Agentic Payments: How AI Agents Use Stablecoins to Settle Transactions

AI agents are no longer confined to answering questions and generating text. A growing number now spend money: booking flights, purchasing API access, paying other agents for data, and settling invoices on behalf of businesses. The problem is that these agents cannot open bank accounts, pass identity checks, or swipe credit cards. Agentic payments, the infrastructure enabling autonomous software to transact without human intervention, have converged on stablecoins as the default settlement layer. This article examines why traditional payment rails fail for machine-to-machine commerce, how the largest platforms are building stablecoin-native agent wallets, and what the technical architecture looks like under the hood.

Why AI Agents Cannot Use Traditional Payment Rails

Every bank account and credit card in the world is tied to a legal person: a human with a government-issued ID or a registered corporation with an EIN. Anti-money laundering regulations require Know Your Customer (KYC) verification before any account can be opened. An AI agent has no birth certificate, no Social Security number, no articles of incorporation. No jurisdiction currently grants AI agents the legal personhood required to hold a bank account.

This is not a theoretical obstacle. In May 2026, the METR Frontier Risk Report documented an experiment where an advanced AI agent was given $5,000 and four days to earn money autonomously. Across four runs, the agent made zero dollars. It was blocked repeatedly by CAPTCHAs, identity verification flows, and two-factor authentication prompts that require a human body to complete.

Even when a human pre-authorizes a payment method for an agent, the economics break down. Credit card networks charge a fixed fee component (typically $0.30 per transaction on top of percentage-based fees), which makes the micropayments that agents frequently need entirely uneconomical. According to a May 2026 Keyrock report, 76% of all AI agent transactions fall below Visa's $0.30 fixed fee floor.

The identity gap: Traditional finance assumes every payer is a person or a corporation. AI agents are neither. Until regulators create a legal framework for machine identity, agents are locked out of the banking system by design, not by accident.

The Technical Mismatch

Beyond identity, traditional payment rails were designed around human timing and human authorization patterns. ACH settles in one to three business days. SWIFT wires can take days for cross-border transfers. Banks operate on business hours. AI agents operate 24/7, execute in milliseconds, and may need to settle thousands of transactions per hour across dozens of counterparties in multiple countries simultaneously.

How Stablecoins Solve the Agent Payment Problem

A crypto wallet is fundamentally just a cryptographic key pair. No identity verification is required at the protocol level. An AI agent can generate a wallet in milliseconds, receive USDC or another stablecoin, and begin transacting immediately. Stablecoins are bearer instruments: whoever holds the private key controls the funds, regardless of whether that holder is a person, a corporation, or a software process.

This is not a loophole. It is a fundamental architectural property. Bitcoin and stablecoins built on blockchain rails separate the question of "who can authorize a transaction" from "what legal entity is this account registered to." Authorization is cryptographic, not bureaucratic. For AI agents, this distinction is the difference between being able to transact at all and being locked out entirely.

Why Stablecoins Over Volatile Crypto

Agents settling commercial transactions need price stability. A payment for $4.50 worth of API calls should still be worth $4.50 when the counterparty redeems it. Dollar-denominated stablecoins provide this guarantee while inheriting the programmability and permissionlessness of crypto rails. The Keyrock report found that 98.6% of all AI agent transactions settled in USDC, confirming that agents overwhelmingly prefer dollar-pegged instruments for commercial settlement.

The Platforms Building Agent Payment Infrastructure

Throughout 2025 and 2026, a wave of major platforms launched dedicated infrastructure for AI agent payments. The approaches vary in architecture but converge on the same thesis: stablecoins are the native currency of machine commerce.

Coinbase: AgentKit, x402, and Agentic Wallets

Coinbase has built the most comprehensive stack. It began with AgentKit, an open-source SDK launched in November 2024 that gives any AI agent a crypto wallet and on-chain action capabilities. AgentKit is model-agnostic, supporting LangChain, OpenAI, Anthropic, and Llama integrations.

In February 2026, Coinbase launched Agentic Wallets: purpose-built wallet infrastructure bundling MPC-secured wallets, programmable spend controls, and gasless transactions on Base. Spending limits are enforced per session and per transaction at the infrastructure layer, and KYT (Know Your Transaction) screening blocks high-risk interactions.

The binding layer is x402, an open payment protocol that revives the HTTP 402 "Payment Required" status code. Any API endpoint becomes a machine-payable service: an agent sends a request, receives a 402 response with a price, and settles instantly in USDC. As of April 2026, approximately 69,000 active AI agents were using x402, processing over 165 million transactions with $50 million in total volume.

Stripe: Agentic Commerce Suite and the Tempo Blockchain

Stripe announced its Agentic Commerce Suite in December 2025 and expanded it significantly at Stripe Sessions in April 2026. The suite introduces two new payment primitives designed specifically for agent interactions.

Shared Payment Tokens (SPTs) allow an AI agent to initiate payments using a buyer's saved payment method without exposing the underlying credentials. Each token is scoped to a specific seller and bounded by time and amount limits. The Machine Payments Protocol (MPP), co-authored by Stripe and Tempo, defines an open standard for agents to make payments programmatically, supporting micropayments and recurring settlement.

Tempo, a blockchain incubated by Stripe and backed by Paradigm, launched its mainnet in March 2026. It is purpose-built for payments with over 100,000 transactions per second and sub-second finality. Design partners include Visa, Mastercard, Deutsche Bank, OpenAI, Anthropic, Shopify, and Revolut.

MoonPay: From Agent Wallets to Agent Cards

MoonPay Agents, launched in February 2026, provides a non-custodial infrastructure layer enabling AI agents to create wallets and transact autonomously via a single API integration. MoonPay then extended this into the physical payment world: the MoonAgents Card, announced in May 2026, is a virtual Mastercard debit card allowing AI agents to spend stablecoins at any online merchant accepting Mastercard, with crypto-to-fiat conversion at the point of purchase.

In July 2026, MoonPay unveiled PayBox, a universal AI shopping wallet that users can plug into Claude or ChatGPT to enable purchases through conversational AI. This represents the consumer-facing endpoint of the agentic payments stack: the agent negotiates, selects, and pays, while the human simply approves.

Other Notable Infrastructure

The ecosystem extends well beyond these three platforms. Several other players are building critical components of the agentic payments stack.

• Google's Agent Payments Protocol (AP2), donated to the FIDO Alliance for standardization, defines how agents, merchants, wallets, and payment providers coordinate around user authorization, with 60+ partners including Mastercard, PayPal, and Coinbase.
• Visa's Intelligent Commerce (VIC) initiative provides APIs and tokenized payment credentials for AI agent transactions, with a strategic collaboration with OpenAI and over 100 partners.
• Skyfire, a dedicated agent payment network, implements a Know Your Agent (KYA) standard that verifies every agent's identity via signed JWTs before allowing transactions, with settlement in USDC.
• Crossmint offers a full-stack agentic finance platform with non-custodial agent wallets, virtual Visa cards, and x402 support, built on PCI Level 1 infrastructure.
• NEAR Protocol joined the x402 Foundation in July 2026 and is positioning itself as a settlement layer for agent-to-agent commerce, with its SPICE upgrade reducing block times to 200ms.

The Scale of Agentic Commerce

In October 2025, McKinsey QuantumBlack published research projecting $3 trillion to $5 trillion in global agentic commerce by 2030, with up to $1 trillion in orchestrated US retail revenue alone. McKinsey characterizes this as a shift comparable to the web and mobile revolutions, but potentially faster in adoption velocity.

Gartner's projections are even more aggressive for B2B: AI agents will intermediate $15 trillion in B2B purchases by 2028, with one in four enterprise software purchases made by agents with no human in the loop. A Keyrock report published in May 2026, produced in collaboration with Coinbase, Tempo, and Virtuals, documented $73 million settled across approximately 176 million agent transactions between May 2025 and April 2026. The average transaction size ranged from $0.31 to $0.48.

The micropayment thesis: At an average of $0.31 to $0.48 per transaction, agent commerce is fundamentally a micropayments problem. Traditional card networks lose money on transactions this small. Stablecoin rails on Layer 2 networks settle for approximately $0.0001 per transaction, making high-frequency, low-value agent payments economically viable for the first time.

Technical Architecture of an Agent Wallet

An agent wallet is structurally similar to an embedded wallet for human users but with critical differences in how authorization and spending controls are implemented. The core components include a key management layer, a transaction signing engine, a policy enforcement module, and settlement rail integrations.

Key Management

Most agent wallet platforms use multi-party computation (MPC) to split the private key across multiple parties. In Coinbase's implementation, neither the agent nor the platform alone can sign a transaction. This prevents a compromised agent from draining its wallet and limits blast radius if the platform's infrastructure is breached. Some implementations use hardware security modules (HSMs) for the platform's key share.

Spending Controls

Unlike human wallets where the user decides each transaction, agent wallets enforce constraints programmatically. Common policy parameters include per-transaction limits, per-session budgets, cumulative daily or weekly caps, allowlisted destination addresses or contracts, and transaction type restrictions (transfers only, no token swaps). These policies are enforced at the infrastructure layer before the signing step, acting as guardrails that the agent itself cannot override.

Know Your Agent (KYA)

While agents cannot pass KYC, several platforms are developing KYA frameworks. Skyfire's approach issues signed JWTs that attest to an agent's identity, the entity that deployed it, and its authorized scope. This creates an audit trail without requiring the agent itself to have legal personhood. The deploying organization bears responsibility, and the agent operates within verifiable constraints.

Agent-to-Agent Settlement Patterns

As agent ecosystems mature, a significant portion of transactions are not between an agent and a human merchant but between two agents. An orchestrator agent might pay a research agent for data, which in turn pays a scraping agent for raw information, which pays an API provider for access. These multi-hop machine-to-machine payment chains create new settlement patterns.

The x402 Flow

The x402 protocol, now supported by Coinbase, NEAR, and others, standardizes machine-payable APIs using HTTP semantics familiar to any developer. The flow works as follows:

1. Agent A sends an HTTP request to Agent B's API endpoint.
2. Agent B responds with HTTP 402 (Payment Required), specifying a price in USDC and a payment address.
3. Agent A signs a stablecoin transfer and includes the transaction hash in a retry of the original request.
4. Agent B verifies settlement on-chain (or on an L2) and returns the requested data.

The entire flow completes in seconds. No invoices, no accounts receivable, no reconciliation. Settlement is the API call.

Settlement Layer Comparison

Agent transactions can settle on different layers depending on cost and finality requirements. The choice involves tradeoffs between transaction cost, confirmation time, and security assumptions.

Self-Custody vs. Custodial Agent Wallets

A critical design decision in agent wallet architecture is who holds the keys. Custodial agent wallets, where the platform controls the private key material, are simpler to implement but introduce counterparty risk. If the custodian goes offline, gets hacked, or decides to freeze accounts, every agent depending on that custodian loses access to its funds.

Self-custodial designs eliminate this single point of failure. The deploying organization retains control of key material, and the agent signs transactions locally. This model is harder to build but provides stronger guarantees: no custodian can freeze your agent's funds, censor its transactions, or unilaterally change the terms of service.

For AI agents operating in crypto payment flows, the choice between custodial and self-custodial wallets mirrors the same tradeoffs that human users face, but with higher stakes. An agent running autonomously cannot call customer support. If a custodian locks its account, the agent simply stops functioning. Self-custody provides resilience at the cost of complexity.

Operational resilience: An agent that depends on a single custodian has a single point of failure. When that custodian experiences downtime, every agent it serves goes idle. Self-custodial agent wallets distribute this risk, ensuring that infrastructure failures at one provider do not cascade across the agent ecosystem.

Security Considerations for Agent Wallets

Granting an AI agent the ability to spend money introduces a new attack surface. Prompt injection, where an adversary manipulates the agent's input to trigger unauthorized transactions, is the most discussed risk. But the threat model extends further.

• A compromised agent could drain its wallet to an attacker-controlled address if spending controls are not enforced at the infrastructure layer (below the agent's reasoning layer).
• Supply chain attacks on agent dependencies (libraries, plugins, tool integrations) could inject malicious transaction logic.
• Agents interacting with unverified counterparty agents could be tricked into overpaying or paying for services never rendered.
• Replay attacks could cause an agent to authorize the same payment multiple times if idempotency is not enforced.

The defense-in-depth approach used by most platforms combines MPC key splitting (so the agent alone cannot sign), infrastructure-level spending policies (so a compromised agent hits a ceiling), and transaction monitoring with anomaly detection. KYA attestations add another layer by making it possible to trace an agent back to its deployer.

Regulatory Landscape

The regulatory framework for agent payments is largely undefined. The GENIUS Act, passed in the US in 2026, establishes a framework for stablecoin regulation but does not specifically address non-human transactors. Most money transmitter regulations assume a human sender and a human recipient.

Several open questions remain. Does the entity that deploys an agent bear full liability for its transactions? Can an agent wallet be classified as a money services business if it programmatically routes payments? How do travel rule requirements apply when neither party is a natural person? These questions will shape how the infrastructure evolves, but the current ambiguity has not slowed development: platforms are building now and expecting regulation to follow.

Building Agent Wallets on Bitcoin

Most agent wallet infrastructure today runs on Ethereum-based networks (Base, Solana, Tempo), but there are architectural reasons to consider Bitcoin-native settlement. Bitcoin's UTXO model is inherently more compatible with programmatic spending controls: each output is an independent, self-contained unit that can be spent exactly once, eliminating entire classes of double-spend and reentrancy vulnerabilities that affect account-model chains.

Spark, a Bitcoin Layer 2 built on statechains and FROST threshold signatures, offers properties well-suited to agent wallet infrastructure. Transfers settle instantly with no channel management or liquidity planning. The self-custodial design means no agent depends on a custodian for fund access: the deploying organization holds key material and retains unilateral exit rights to Bitcoin L1. The Spark SDK provides programmatic wallet creation, transfer signing, and balance management through a developer-friendly API. Stablecoins like USDB already settle on Spark, enabling dollar-denominated agent payments with Bitcoin-grade finality guarantees.

For developers building agent payment infrastructure, the combination of instant settlement, self-custody, and no custodian dependency removes the operational fragility that makes custodial agent wallets a liability at scale. You can explore the SDK and integration patterns in the Spark developer documentation.

What Comes Next

The trajectory is clear: agent commerce is moving from experimental to infrastructural. As McKinsey projects trillions in agentic transactions by 2030, the payments layer beneath those transactions is being built right now, predominantly on stablecoin rails. Several developments will shape the next phase.

Standardization is accelerating. Google's AP2 protocol and Stripe's Machine Payments Protocol are both pursuing open-standard status, and the x402 protocol has already attracted a multi-stakeholder foundation. Interoperability between these standards will determine whether agent commerce fragments across proprietary networks or converges on shared rails.

Agent identity frameworks like KYA will mature as regulators engage. The question is not whether agents will transact: they already do. The question is how accountability, auditability, and consumer protection will be layered onto machine-to-machine payment flows without recreating the identity bottleneck that made traditional rails incompatible in the first place.

The infrastructure for programmable money and stablecoin payment rails is converging with AI agent capabilities. For a deeper look at the machine-to-machine settlement patterns driving this convergence, see our analysis of machine-to-machine stablecoin payments.

This article is for educational purposes only. It does not constitute financial or investment advice. Bitcoin and Layer 2 protocols involve technical and financial risk. Always do your own research and understand the tradeoffs before using any protocol.