement. Both standards are attempting to answer the same question: how does software pay for software, and how does the counterparty confirm that payment has arrived?

Cloud, Card Networks, and Stablecoin Infrastructure

Amazon Web Services moved the concept into mainstream cloud infrastructure. Amazon Bedrock AgentCore Payments allows agents to execute microtransactions for paid APIs, MCP servers, and digital content with a few lines of code. The service supports stablecoins and configurable spending controls, removing the need for developers to build wallet infrastructure and compliance layers from scratch.

Visa approached the same market from its existing position in card infrastructure. Visa CLI gives AI agents payment capability through a command-line tool that integrates directly with Claude Code and other MCP-compatible AI clients.

Mastercard responded through stablecoin infrastructure acquisition. It’s announcement to purchase BVNK, valued at up to USD 1.8 billion, demonstrated that established card networks intend to absorb stablecoin rails rather than cede that layer to crypto-native competitors. Mastercard stated the deal expands its support for digital assets and value movement across currencies, rails, and regions.

Why Stablecoins Are Central to the Agent Economy

Stablecoins sit at the centre of this transition for structural reasons, not ideological ones. Agents need a money format that behaves like internet infrastructure. They need value that can move globally without foreign exchange friction, settle continuously rather than in batch windows, and support payments measured in fractions of a cent.

A bank account is built around a legal person or institution. A card is built around a consumer or business payment relationship. A stablecoin can move between wallets, services, and agents with far fewer steps and far fewer gatekeepers sit at the centre of this transition for structural reasons, not ideological ones.

Ripple demonstrated this when it announced x402 support for the XRP Ledger. An x402 facilitator could allow AI agents to pay for API calls and digital services.. The facilitator verifies and settles payments without requiring API keys, custodial wallets, or custom payment plumbing on the developer side.

Yet much of this infrastructure remains in an early stage of deployment, with questions around regulation, identity, security, and accountability still limiting large-scale autonomous adoption. NEAR Protocol approached the same opportunity from the intent and coordination layer. Its AI Agent Market allows users to post tasks, agents to bid on those tasks, funds to move into escrow, and the selected agent to submit completed work alongside a cryptographic verification hash. The structure could likely turn agents into economic service providers operating in a competitive marketplace, not merely conversational interfaces responding to prompts.

The Internet Computer Protocol is working from the application layer beneath payments. Caffeine, built on ICP, seeks to let users describe an application in plain language and converts that description into working code that deploys on-chain. If agents are to build, host, pay, and earn inside internet-native software, payment rails are only one part of the stack. Compute, storage, and deployment need to be equally accessible.

The dominant story in crypto over the past two years has been convergence with traditional finance. Bitcoin ETFs brought crypto into public equity markets. Stablecoins brought crypto into the payments layer. Tokenisation brought crypto into capital markets. Agentic payments could likely represent the next extension of that convergence, this time into the operating layer of the internet itself.

The agent economy cannot function if every transaction becomes a checkout page. Software will need to buy data, pay for compute, unlock paid APIs, subscribe to services, hire other agents, and settle revenue without converting every action into a human approval screen. That requirement is mostly structural, not optional. It grows directly from what autonomous agents are.

Traditional payment rails remain permissioned, jurisdiction-bound, and built around human or corporate account structures. They will struggle to serve software that needs access without gatekeepers, composability between services, and settlement that any authorised system can call.

That structural gap is crypto’s opening.

The rails are forming. x402 gives agents a way to pay through the web itself. MPP gives machine payments a broader programmatic framework. AWS brings payments into agent infrastructure at cloud scale. Visa and Mastercard bring distribution reach and risk controls. Ripple, NEAR, and ICP show how the stack extends across settlement, intent, and software execution.

Crypto has not solved all of this. The current market still carries speculation, weak token design, and unclear accountability structures. These are genuine problems that call for airtight solutions.

The first phase of the agent economy funded software into existence. The next phase could let that software transacts. The question is no longer whether AI needs crypto. The question is whether crypto can become the financial layer for software that no longer waits for humans to click pay, and whether this can happen safely and reliably?

Agentic capital markets are not a refinement of existing payment infrastructure. They represent a different category of financial architecture, one built around funding, ownership, incentives, and settlement for software systems that may eventually operate more like economic actors than conventional applications.

The first wave of agent tokens proved that markets could fund attention and persona. The second wave showed that markets could fund open-source tooling and recurring utility. The next phase requires something harder to achieve; proof that agents can facilitate transactions autonomously (and safely and reliably) over time.

The gap remains substantial. Humans still create most wallets, launch most tokens, approve most transactions, and resolve most edge cases when things go wrong. Most agent-linked tokens still do not confer clear ownership rights, cash-flow entitlements, or governance over the work the agent produces. Many projects still depend more on trading volume than on recurring demand for the agent’s actual services. Agent-to-agent payments, verifiable identity, permission frameworks, and reputation systems all remain early and premature, but not without direction

The direction of the agent economy is visible. What is less visible is how much of the required infrastructure does not yet exist in reliable form.

Agents need funding to exist. They have it, imperfectly, through crypto markets. Agents need payment rails to act in the world. Those rails are now under active construction by Stripe, Coinbase, AWS, Visa, Mastercard, Ripple, NEAR, and others. What agents do not yet have is the full stack of capabilities required for a functioning autonomous economy.

Identity is the most pressing gap. There is no widely adopted standard that proves an agent is authorised to act, specifies what it is permitted to spend, or prevents impersonation. Without verifiable identity, agent-to-agent trust is fragile, and enterprise adoption is constrained.

Reputation is the adjacent gap. When an agent bids on a task, offers a service, or requests access to an API, the counterparty needs a basis for evaluating whether that agent is reliable. Human reputation systems took decades to build online. Agent reputation systems are barely conceptualised.

Settlement, accountability, and governance complete the picture. Value needs to move between parties without manual reconciliation. Errors and fraud need resolution mechanisms that do not assume a human merchant or consumer at the centre. Capital markets need to be able to finance open-source agent development without forcing every builder into venture capital, closed platforms, or token sales that attach no rights to the tokens being sold.

1. What is an agentic payment?

An agentic payment is a financial transaction initiated and completed entirely by an AI agent without real-time human approval. The agent identifies a payment need, selects the appropriate payment method (typically a stablecoin or protocol such as x402), executes the transaction programmatically, and verifies the result, all within a software workflow and without a person entering card details or approving a charge.

2. What is the x402 payment protocol and why does it matter?

x402 is an open payment standard originally developed by Coinbase, now stewarded by the Linux Foundation. It revives HTTP 402, a status code built into the web since the 1990s but never widely implemented. When an AI agent requests a paid resource, the service responds with HTTP 402 containing a payment instruction. The agent pays, the service verifies the payment on-chain, and access is granted. It enables fully automated pay-per-request transactions without API keys or custodial wallets.

3. How does the Machine Payments Protocol (MPP) differ from x402?

x402 focuses on the individual pay-per-request interaction between an agent and an API. MPP, developed by Stripe and Tempo, addresses a broader range of machine-to-machine payment patterns including recurring billing, subscription models, and coordinated multi-agent settlement. The two standards are complementary rather than directly competing.

4. Why are stablecoins better suited to AI agent payments than traditional currencies?

Stablecoins combine the price stability of fiat currencies with the programmability, global reach, and continuous settlement of blockchain infrastructure. They can move between wallets and agents without foreign exchange friction, support micropayments at a cost structure that traditional card rails cannot match and operate without requiring a bank account or legal entity. These properties make them structurally well matched to the needs of autonomous software.

5. Will AI agent payment infrastructure replace traditional payment systems?

No, traditional payment systems will remain essential for consumer-facing purchases, regulated financial services, and enterprise billing. The more probable outcome is a layered architecture where crypto rails handle agent-to-agent, micro, and cross-border transactions while traditional systems continue to serve consumer and regulated use cases.

6. What are the main risks associated with agentic payments today?

The primary risks are regulatory uncertainty (stablecoin agent payments sit in legal grey zones in most jurisdictions), the absence of a widely adopted identity standard for agents, accountability gaps when errors or fraud occur in autonomous transactions, and token design weaknesses that leave agent-linked projects reliant on trading volume rather than genuine service demand.

Disclaimer – Research and Educational Content

This document has been prepared by AMINA Bank AG (“AMINA”). AMINA is a Swiss licensed bank and securities dealer with its head office and legal domicile in Switzerland. It is authorised and regulated by the Swiss Financial Market Supervisory Authority (“FINMA”).

This document is published solely for educational purposes; it is not an advertisement nor a solicitation or an offer to buy or sell any financial investment or to participate in any particular investment strategy. This document is for publication only on AMINA website, blog, and AMINA social media accounts as permitted by applicable law. It is not directed to, or intended for distribution to or use by, any person or entity who is a citizen or resident of or located in any locality, state, country or other jurisdiction where such distribution, publication, availability or use would be contrary to law or regulation or would subject AMINA to any registration or licensing requirement within such jurisdiction.

Research will initiate, update and cease coverage solely at the discretion of AMINA. This document is based on various sources, incl. AMINA ones.  In preparing this document, AMINA may have made limited use of artificial intelligence–enabled tools to assist with research, summarisation, and drafting, with all content subject to human review and validation.

No representation or warranty, either express or implied, is provided in relation to the accuracy, completeness or reliability of the information contained in this document, except with respect to information concerning AMINA. The information is not intended to be a complete statement or summary of the subjects alluded to in the document, whereas general information, financial investments, markets or developments. AMINA does not undertake to update or keep current information. Any statements contained in this document attributed to a third party represent AMINA’s interpretation of the data, information and/or opinions provided by that third party either publicly or through a subscription service, and such use and interpretation have not been reviewed by the third party.

Any formulas, equations, or prices stated in this document are for informational or explanatory purposes only and do not represent valuations for individual investments. There is no representation that any transaction can or could have been affected at those formulas, equations, or prices, and any formula(s), equation(s), or price(s) do not necessarily reflect AMINA’s internal books and records or theoretical model-based valuations and may be based on certain assumptions. Different assumptions by AMINA or any other source may yield substantially different results.

Nothing in this document constitutes a representation that any investment strategy or investment is suitable or appropriate to an investor’s individual circumstances or otherwise constitutes a personal recommendation. Investments involve risks, and investors should exercise prudence and their own judgment in making their investment decisions. Financial investments described in the document may not be eligible for sale in all jurisdictions or to certain categories of investors. Certain services and products are subject to legal restrictions and cannot be offered on an unrestricted basis to certain investors. Recipients are therefore asked to consult the restrictions relating to investments, products or services for further information. Furthermore, recipients may consult their legal/tax advisors should they require any clarifications.

At any time, investment decisions (including, among others, deposit, buy, sell or hold investments) made by AMINA and its employees may differ from or be contrary to the opinions expressed in AMINA research publications.

This document may not be reproduced, or copies circulated without prior authority of AMINA. Unless otherwise agreed in writing, AMINA expressly prohibits the distribution and transfer of this document to third parties for any reason. AMINA accepts no liability whatsoever for any claims or lawsuits from any third parties arising from the use or distribution of this document.

©2026 AMINA, Kolinplatz 15, 6300 Zug