Share
Four payment architectures for AI agents have been launched in the past year. Coinbase, Stripe, Google, and Visa have each staked positions in what is becoming a new payment stack.
AI agents have settled more than $73 million across 176 million transactions. Incumbents have spent more than $8 billion on acquisitions to secure their layer of the infrastructure. The settlement economics work. The trust infrastructure does not.
That gap is the defining tension in machine-native payments right now, and the industry is not talking about it clearly enough.
The four architectures that emerged over the past twelve months are not purely competing. They are stacking. Coinbase built x402, a crypto-native protocol that turns stablecoin wallets into universal API keys.
Stripe and Tempo launched MPP, a payment-method-agnostic standard routing cards, crypto and Lightning through a single HTTP flow. Google assembled AP2, an authorization layer allowing users to delegate spending power to agents via cryptographic mandates. Visa extended existing card rails to provision AI-ready tokenized credentials.
Each architecture solves a different layer of the same problem: how does an agent pay for something without a human approving every transaction?
The answer each company has converged on is delegation, giving agents bounded, pre-authorized access to funds within defined parameters. However, the parameters are the hard part.
A report examining agent payments infrastructure published by Keyrock identifies the real question as not which protocol wins, but which companies capture the most layers and therefore attract the most value.
Coinbase and Stripe each span five of six layers in the emerging stack. Coinbase controls settlement through Base, wallets through Agentic Wallets, routing through internal infrastructure, the protocol through x402, and governance as an AP2 partner. Stripe mirrors that coverage through Tempo, Privy, Bridge, which was acquired for $1.1 billion, MPP and its own compliance infrastructure.
Vertical integration at this pace means the companies building the payment rails are also defining the trust assumptions embedded in them. That is worth examining carefully.
The economic case for blockchain rails in agent payments is not ideological, but it is arithmetic. Across 176 million x402 payments, the median transaction falls between $0.01 and $0.10, with 76% of activity below the $0.30 card-fee floor. A fixed processing fee of roughly 30 cents makes sub-dollar payments structurally uneconomical on traditional rails. An agent paying three cents for a weather API call cannot route through Visa.
Coinbase for Agents is live!
AI systems can now place trades, manage cash & pay for APIs autonomously. Built on MCP + AgentKit + x402 protocol. Sandboxed sub-portfolios with hard parameter caps. Agent-driven finance just moved from dev labs to retail accounts.…
— Dr Efi Pylarinou (@efipm) June 16, 2026
Layer-2 stablecoin settlement costs approximately $0.0001 per transaction. That cost differential did not emerge from a philosophical preference for decentralization. It emerged because the micropayment use case that agent commerce requires was economically impossible on existing infrastructure.
The consequence is that crypto rails have already won micropayments by default, without a competitive fight. Of those 176 million payments, 98.6% settled in USDC. That concentration validates Circle’s positioning as the default settlement asset for machine commerce and simultaneously creates a single point of failure the industry is not discussing openly.
The entire agent payments ecosystem currently depends on one stablecoin issuer’s reserve management, regulatory standing and technical infrastructure.
Here is the failure mode that does not have a clean answer yet.
An AI agent is delegated $500 to manage a user’s cloud infrastructure costs for the month. The agent misinterprets its task parameters, or hallucinates a cost optimization strategy, or encounters an adversarial prompt injected through a third-party API response.
It commits $4,200 in payments before the anomaly is detectable. The settlement layer has already finalized the transactions. Blockchain finality, the property that makes crypto rails fast and cheap, is also what makes overspending irreversible.
The user approved the agent, the protocol processed valid transactions, and the wallet enforced spending limits. Yet if an AI agent makes costly mistakes, no clear party bears responsibility. Existing regulations do not assign liability, and insurance markets have yet to address the risk at scale.
The issue is no longer theoretical. AI agents already account for 37% of all Safe transactions on Gnosis Chain, exceeding 75% on peak days. Adoption is accelerating, but trust and accountability frameworks remain underdeveloped.
Three technical approaches are converging on the authorization layer problem. Each addresses a different dimension of the question of how agents spend money safely. None of them fully resolves the overspending liability question.
ERC-4337 account abstraction and EIP-7702 are the most discussed. EIP-7702, included in the Ethereum Pectra upgrade, allows an externally owned account to temporarily delegate execution to a smart contract. In agent payment contexts, this means a user wallet can grant an agent session-level spending authority, defined by time, counterparty whitelist, and value cap, without permanently transferring custody. The session expires. The agent’s authority is bounded and revocable.
The limitation is that EIP-7702 defines the envelope but not the enforcement logic. A session key authorizing up to $500 in cloud infrastructure payments does not automatically detect that a specific payment is the result of a hallucinated cost model rather than a legitimate optimization. The cryptographic authorization is valid. The economic decision behind it may not be.
Session keys, as implemented in protocols like Privy and across several smart account standards, attempt a tighter constraint by scoping authorization to specific contract addresses and function signatures. An agent with a session key scoped to a specific DEX router and a specific token pair cannot sign a transaction to an arbitrary address even if it attempts to. This is meaningful progress. It narrows the attack surface and limits the blast radius of a misaligned agent action.
But the design assumption embedded in session keys is that the risk is unauthorized access, an agent doing something outside its permission scope.
The harder problem is an agent doing something inside its permission scope that is wrong. An agent authorized to spend up to $500 per day on API calls can legitimately exhaust that limit in 90 seconds if its cost model is broken. The session key did its job. The outcome was still a failure.
ERC-8004, sometimes called the AI Agent Authorization standard, takes a different approach by proposing structured intent declarations that agents must submit before executing a payment.
AI agents need identity, not just a wallet.
Who built it. What it did. Why it can be trusted.ERC-8004 gives AI agents an on-chain identity that persists beyond a single session.
OpenClaw Bootcamp integrates this by default.
Apply →https://t.co/chKxrS4LSW https://t.co/eL3Su9npCU— LazAI Network (@LazAINetwork) June 15, 2026
The intent is logged onchain before settlement, creating an auditable record of the agent’s stated reasoning for a transaction. This does not prevent overspending but it creates a forensic layer that makes post-hoc liability attribution more tractable. If the agent’s declared intent was “pay for weather API, estimated cost $0.03” and the actual settlement was $47, the delta is legible onchain.
The problem is that intent declarations depend on the agent accurately representing its reasoning. A hallucinating agent does not know it is hallucinating. Its intent declaration will be internally coherent and factually wrong simultaneously.
MiCA, the GENIUS Act and the EU AI Act all reached enforcement proximity within weeks of each other in mid-2026. None of them address autonomous machine-to-machine transactions in a way that resolves the liability question.
MiCA regulates stablecoin issuers and crypto asset service providers. It does not define the legal status of an AI agent as a transaction originator, or assign liability when an agent-initiated transaction causes financial harm.
The GENIUS Act established the first federal licensing framework for payment stablecoin issuers in the United States when signed into law in July 2025, with implementing regulations due by mid-2026. It governs what stablecoins can be used for payments. It does not govern who is responsible when an agent uses a compliant stablecoin to make a payment that turns out to be wrong.
While the EU AI Act classifies AI systems by risk, its financial provisions focus on areas such as credit scoring and insurance, not autonomous AI agents making payments.
This leaves a regulatory blind spot. AI agents are already moving real capital, but clear rules on accountability and liability have yet to emerge.
The bottom-up trajectory is the most defensible framing for how machine-native payments scale from here. Crypto rails have won micropayments by default. As transaction volume builds and trust infrastructure matures, progressively larger payment values will migrate onchain. The constraint is not settlement capacity or protocol design. It is the trust layer.
Three things need to happen before that migration can reach transactions of meaningful size:
The payment infrastructure works. Agents can already settle transactions faster and cheaper than any previous payment architecture. The question is not whether machine-native payments will scale. It is whether the people building the rails are solving the right problem.
Right now, most of the engineering attention is on the settlement layer. The liability layer is where the actual risk sits.
Until agents stop hallucinating, they probably should not have unsupervised access to anyone’s funds. The identity primitives exist to constrain that access. The industry has not yet agreed on what constrained actually means.
Share
preferred on
