The past week produced a small but meaningful shift in the AI conversation.

For most of the past year, discussion around AI has centered on capabilities. The focus has been on what models can produce and automate: writing code, analyzing datasets, generating content, assisting with research, and performing pieces of professional work that previously required human labor.

Those developments have been impressive, but they have largely remained inside the domain of software productivity.

This week the conversation moved one layer deeper.

The emerging question is not only what AI agents can think or generate. It is whether they can transact.

That question exposes a structural limitation that has been easy to overlook. AI agents cannot meaningfully participate in the economy without payment infrastructure. They can recommend purchases, coordinate workflows, and plan transactions, but without wallets, identity systems, and settlement rails they remain advisory systems rather than economic actors.

Over the past week, several infrastructure developments suggested that this constraint is beginning to change.

Developers and infrastructure companies are now experimenting with agentic payment systems: frameworks that allow AI agents to hold wallets, execute payments, and interact directly with digital services.

If that capability develops further, something subtle begins to shift in how the internet functions. The network stops being defined primarily by humans operating software and begins to include software transacting directly with other software.

Once that dynamic appears, the underlying financial rails start to matter in a different way.

Part I: the agent wallet moment

Several developments this week pointed in the same direction.

AI agents are getting wallets

Coinbase CEO Brian Armstrong summarized the shift in a short comment on X:

“Every AI agent deserves a wallet.”

Armstrong made the remark while discussing internal Coinbase experiments in which AI agents hold stablecoin wallets and operate somewhat like “digital employees,” capable of executing certain payments autonomously.

The idea initially sounds futuristic, but the underlying problem is fairly practical.

If an AI agent is responsible for coordinating infrastructure, purchasing data, paying API providers, or routing tasks across services, the workflow eventually reaches a familiar bottleneck. At some point the system needs to move money.

Traditional financial infrastructure was never designed for non-human actors. Opening financial accounts requires legal entities, identity verification, regulatory compliance procedures, and institutional relationships. These processes assume a person or an organization operating within a defined jurisdiction.

Crypto wallets operate under a different set of assumptions. They function as programmable financial identities that can be created instantly and controlled directly by software.

In the context of AI agents, that distinction starts to look less ideological and more architectural.

The x402 standard and machine-to-machine payments

One of the most widely discussed infrastructure developments this week was Coinbase’s x402 protocol, a proposed standard designed to give AI agents programmable wallets and allow them to execute micropayments autonomously.

Developers demonstrated agents using the protocol to perform tasks that involve small automated payments during software workflows. In several examples, agents paid for blockchain data through Alchemy APIs, executed microtransactions in USDC on Base, and paid incrementally for computation during autonomous processes.

In one demonstration, an agent queried blockchain data and immediately paid for the request using USDC on Base through x402. The request, payment, and execution occurred within the same automated interaction.

In traditional financial systems that workflow would require billing infrastructure, user accounts, invoices, and delayed settlement. Programmable payment rails collapse those steps into a single event that software can execute directly.

Welcome to x402 - Coinbase Developer Documentation

Coinbase Developer Documentation

Bitcoin and Lightning enter the conversation

The experimentation around agentic payments is not limited to Ethereum-based ecosystems.

Developers have also demonstrated autonomous workflows using Bitcoin and the Lightning Network, where agents receive funds and automatically route payouts across multiple destinations based on predefined rules.

One demonstration showed an agent receiving USDT and distributing funds between a bank account and a BTC wallet according to programmed allocation logic.

These experiments remain early, but they illustrate a broader point. Once payments become programmable primitives, they can be embedded directly inside automated systems. Financial transactions begin to function like software instructions rather than administrative processes.

That shift is what makes agent-based commerce technically possible.

Even TradFi is building agent payment rails

The same shift is beginning to appear within traditional payment infrastructure.

This week Mastercard announced Verifiable Intent, a system that uses cryptographic proofs to validate AI agent purchases and reduce fraud in autonomous transactions. The framework is designed to confirm that when an AI agent executes a purchase, the action corresponds to a previously authorized user instruction.

In practical terms, the system attempts to ensure that agent actions remain cryptographically linked to human intent.

Observers quickly noted that the architecture shares conceptual similarities with blockchain verification models, where transactions are validated through cryptographic mechanisms rather than relying exclusively on institutional trust.

The boundary between traditional payment systems and crypto infrastructure is starting to blur.

Part II: the machine economy thesis

Taken together, these developments point toward a broader shift.

AI agents are gradually moving from software tools toward economic participants operating within digital markets. For agents to function in that role they require three foundational capabilities: identity, access to services, and payment rails.

Most AI systems already possess the first two components. They can authenticate to services, call APIs, and coordinate workflows across digital infrastructure.

Payments remain the missing layer.

This is one reason crypto infrastructure keeps appearing in discussions about agent architecture. The explanation is not ideological enthusiasm for digital assets. It is the practical reality that crypto wallets are currently the only widely available financial primitive that software can access directly and control programmatically.

Why fiat systems struggle with agents

Traditional financial systems evolved around assumptions that center human participants.

Opening financial accounts requires legal identity, institutional relationships, regulatory compliance procedures, and geographic jurisdiction. These processes work reasonably well for individuals and corporations but translate poorly to autonomous software systems.

When an AI agent executes a transaction, several questions emerge immediately. Who owns the account? Who authorizes the transaction? Who bears legal responsibility for the resulting activity?

Crypto infrastructure approaches financial identity differently. Wallets function as programmable identities controlled through cryptographic keys rather than institutional onboarding processes.

An AI agent can generate a wallet, sign transactions, and interact with services across the internet without requiring institutional approval.

From an architectural perspective, that model aligns naturally with machine-mediated commerce.

Programmable payments unlock new business models

If AI agents can transact autonomously, a different set of economic models begins to emerge.

Agents could pay directly for API calls, purchase computing resources dynamically, negotiate services across platforms, or route payments between counterparties involved in a workflow.

These interactions lend themselves to granular pricing models. Instead of subscription billing or fixed contracts, services can charge per task, per query, or per unit of computation.

Traditional financial systems struggle with this type of microtransaction environment because the operational overhead of billing and settlement is relatively high. Programmable payment rails reduce that overhead and allow financial transactions to occur at the same speed and scale as software operations.

Why this matters for Open Money

The Open Money thesis has always focused on financial primitives that operate without centralized gatekeepers.

The emergence of agent-based commerce extends that idea in a new direction. Instead of building permissionless financial infrastructure solely for human users, the system must also support software participants operating autonomously inside digital markets.

Supporting that environment requires infrastructure capable of instant settlement, programmable payments, global accessibility, and machine-readable verification.

These properties align closely with the capabilities many crypto systems were originally designed to provide. The difference now is that the demand may increasingly come from machines rather than people.

Stress testing financial infrastructure

Another way to understand this moment is as a stress test for existing financial architecture.

Human financial systems developed in a world where transactions were relatively infrequent, settlement processes were slow, and identity was anchored to institutions such as banks and governments.

The emerging agent economy assumes a different environment. Transactions may occur continuously, globally, and at machine speed. Many of those transactions will be generated automatically as part of software-driven workflows.

Financial infrastructure that cannot accommodate those conditions becomes friction. Infrastructure that supports automated, programmable transactions becomes the default layer through which machine-mediated commerce operates.

Research backlog

Several open questions follow from this shift.

It remains unclear which financial identity frameworks will prove most effective for autonomous agents. Different approaches are emerging across crypto wallets, decentralized identity systems, and institutional payment networks.

There is also uncertainty around which settlement rails will dominate machine-to-machine commerce. Stablecoins, Lightning payments, and tokenized bank deposits each represent plausible architectures.

Regulatory frameworks will also need to adapt. Financial regulation historically assumes identifiable individuals or corporations. Systems in which autonomous agents initiate transactions introduce new questions about authorization, accountability, and compliance.

Finally, the development of an agent economy raises questions about where new infrastructure chokepoints might appear. Wallet providers, infrastructure platforms, and agent marketplaces could all become coordination layers within this ecosystem.

Closing thought

For years, advocates of programmable money argued that software-native financial infrastructure would eventually enable new forms of economic coordination.

The claim often sounded abstract because the applications were unclear.

AI agents may provide the first environment where that infrastructure becomes operationally necessary. When software begins transacting directly with other software, payment rails are no longer theoretical components of financial architecture.

They become part of the operating system of the internet.